Home

Arabic Typed Text Recognition in Graphics Images (ATTR-GI)

1. ip amm correlation in 1 out 0 2 1 8 5 2 29 correlation in 0 89 out 0 1 Appendix C Image Groups Table C 1 group A includes different style texts were written on white background 3 2 correlation in 0 91 8 correlation in 0 84 out 0 15 5 E Chay The quick brown fox jumps The quick brown fox jump correlation in 0 98 6011 8 correlation in 0 99 4 7
2. gagang Bale af jajal n ae WWW SIMPLESITES PS 0592331117 00970592331117 2 Direct Way 77 FB xavertisingl PR IMedia Palestine Ramallah Tok 970 2242 1004 Collular 972 597216778 972 699 427 698 SF t Way rg in Palestine Ramallah Tol 470 22621004 Calluler 972 607 210778 972 509 427 0
3. correlation 1921 out 0 correlation in 0 96 out 0 032 correlation in 0 94 01 3 10 0 lal jie correlation in 0 75 4 11 303 ane oe 6658 DS miha Be 7 _ VY amp correlation in 0 99 out 0 001 12 9 nia yay Jaai y ee m 3 tla correlation in 0 87 out 0 12 13
4. id au 7 91 5 FM q Lal aiall 0 8 3 B Figure A 5 Example takes by ABBYY Screenshot Reader by click Enter key for Arabic page 84 Arabic Typed Text Recognition in Graphics Images ATTR GI G Ar The result Microsoft Word Sex 07 Page Layout References Mailings Review View Add Ins AW 2 Je 2 E FA EUS aanbcc AaBbC AaBbce AAB 4aBbce 4 Pate AE B 2 U abe x x A EE 1 No 5026 Heading Heading 2 Title Subtitle Subtle Em eine lt Clipboard g Font 2 Paragraph _ jl Styles OOOO 0 2 Editing m EERE DCSE SIS 1 ia 11 4 2012 9 10 4 2012 C Language Editor Read Only Automatically select languages from the Following list English French German Italian Spanish Specify languages manually English H 5 Natural languages Abkhaz Adyghe Afrikaans Agul Albanian Altaic Armenian Eastern Dictionary support Armenian Grabar Dictionary support Armenian Western Dictionary support Avar Aymara Azerbaijani
5. 449 daad Ey correlation in 0 92 out 0 071 correlation in 0 82 out 0 17 15 alata correlation 1920 77 out 0 22 16 correlation in 0 99 out 0 0012 17 1 1 4 1
6. 9 3 a a a a a A ppan correlation 1820 97 0 4 correlation in 1 out 0 correlation in 0 73 00 6 106 Arabic Typed Text Recognition in Graphics Images ATTR GI Table C 2 group B includes different style texts were written on background filled with one color rather than white correlation in 1 out 0 correlation in 1 out 0 correlation in 0 96 out 0 034 correlation 1821 out 0 correlation in 1 out 0 ic A wiwi Ra des 1 6ge3 ual correlation in 0 99 out 0 035 correlation in 0 96 6011 7 correlation in 0 81 out 0 18 correlation in 0 99 out 0 001 gal correlation in 0 98 00 6 11 yd correlation 1821 out 0 correlation 1820 87 out 0 12 correlation 1821 out 0 correlation in 0 92 00 3 15 A 28 haja correlation in 1 out 0 107 Arabic Typed Text Recognition in Graphics Images ATTR GI Table C 3 group C includes d
7. Tt Toph min ian oe es 5 2o iho ay Mobo 61 Eemi ha F ae mari ari rir er M m i aa 70 l Wa PET 1 uaaa att EEE we 5 p ee BE o ms O 113 Arabic Typed Text Recognition in Graphics Images ATTR GI ly 114
8. s 8 5 2013 gt Gj s giai Jio 2012 2011 pial Figure 3 9 show samples after applying wiener2 and rangefilt filters 36 Arabic Typed Text Recognition in Graphics Images ATTR GI Connected label segmentation Our segmentation H BT 209059231131 Aig gual wes Wau Direc G aaa LER et RST 7522555 322 an 8 T AM apadi 8 5 2083 20127 2011 SME Sjaak ALS alal 2011 201217 8 5 2093 Figure 3 10 example to show localization result 37 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm
9. 8 P Sa correlation in 1 out 0 correlation in 1 out 0 correlation in 1 out 0 correlation in 0 84 out 0 15 correlation in 0 99 out 0 0077 correlation in 0 87 out 0 12 correlation in 0 98 out 0 014 20 Be correlation in 0 99 1 21 ty sipa yss correlation 1 20 89 6011 1 correlation 1820 79 out 0 2 correlation in 0 68 0 1 24 correlation 1820 88 6011 1 25 5 gone 2 men Bas correlation in 0 89 out 0 1 109 Arabic Typed Text Recognition in Graphics Images ATTR GI 26 correlation 1820 9 out 0 098 27 correlation 1820 95 011 46 28 x Magee F me correlation in 0 57 out 0 42 correlation in 0 79 2 correlation in 1 out 0 correlation
10. For collecting information from sub images we will use a simple technique This technique is read the image line by line Then check each pixel 43 Arabic Typed Text Recognition in Graphics Images ATTR GI in each line if it is black or white After that pick the number of each color in separate variables So we create a two dimension array contains three attributes column and Z row to connect each number of colors in each line Where Z is the number of rows in the sub image These attributes are row number number of white pixels and number of black pixels See algorithm 3 8 This algorithm is used because we need to exchange white to black background to make connected label algorithm worked in an efficient manner as we explained in section 3 2 2 1 So there are some region of crop images still has a black background not all of them So we need this algorithm to get information For example check the number of black and white pixels If black is greater than white in the region of crop image then you can swap the color for this region image only not for all sub images to be white background and so on Another usage is finding the maximum row that has large black pixels Also find the peak for the frequency of maximum small black pixels to know if the statement is Arabic or English You can see more usage for this algorithm in next section Algorithm 3 8 collect information Input BW3 region of Binary image
11. crop Distance between region height 1 halfline L imx 24 and half line 2 defhalf imx halfline 3 detbases tax Bas Distance between region height 4 If defhalf lt defBase 5 ratio L defhalf def Base 4 6 Else 7 ratio L defBase defhalf 8 End 9 ratio ratio Base Pseudo Code for Algorithm 3 14 Set of Rules Input 1 objarea is the ratio of foreground area to the region area 102 Arabic Typed Text Recognition in Graphics Images ATTR GI Input 2 Base the value of base line Input 3 halfline the value of half line Input 4 ratio the range could be found the base line below the half line Input 5 accum two dimension array includes in first column the times of strait line of black pixels in the base line are redundant Input 6 val the value in the second column of accum array Input 7 C the position in accum that contain maximum peak value Input 8 start_min_row start_min_col end_max_row end_max_col vectors these vectors contains the origin coordinates from the origin image Input 9 BW3 thinning image crop Input 10 12 blank BW image with white background and has the same size of original RGB image Output I2 output result image this include Arabic text with some unwanted data 1 Apply Pseudo Code for algorithm 3 7 select text region to return obj start_min_row start_min_col end_max_row and end_max_col 2 Convert white background to black 3 BW3 thinni
12. hua dana awa dered ia se 15 2 3 Optical Character Recognition Optical Text Recognition Applications 21 6 2006 E E A Rees ae aS 24 Research Methodology and TechniqUES sasassssasessssesessssesessssnns 24 3 1 Research Methodology ccccceccsececceeceeceeceesueceeeaeesueseecueceeseeseesaeeseeseess 24 3 2 TECHNIQUES Sa 27 2 2 1 6 66551 8 Seas o e o 2 Meee aia 27 3 2 22 6 9 29 3 2 3 POSt PrOCESSING orreri sa e EEEn 53 3 3 Implementation ssn 55 Chapter 000 58 Evaluation and DisCUSSION 22222 222 eae sane ean 58 4 1 19 11 seis a ahi RS 58 4 1 1 Experimental Settings 222 2 2 58 Arabic Typed Text Recognition in Graphics Images ATTR GI 63 2222220 Experimental results 4 1 2 Al 2 DISCUSSION e iini e Rea phe 71 Chapter Five 0000000 76 Conclusion and Future Work cccceceee eee eee ener eee eee ee eae ne
13. in 0 97 out 0 027 correlation in 1 out 0 correlation 1820 93 0 6 correlation 1820 99 out 0 005 110 Arabic Typed Text Recognition in Graphics Images ATTR GI In the next table we present the original image but in black white mode by marking the discovered text with red color Table C 5 set of results images for show i lt rey ec et 2 111 Arabic Typed Text Recognition in Graphics Images ATTR GI 27 25 26 PER e 1 Ay de ka 3 FN cei ane ny er oe 5 die Und hed 1 rer Warn ad oom 112 Arabic Typed Text Recognition in Graphics Images ATTR GI ely ad
14. sass 32 Figure 3 6 example 1 Apply connected label on colored background 33 Figure 3 7 example 2 Apply connected label on colored and white background 33 Figure 3 8 example 3 Apply connected label on black background 34 Figure 3 9 show samples after applying wiener2 and rangefilt filterS 36 Figure 3 10 example to show localization result sss 37 Figure 3 11 clearGraphic filter that represents the base structure for our work 56 Figure 3 12 structure represents the steps of preprossiNng ccseeeeeeeeeeeeeeeeeeeees 57 Figure 3 13 structure represents the two ways of 2051210551118 57 Figure 4 1 comparison between 50 and 90 image according accuracy and error ratio using first EVALUATION 66 Figure 4 2 comparison between 50 and 90 image according number of black pixels all in and out USING second EVAlUATION cccccecceececseeceeceeceeceeceesaeeseeseesueceesaeseesaesensaeeseenas 67 Figure 4 3 comparison between 50 and 90 image according accuracy and error ratio using second evaluation cece cece reece reenter ns 67 VIII Arabic Typed Text Recognition in Graphics Images ATTR GI Figure 4 4 comparison between groups A B C and D according number of black pixels all in and Out using second evaluation
15. 100 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 Arabic Typed Text Recognition in Graphics Images ATTR GI While b lt x If Bline b 0 b b 1 Else If not Bfinish bb Bline b accum bc 1 bb Bfinish 1 sumc 0 End If Bline b bb accum bc 2 sumc be bc 1 bb Bline b accum bc 1 bb sumc 1 End b b 1 End End If b gt x accum bc 2 sumc End Pseudo Code for Algorithm 3 12 find peak and its position Input 1 accum x 2 is a two dimension accumulator first column is the length value and the second column is how many times this value redundant in Bline vector use Pseudo Code for algorithm 3 11 Input 2 bc value where bc is length of accum two dimension array Output 1 maximum peak value Output 2 C the position of maximum peak 1 peak 1 101 Arabic Typed Text Recognition in Graphics Images ATTR GI 2 C 1 3 For ui from 1 to bc 1 4 peak accum ui 2 5 C ui 6 For uj from ui 1 to bc 7 If peak lt accum uj 2 8 peak accum uj 2 9 6 10 End 11 End 12 End Pseudo Code for Algorithm 3 13 find half line and Base line ratio Input 1 imx is the height region crop Input 2 Base value for the region crop use Pseudo Code for algorithm 3 9 Output 1 halfline the value of half line Output 2 ratio range that maybe found place of base line in the region
16. 67 Arabic Typed Text Recognition in Graphics Images ATTR GI E total black pixels E black in E black out group D 140000 120000 100000 80000 60000 40000 20000 group B group A Figure 4 4 comparison between groups A B C and D according number of black pixels all in and out using second evaluation groupA E accuracy Merror 7 group B E accuracy W error 3 group C E accuracy W 11 group D E accuracy W error 10 Figure 4 5 comparison between groups A B C and D according accuracy and error ratio using second Arabic Typed Text Recognition in Graphics Images ATTR GI 120 00 100 00 80 00 E Group A 60 00 Group B 40 00 E Group C PE 0 8 E Group D 0 00 Figure 4 6 comparison between groups A B C and D according accuracy and error ratio using first evaluation Group A Group B E Macro accuracy W error E Macro accuracy W error Group C Group D E Macro accuracy W error E Macro accuracy W error Figure 4 7 comparison between groups A B C and D according accuracy and error ratio using first evaluation 69 Arabic Typed Text Recognition in Graphics Images ATTR GI 90 image correct absence missing unexpicted correct result Figure 4 8 compare between values of correct result correct absence unexpected and missing results for all 90 images 50 image correct absence missing
17. Avgmaxrow 5 If object_array fg mincol object_array fc mincol lt Avgmincol 6 Object_array fg 1 Object_array fc 1 7 Else 8 Label label 1 9 Object_array fg 1 label 10 End 11 Else 12 Label label 1 13 Object_array fg 1 label 14 End 15 End 16 Sort object array label sort all elements in object_array by label in ascending order 95 Arabic Typed Text Recognition in Graphics Images ATTR GI Pseudo Code for Algorithm 3 6 Localization process step 4 find minimum and maximum row and column after new label classification Input 1 ascending sort object_array by label use Pseudo Code for algorithm 3 5 Output 1 start_min_row vector is a vector that contains all minimum rows for after classification labels Output 2 start_min_col vector is a vector that contains all minimum columns after new classification labels Output 3 end_max_row vector is a vector that contains all maximum rows after new classification labels Output 4 end_max_col vector is a vector that contains all maximum columns after new classification labels 1 Searchforlabel 1 2 Px 1 3 While Searchforlabel lt z and Px lt z 4 start_min_row Searchforlabel object_array Px minrow 5 start_min_col Searchforlabel object_array Px mincol 6 end_max_row Searchforlabel object_array Px maxrow 7 end_max_col Searchforlabel object_array Px maxcol 8 For similar from Px 1
18. Vol 8 ISSN 1810 6366 85 108 2 Ahmed M E Mohamed A 2000 A Graph Based Segmentation and Feature Extraction Framework for Arabic Text Recognition 3 AIM 2001 Optical Character Recognition Technical paper A M Inc 634 Alpha Drive 2 10 4 Amin A S Fischer T Parkinson and R Shiu 1996 Fast Algorithm for Skew Detection IS amp T SPIE symp On Elec Image San Jose USA 65 76 5 Anthony C Mohammed B Neil B 2001 An Arabic optical character recognition system using Recognition Based Segmentation Pattern Recognition 34 215 233 6 El Mahallawy M S 2008 A LARGE SCALE HWM BASED OMNI FONT WRITTEN OCR SYSTEM FOR CURSIVE SCRIPTS Giza Egypt Faculty of Engineering Cairo University 7 Jain A K Yu B 1998 Automatic Text Location in images and video Frames Pattern Recognition Vol 31 No 12 2055 2076 8 Kareem D and Ossama 2007 Retrieving Arabic Printed Document a Survey 9 KASMIRAN J and MOHAMED A 2002 A SURVEY AND COMPARATIVE EVALUATION OF SELECTED OFF LINE ARABIC HANDWRITTEN CHARACTER RECOGNITION SYSTEMS Jurnal Teknologi 36 E Jurnal Teknologi 36 E Jun 2002 1 18 1 18 10 Keechul J Kwang I K Anil K J 2004 Text Information Extraction in Images and Video A Survey 1 35 11 Kim K C Byun H R Song Y J Choi Y W Chi S Y Kim K K Chung Y K 2004 Scene Text Extraction in Natural Scene Images using Hierarchical Feature C
19. Some of them use precision and recall rate Other people use detection rate and accuracy And some other can t be evaluated because the subject is very 74 Arabic Typed Text Recognition in Graphics Images ATTR GI difficult So we use precision and recall to compare our work result with other works And make correlation evaluation to make sure our result And we see the correlation rate is suitable evaluation in this subject area because may contain character word or string region which is make the rate absence of character equal the rate absence of string The experimental results for precision and recall are compared to work in 21 where their precision is 72 8 and their recall is 78 54 for 50 samples But in our work precision for 50 images is 86 90 and recall is 89 46 But we need to say here in 21 Works to extract Farsi text from video not from static image Our algorithm tries to extract the text from photo image and its work well As well as for different font type and size 75 Arabic Typed Text Recognition in Graphics Images ATTR GI Chapter Five Conclusion and Future work This chapter is the last one that presents our conclusion as a summary for all our work Then present our future work 5 1 Conclusion In this section we conclude our research by outlying our proposed method and summary its primary performance indicator and point out its limitation Our approach is applied on colored com
20. and this step might be done manually Second search for available OCR systems which can integrate with our module Candidate systems are e ABBYY is considered one of the top 5 programs in text recognition e Gimp a free and open source image manipulation program This requires a thorough analysis of these systems and a process of tailoring our module to integrate with the system of choice Our algorithm for extracting text written in horizontal way from colored images which prepared by PowerPoint Photoshop or any application capable to design an image passed through several variations before reaching the final goal In this research we propose a new technique for automatically detecting text in graphics documents and preparing them for OCR processing Our detection approach is based on finding regions in the input image with high density of text features When we started work in this research we thought to get these features by collecting information about angles lines and curves To help us for detection regions are subsequently processed to enhance the likelihood for successful character recognition by existing OCR techniques But through our research we could not find any method to give these text features 26 Arabic Typed Text Recognition in Graphics Images ATTR GI So our approach to solve the problem of recognizing text from colored images is depending on generates a set of rules based on comparisons between t
21. correlation correlation 1 correlation i correlation in 0 96 out 0 038 in 0 82 out 0 17 in 2 1020 93 6011 6 1820 84 00 15 in 0 84 out 0 15 17 16 18 19 74 3 i iawn lt alg pe correlation in 0 88 out 0 11 correlation in 0 99 out 0 004 correlation in 0 77 out 0 22 correlation in 1 out 0 correlation in 0 81 out 0 18 21 correlation in 0 97 out 0 023 correlation in 0 82 out 0 17 108 Arabic Typed Text Recognition in Graphics Images ATTR GI Table C 4 group D includes different style texts were written on picture background correlation in 0 99 009 correlation in 0 92 out 0 075 correlation in 0 79 0101 2 correlation in 0 88 0116 1 5 1 is Gant all 0 correlation in 0 98 6011 7 correlation 1820 86 6011 73 correlation in 1 out 0 correlation in 0 92 0116 1 correlation 1820 99 011 14 10 correlation 1921 out 0 11 correlation in 0 91 6011 9 correlation in 1 out 0 13 Sez sas 7 5
22. some researches on Arabic Chinese and Japanese character recognition 9 Arabic Typed Text Recognition in Graphics Images ATTR GI Effective text recognition techniques are widely used such as for indexing and retrieval of document images and understanding of text in pictorial images or videos 34 In other words text recognition allows translating images of text as scanned documents or images of natural scenes containing signs into actual text characters This is only true for clean non distorted images 18 From above definitions there is no specific topic called text recognition as itself Searching on text recognition is almost take results on optical character recognition with little text recognition That means the two titles are the same meaning and contains the same methodology to recognize finally characters in text image image that already contains text The text recognition step is used as text localization to know the first location of the text only assuming the text exists in the image Then complete the rest of recognition which are separation segmentation then finally character recognition or word recognition Most researches in text recognition or OCR depends on comparing between set of words characters respectively to identify the word or character in the scanned paper or graph image but not necessary to know if this region is a text before recognized it So it takes a lot of times for comparing process and if
23. 3 2 Localization process step 1 Collect information Input Binary image BW I Output 1 localized each item in I via dashed rectangle shape Output 2 ascending sort object_array by minimum row 1 bw3 is a computation distance of BW 2 find the labels for each elements in bw3 3 Put any other elements in the image like background to zeros 4 Define 2D object_array with six items to save element number maximum row maximum column minimum row minimum column and last one to save new label 5 Draw rectangle shape with these parameter Mnc Mrr width height line width line style 6 Sort 2D object_array in ascending order depending on min_row 38 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 3 compute difference and mean for minimum row Input 1 ascending sort object_array by minimum row use algorithm 3 2 Input 2 z number of items Input 3 kind variable it is maybe max or min row or col Output 1 mean for input 1 Output 2 difference_array1 vector for input1 1 Define Difference__array by length z 2 Each index in Difference_array contains the value of difference between two adjacent items in one column of max min row col in object array 3 Now compute summation for all items in Difference_array vector 4 Find the average and get the floor value 39 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 4 Localization process step 2 give label f
24. Ais a pointer in the first item ii Bis another pointer to the next item iii Check the difference of maximum row between the two pointers is less than or equal the average of maximum row then 1 Check the difference of minimum column between the two pointers is less than or equal the average of minimum column then 2 B get the same label of A 3 Else increment the label value 4 Label of B equal new label value iv Else increment the label value v Label of B equal new label value 2 Sort object array label gt sort all elements in object_array by label in ascending order 41 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 6 Localization process step 4 find minimum and maximum row and column after new label classification Input 1 ascending sort object_array by label use algorithm 3 5 Output 1 start_min_row vector is a vector that contains all minimum rows for after classification labels Output 2 start_min_col vector is a vector that contains all minimum columns after new classification labels Output 3 end_max_row vector is a vector that contains all maximum rows after new classification labels Output 4 end_max_col vector is a vector that contains all maximum columns after new classification labels 1 Each vector in output 1 2 3 and 4 initialized with values of first item in 2D object_array 2 Make a search for each label to find minimum and maximum row and column to ident
25. Base 1 For mp 2 to imx1 If maxm lt full info img mp 3 maxm full info img mp 3 Base mp end End oe Pseudo Code for Algorithm 3 10 find lengths of adjacent sequence black pixels in Base line Input Base line for BW3 region of Binary image crop use Pseudo Code for algorithm 3 9 Output descending Bline vector this vector contains length of adjacent sequence black pixels in base line 1 zx zy size BW3 2 x 1 3 For m from 1 to zy 4 Bline m O 99 Arabic Typed Text Recognition in Graphics Images ATTR GI 5 End 6 sumb 1 7 For L from 1 to zy 1 8 If BW3 Base L 0 and BW3 Base L 1 0 9 sumb sumb 1 10 End 11 If BW3 Base L 0 and BW3 Base L 1 1 12 Bline x sumb 13 x x 1 14 Sumb 1 15 End 16 End 17 Descending sort for Bline Pseudo Code for Algorithm 3 11 find accumulator for lengths of adjacent sequence black pixels in Base line Input 1 descending Bline vector this vector contains length of adjacent sequence black pixels in base line use Pseudo Code for algorithm 3 10 Input 2 x value where x is length of Bline vector Output accum x 2 is a two dimension accumulator first column is the length value and the second column is how many times this value redundant in Bline vector 1 accum x 2 0 2 bc 1 3 b 1 4 sumc 1 5 Bfinish 0 6 For m from 1 to x 7 accum m 1 0 8 accum m 2 0 9 End
26. TP TN Accuracy Equation 5 TP TN FP FN FP FN Error Equation 6 TP TN FP FN Any known measure for binary evaluation can be used here such as accuracy precision and recall The calculation of these measures for all labels can be achieved using two averaging operations called macro averaging and micro averaging These operations are usually considered for averaging precision recall and their harmonic mean F measure in information retrieval tasks The macro averaged and micro averaged versions of B TPA FPA TNA FNA for label A are calculated as follows Bmaro 1 q YJ_ B TPATNA FPA FNA and Bmicro B XL TPA XL FPA D TNA YL FNA 30 Therefore the additional criterion measures are Micro Precision Micro Recall Macro Precision and Macro Recall So we can explain it as follows Micro average precision is tpi YL tit DL fpi Equation 7 59 Arabic Typed Text Recognition in Graphics Images ATTR GI Micro average Recall is Xj tpi DL tpi XL fni Equation 8 Macro average Precision is q 5 Zani Equation 9 Macro average Recall is q 5 Heat Equation 10 Where Pi is the precision for each element and Ri is the recall for each element 60 Arabic Typed Text Recognition in Graphics Images ATTR GI 7 The evaluation steps that used in first evaluation technique 1 After applying text region segmentation algorithm some words are still has alone segmentation
27. Text candidate extraction Combining Select verifying region Stroke features and SVM Verification High level Text feature Skew Predictive variation 1 Image enhancement Text region Figure 2 1 the proposed method for Kim et al in 11 Michael et al in 18 develop a framework based on identifying the text place automatically in the image by using Maximally Stable Extremal Regions MSER detection 16 This is the best way that interest in Region identification in computer vision according evaluation in 20 This method depends on Continuous geometric transformations and is invariant to affine intensity changes Farther more identify the text in different size The new approach for them is depending on increase the accuracy for word recognition with 87 05 instead of 68 58 which based on ICDAR 2005 So they connected with web search engine like Google and Yahoo This way is automatic but need a large number of words to give a correct recognition But in our thesis we based on the external features to recognize the text Michael et al in 19 performs a detailed analysis of multilingual text characteristics including English and Chinese They work on a comprehensive efficient video text detection localization and extraction method Text detection based on edge detection local thresholding and hysteresis edge recovery Text localization performed to identify text reg
28. algorithm couldn t detect it 8 If we make our conditions more restriction we give a results are absolutely Arabic with little uncorrected results But the restricted conditions effect on the ratio of retrieved Arabic text Therefore we make our conditions have little flexibility to retrieve the highest amount of Arabic texts with some uncorrected and absence results was observed 9 Our algorithm works correctly on a small size font The font size is approximately from eight to thirty Also works on bigger than thirty but maybe lost some of them We think that back to the font type if it is normal font with simple format or it has word art format We look that if the text has word art format the ratio of missing is very big 10 From experiment if the image contains only Arabic text without any English text and has clear background color We mean in clear background the background has one color not picture and there is no part of pictures near to the text then the segmentation process works 72 11 12 13 14 Arabic Typed Text Recognition in Graphics Images ATTR GI successfully Because the value of distance is consider as a factor in the segmentation process If all factors are available then recall precision and accuracy are equal to one From our experience in this area If we relied on our work on edge map the segmentation process works successfully one hundred percent Because the high intensity gives to
29. and High level Low level contains three image features that used 1 to find local variation of intensity 2 colors and 3 to find color continuity of the same text area High level is used to verify the candidate text region by examining stroke composition Finally verify character recognition as shown in figure 2 1 Also used SVM that works on variant size for verification by input all previous features to SVM to classify the local area is text or non text Their method is just for extracts and verifies longest text regions before the final text recognition with OCR They found that the average of results be increased in the colored image to become 88 6 But text image become 85 5 also test their model on Camera but did not work efficiently So they need to improve their model to work in a proper way with Camera But not experiment on Arabic Text This emphasizes the process of integration might be done and this technique suitable for special purpose and might not give a good results in scanning images 11 Arabic Typed Text Recognition in Graphics Images ATTR GI Scanning Image Intensity variation Color variation Color Continuity ee Edge detection HorL VerL color variation Color Reduction Image features 1 Long line detection AND Operation Color Clustering Text candidate extraction Text candidate extraction
30. as word segmentation and the most formed text region But little words or statements do not segment from the beginning of using the connected label method So we count correct text and word segment region by count Arabic text region one time and other time just count the Arabic word region for each image Then count the correct visible statements words in output result as correct result TP We do not count the statements or words are not labeled from the beginning either Arabic or English after using the connected label method Also don t regard to paragraph region Because our focus on statement region only But if Arabic paragraph region are presenting in output result we are counting as correct result TP Some items in the origins image represent in the output result but it should not be represented like English text or shapes So we are counting as unexpected output FP The selected statements or pictures or shapes are correct absence from the output result we are counting as correct absence TN Some items are selecting from the beginning but some rules incorrect prevent it from appearance in the output result and we are counting as missing FN Then we are counting the correct results missing results unexpected results and correct absence results for each image to compute precision recall and accuracy for each image Then compute micro and macro precisions and recall for all images see table 4 9 and 4 10 61 Arab
31. automatic text word segmentation and solves many problems in colored image until reach the final goal This goal is creating filter to extract text form colored images The second section presents the implementation structure for the main functions used to build the filter But before discussing these sections we will present a research design and methodology to get brief details 3 1 Research Methodology Our research will be composed of three main stages first define most suitable set of features for Arabic Second we will develop our Arabic text recognition model Third evaluate our newly developed model and integrate it into an OCR system to evaluate its performance within the overall OCR process These stages and their details are depicted in figure 4 and are further explained in the following two subsections Stage 1 Define Arabic features Text Detection Post processing Evaluate above stage System Integration Figure 3 1 the model for research design 24 Arabic Typed Text Recognition in Graphics Images ATTR GI Stage 1 Define most suitable set of features for Arabic In this stage we try the most suitable features for Arabic Like 1 seeking process must be done from right to left 2 base line to identify the orientation of Arabic text 3 to decide if this region is an Arabic text or not may use some methods to calculate the number of corners vertical horizontal and curved lines Stage 2 Ar
32. background filled with one color rather than white aha 107 Table C 3 group C includes different style texts were written on background filled with inves enn eda een 108 tied variant eradient COlOM si Table C 4 group D includes different style texts were written on picture background gio E E O E E E CD care taupe eueeeneeeamne gees 109 Table C 5 set of results images for SHOW ccccccecccecececececesececceceeececeteeeteeseeeseeetees 111 Arabic Typed Text Recognition in Graphics Images ATTR GI Chapter 1 Introduction This chapter contains nine sections We will declare some definitions for abbreviations or terms that are related to our work starting from OCR definition until we even got a way to draw the appropriate definition of AOTR Then we display problem statement objective scope and limitation significance of thesis and research format respectively 1 1 Optical Character Recognition Optical Character Recognition abbreviated OCR defined in 44 and 49 as the process to convert the scanning image of papers to new documents can be easily used with computers for manipulations This process can dealing with handwritten or printed characters Also it is useful for commercial and education to simplify for getting information like credit card in the Banks or shops OCR used in two ways according the components of
33. containing homogeneous text lines of the same orientation is processed for text recognition However this zoning approach cannot handle documents that don t have homogeneous text Arabic Typed Text Recognition in Graphics Images ATTR GI line such as artistic documents pictorial image with text raster maps and engineering drawings In 10 Keechul et al defined Text Information Extraction TIE as a process might be used in set of scenes or in fixed image It goes through the following phases 1 Text detection to determine the presence of text in sequence of images 2 Text localization to determine the location of text in the image and generating bounding boxes around the text 3 Text tracking used to reduce the processing time for text localization and to maintain the integrity of position across adjacent frames 4 Text extraction is the stage where the text components are segmented from the background 5 Enhancement the text separated from image required an enhancement because the text region usually has low resolution and is prone to noise After that the text was extracted ready for using in OCR technology as shown in figure 1 1 images or video clips Text Detection Text Localization Text Tracking Text Extraction and Enhancement 1 Recognition OCR Text Figure 1 1 the architecture of TIE system In chapter related works we show that there is a significa
34. crop Output full info matrix is a Tow dimension array with three column find the size for each crop object b count black pixel in each row w count white pixel in each row save row number into first column of full info matrix save b into full info matrix into second column SY N FP save w into full info matrix into third column 44 Arabic Typed Text Recognition in Graphics Images ATTR GI The above mentioned algorithm keeps information for all regions of crop images which are stored in full info matrix array Now we will illustrate the Calculation Arabic Base Line algorithm This algorithm based on make a comparison between the amounts of black pixels for each row in the sub image The amounts of black pixels saved in column three in full info img array Assume the first value in full info img is a maximum value and name it maxm Also assume Base variable is the first row Then start the search process from the second row to last row If maxm less than the value of black pixels in the next row change the maxm value to the new value of the next row see algorithm 3 9 45 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 9 Calculation Arabic Base Line Input two dimension array full info img use algorithm 3 8 Output The Base Line Number 1 make a search to find row that contains maximum black pixels 2 compute half line 3 check if this row is below the half line Alg
35. postprocessing 0 method 52 Arabic Typed Text Recognition in Graphics Images ATTR GI 3 2 3 Post Processing In this section we have two post processing depending on the output result One of them for OCR system with white background and black foreground but size differs from image to other The other output is for users to be easy to show the correct text recognition The text marked with red color on the RGB image with only black and white colors as a copy from the RGB image 3 2 3 1 Result Image for OCR After finish the step of text detection the output image has some items shapes and English text should be not appear on the result So we decide to take the result and inserts on more process as a trial step to give OCR system a clear image result as much as possible Thus for this we take the output image 12 from algorithm 3 2 2 3 2 1 Then make segmentation on thinning image with big distance to select long text region Then put some rules that calculate the ratio of black pixels to white pixels We have to because ratio of black pixels after thinning in Arabic text is less than white pixels in the shapes and English text These ratios are selecting by manual experiments give good results But we go in a close end because if put excess conditions maybe lose some items we need it to appear See algorithm 3 15 53 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 15 OCR Post Processing Inp
36. segmentation process to identify the localization of objects 3 2 2 Text Detection This section is the most important point in our research and it refers to the primary aim to reach final goal Three steps are discussed namely text localization collect information and text type recognition 29 Arabic Typed Text Recognition in Graphics Images ATTR GI 3 2 2 1 Text Localization Method This section depends on segmented the image and extract all items that are included after converted the image from RGB to binary then identify the localization of Arabic text It is important to say this step is most difficult because it takes a lot of time and variant trials to find a way for solving this problem In the beginning we thought to create a two dimensions array that holds information for the entire binary image The information as follows line number number of white pixels and number of black pixels for each line Then we used this information as input data in order to enable us to deal with image So we apply three stages Stage 1 the first one is segment the image horizontally row by row considering that the text was written in horizontal way A long horizontal white line is calculated i e number of white pixels equal image width But this way faces a problem which is the segmented image might include more than one item like picture with another picture or picture with text the text here could be an English or
37. starting and ending points of the character and usually facilitate writing separate characters not complete words But off line OCR deals with scanned images 1 2 Optical Text Recognition Optical Text Recognition or Text Recognition abbreviated OTR TR and OCR both of them are an active area in both academic research and commercial software development 34 Farther more shared with general steps for recognition process like 1 preprocessing prepare the image to be a binary image not colored image to make the work on image is easier 2 text extraction that s mean get the pixels that performs character text to be an independent part from the rest of the image 3 segmentation if we have a long text with white spaces then segment each word or sub word as a single image 4 separation in a single word that does not have a white space make each character as an individual character in a single image and 5 character recognition identify the character by comparing it with a set of character for recognizing process But in OTR focus on the steps before character recognition which that s mean the steps from 1 to 3 In other words OTR based on recognition of text at first by significant the text region with rectangle shape If the development completes the process with steps 4 and 5 then they are going to recognize separate characters and this is OCR Most researches in OCR or OTR have been on Latin base characters with
38. to z 9 If object_array similar 1 Searchforlabel 10 Px Px 1 11 If object_array similar minrow lt start_min_row Searchforlabel 12 start_min_row Searchforlabel object_array similar minrow 13 End 14 If object_array similar mincol lt start_min_col Searchforlabel 96 Arabic Typed Text Recognition in Graphics Images ATTR GI 15 start_min_col Searchforlabel object_array similar mincol 16 End 17 If object_array similar minrow lt end_min_row Searchforlabel 18 end_max_row Searchforlabel object_array similar maxrow 19 End 20 If object_array similar minrow start_min_row Searchforlabel 21 end_max_col Searchforlabel object_array similar maxcol 22 End 23 End 24 End 25 Searchforlabel Searchforlabel 1 26 Px Px 1 27 End Pseudo Code for Algorithm 3 7 Localization process step 5 select text region Input 1 start_min_row vector is a vector that contains all minimum rows for after classification labels use Pseudo Code for algorithm 3 6 Input 2 start_min_col vector is a vector that contains all minimum columns after new classification labels use Pseudo Code for algorithm 3 6 Input 3 end_max_row vector is a vector that contains all maximum rows after new classification labels use Pseudo Code for algorithm 3 6 Input 4 end_max_col vector is a vector that contains all maximum columns after new classification labels
39. unexpicted correct result Figure 4 9 compare between values of correct result correct absence unexpected and missing results for first of 50 images 70 Arabic Typed Text Recognition in Graphics Images ATTR GI We experiment our work on ninety images These images are collected from web pages like Google search and Facebook in random way We are select images have a variant size font 4 2 Discussion We have some observations on the algorithm works and its results especially on micro recall micro precision correct result missing unexpected and correct absence values Our observations are representing as follows 1 From figure 4 2 and 4 4 you can see that by second evaluation the amount of black pixels is should equal to the actual black pixels that should be represents in the output result are very close And the rate of missing and unexpected black pixels is very small Also from figure 4 3 and 4 5 you can see that by second evaluation accuracy is around 90 and the rate of error is under 11 This means our proposed solution is deal will and can discover Arabic base text in a proper way 2 From table 4 5 we can see the maximum accuracy is in group B because if you look to the pictures in group B you can observe that the font style in each image is mostly the same This means that if the image contains Arabic text with the same font type and font size the success ratio will be increase 3 The Clear
40. z bw label bw3 Note L1 is the label number and 2 is the total number of all labels 3 Put any other elements in the image like background to zeros 92 Arabic Typed Text Recognition in Graphics Images ATTR GI 4 For j 1 to z 5 r c find L1 j 6 Object array j 1 0 7 Object_array j 2 j 8 Object_array j 3 Maximum_row 9 Object_array j 4 Minimum_row 10 Object_array j 5 Maximum_column 11 Object_array j 6 Minimum_column 12 Width object_array jjj 5 object_array jjj 6 13 Height object_array jjj 3 object_array jjj 4 14 If width gt 0 and height gt 0 15 Draw rectangle shape with these parameter Mnc Mnr width height line width line style 16 End 17 End 18 Sort object_array in ascending order depending on min_row Pseudo Code for Algorithm 3 3 compute difference and mean Input 1 ascending sort object_array by minimum row use Pseudo Code for algorithm 3 2 Input 2 z number of items Input 3 kind variable it maybe max or min row or col Output 1 mean for input 1 Output 2 difference_array for input1 1 Difference__array z 0 2 sum 0 3 For ost from 2 to z z is the number of items 4 Difference_array ost object_array ost kind object_array ost 1 kind 5 sum sum Difference_array ost 93 Arabic Typed Text Recognition in Graphics Images ATTR GI 6 End 7 Avg L mean Difference_array Pseudo Code for Algorithm 3 4 Localization p
41. 2222220 0 ns 68 Figure 4 5 comparison between groups A B C and D according accuracy and error ratio USING SECON CVAalUAtTION c cccceccecceeceeceeceeceueseesueceeceeseeseseusaesseeseeseeseeseesensaesaeeaenes 68 Figure 4 6 comparison between groups A B C and D according accuracy and error ratio USING first CVAlUATION reiner sesi asset a nex aa ala 69 Figure 4 7 comparison between groups A B C and D according accuracy and error ratio using first CValUatiON seas SRSA eve weiwexdeiciiee a 69 Figure 4 8 compare between values of correct result correct absence unexpected and missing results for all 90 images cccccceccecceeeeeceeceeceeceeseeseueaeesueceeseeseesuesensgesaeesenes 70 Figure 4 9 compare between values of correct result correct absence unexpected and missing results for first Of 50 0 0060 5 0 0606 20388 70 Figure 4 10 Arabic text with parts of characters wrote in above and below the base line a Has characters Geem gives long height b All characters under the base line have the same height c All characters wrote on the base 110612222215 74 Figure 4 11 English text with parts of
42. 3 Swap between black and white in I4 Now apply thicken skeleton on I4 Swap again between black and white in 14 rgb Convert 14 image to RGB image I2 Convert I2 image to RGB image PN Se oS VN Change foreground color in 14 to red color 10 Img multiply 12 with rgb This is our idea to solve the problem of extracting Arabic text images from the colored background image for more details refer to Appendix B 3 3 Implementation Our work is implemented by MATLAB 2008 Our functions are clearGraphic preprocessing and postprocessing The clearGraphic is the base filter that use preprocessing and postprocessing functions and returns a black red white image as shown in figure 3 11 Other functions have one image parameter and contain set of steps to give an output binary image 55 Arabic Typed Text Recognition in Graphics Images ATTR GI 12 is a thinning binary image but contains text with some unwanted data which needs more processes So 12 pass throw the postprocessing and return 13 After that clearGraphic take 13 where 13 is a thinning black white image after removing a large amount of unwanted data Then take I3 and apply thickens method three times to be bold as much as possible and then convert it to RGB image to give red color for foreground and background still white Additionally 14 is defined to be a binary black white copy from the origin image then convert it to colored black white ima
43. 300 dpi The average of correct script classification rates were 95 2 and 94 1 respectively The average correct recognition rates were 94 8 and 88 9 respectively Another experiment was performed on pages were printed by laser printer with 300 dpi The average of correct script classification rate was 94 6 and the average correct recognition rate was 93 4 but they graphic images are not included in their experiments Anthony et al in 5 proposed a serious problem on Arabic recognition that is segmentation This problem comes from the Arabic text written with overlapping rather than the other language which come separate when written by computer as a printed character But there are 28 Arabic characters where each character can be used between two to four ways i e the shape of character in the beginning of text is differing from the middle or in the last So the structure they used for Arabic OCR system is at first scan the Arabic documents from right to left Second give an image acquisition Third preprocessing which used Binarization to convert image to black and white which make the process on image is easy because it does not cost a lot of measurements and easy to remove noise In addition smoothing is used to fill the gap between pixels that lost colored word segmentation character fragmentation combination feature extraction classification call it feedback loop finally recognition results and user interface The accu
44. Arabic Whatever go to second stage Stage 2 we developed a code to segment a sub image vertically to separate between items in one sub image By saved an information for each sub image separately but this time read an image column by column to have a two dimension array includes column number number of white pixels and number of black pixels To find long vertical white pixels equal to high of sub image But another time this way face a problem with sub image includes more than one item have an interlacing between of them the interlacing means that between the two items there are no white column to allow us to apply vertical separation process So we go to next stage Stage 3 a connected label algorithm is used to label close items with same number depending on identify a small ratio of distance To be easy to know the coordinates of each item and then separated as rectangle form The value of distance is chosen by select and test different values until we found the 30 Arabic Typed Text Recognition in Graphics Images ATTR GI nearest value for extracting process which gives a nearest connected item in one line If minimize the distance value as much as possible can get each item alone according the long of the space between items So we found a bwdist method to identify the distance value and choose it to be four and bwlabel method which helps us to separate between items in the same sub image based on that dist
45. Cyrillic Azerbaijani Latin Bashkir Dictionary support IRacn e 000000000 Figure A 7 supported language window When tried Quick tasks which is the second process in ABBYY FineReader we test some examples and include three samples with different situation and found an excellent extraction for English text from image too There are three cases Case one if image contained Arabic and English with 85 Arabic Typed Text Recognition in Graphics Images ATTR GI white background a conformation message appear because the resolution of image is less than 400 dpi and the program found characters cannot be recognized because the data set for Arabic characters does not include in the system to use it for character reorganization but detect an English characters or numbers see figure A 10 A 13 So the result will appear numbers and characters of English but garbage for Arabic characters or numbers The second case if the image containing only Arabic text the system cannot be processed and present a confirmation message it does not detect any character So the result is an image as you see in figure A 14 A 16 The third case if image contained Arabic text with format style the program extracts it as picture see figure A 13 A 19 Set Program Access and Defaults ee Windows Catalog Windows Update Accessories 5 Microsoft office rie i EG Documents eDocFile LEAD
46. For best results set the resolution to 400 Figure A 18 conformation message because resolution is less than 400 Document5 Microsoft Word 2 i Home Inset Pagelayout References Mailings Review View ABBYY FineReader11 Acrobat amp cut z lt FFE 5 4A Fina Calibri 14s 3 3 aanbcc AaBbC AaBbce AAB 4 6 2 9 3 a a AEN Farai B 72 U be x x Aa HESE T No Spaci Heading1 Heading2 Title Subtitle Subtle Em rie EEE 1 Clipboard G Editing uljlftuo 1C MARKETING PLANNING Fage t of2 Words 52 B English U S EEE TTS Figure A 19 result for third example take in figure A 15 91 Arabic Typed Text Recognition in Graphics Images ATTR GI Appendix B Pseudo Code In Appendix B we represent a pseudo code in MATLAB for each algorithm in chapter three Pseudo Code for Algorithm 3 1 threshold for edge map Input colored image I Output binary image 1 X grayscale image I S wiener2 X K rangfilt S e N Level graythresh K 5 BW convert grayscale image to black white with K Level Pseudo Code for Algorithm 3 2 Localization process step 1 Collect information Input Binary image BW Output 1 localized each item in I via dashed rectangle shape Output 2 ascending sort object_array by minimum row 1 bw3 bw distance BW 2 L1
47. Graphic filter gives high accuracy if the complex image includes text with clear distance between lines so it does not have word close and the words in the same line have the same style Thus the ClearGraphic filter is working correctly on a single statement and sometimes on a single word 4 The ClearGraphic filter works on different style of images therefore the text should be clear and the interior filters could discover the text 71 Arabic Typed Text Recognition in Graphics Images ATTR GI 5 Has been deleted group of images because the actual output should be a clear image without any black pixels so the accuracy and error for it is NAN 6 Sometimes the correct output could be 50 or less And unexpected output more than 50 ffor more than one reason First one is the image include different font size specially 12 or less with 40 or more So the distance factor in ClearGraphic filter should be affected and the font with small size only shows response Another reason after applying interior filter there are some connection between lines This connection makes two lines as one line and in this case we cannot detect if it is a text or not 7 We found ClearGraphic filter can detect some text wrote in diagonal rotation because the rotation angle is not large But does not work on text write by word art because after using the interior filters the words have outline border without filling the inside for this reason our
48. ITES PS 0592331117 00970592331117 Page t 0 1 Words 21 lt 35 English U S EO 2 ue Figure A 13 the result for first example used in figure A 8 88 Arabic Typed Text Recognition in Graphics Images ATTR GI Open Image lt Look in images 0 mG _1232012537524M jpg IMG_1232012537524Maa jpg My Recent E pali jpg Documents pal jpg N untitled2 bmp N untitled3 bmp Desktop O untitled4 bmp untitledS bmp 5 N untitled bmp My Documents My Computer File name untitled5 bmp My Network Files of type Common Image Files Places Page range All Pages V Detect page orientation V Enable image preprocessing M Split Facing pages Figure A 14 the second example containing only Arabic text with background color Open and convert images PDF file Opening the document Opening the document in Microsoft Word File C Documents and Settings lalsaedi Desktop images images untitled5 bmp Al page 1 The resolution of the source image is too small and the image has been stretched 4 Page 1 There are no objects to recognize on this page no text table or barcode areas have been detected Figure A 15 conformation message represent the program can not detect any object to recognize it 89 Arabic Typed Text Recognition in Graphics Images ATTR GI vo D
49. Input 6 val the value in the second column of accum array Input 7 C the position in accum that contain maximum peak value Input 8 start_min_row start_min_col end_max_row end_max_col 51 Arabic Typed Text Recognition in Graphics Images ATTR GI vectors these vectors contains the origin coordinates from the origin image Input 9 BW3 thinning image crop Input 10 12 blank BW image with white background and has the same size of original RGB image Output I2 output result image this include Arabic text with some OMNI Bw Mm 15 16 17 unwanted data Apply algorithm 3 7 select text region to return obj start_min_row start_min_col end_max_row and end_max_col Convert white background to black BW3 thinning to image_crop Farea Find foreground area Imagearea Find image_crop area width height Objarea Farea Imagearea Apply algorithm 3 13 to find halfline and ratio Apply algorithm 3 11 to find val and C Size I2 size I I is the origin image 12 1 gt ie 72 has white background if Objarea gt 0 01 and Objarea lt 0 16 if Base gt halfine and Base lt hafline ratio if accum C 1 gt 1 and val gt 3 Copy the segmentation part from the region image crop to its position in 12 output image End End End All of the previous algorithms are integrated into one method namely preprocessing Now in the next section we will describe the
50. Islamic University of Gaza Faculty of Information Technology Deanery of Postgraduate Studies Arabic Typed Text Recognition in Graphics Images ATTR GD This thesis Submitted to the Faculty of Information Technology Islamic University of Gaza In Partial Fulfillment of the Requirements for the Degree of Master of Information Technology Prepared By Lamiya Mohmmed El_Saedi Supervisor Dr Ashraf Alattar September 2013 Arabic Typed Text Recognition in Graphics Images ATTR GI SALON PD 3 gt 20 e e Ane AWS Arabic Typed Text Recognition in Graphics Images ATTR GI Pedieation I dedicate this research To the spirit of my dear mother Asma a Syam To my dear father Mohmmed El Saedi To all my family and all my friends To everybody who prayed for me To my university To my our colleagues I dedicate this research Arabic Typed Text Recognition in Graphics Images ATTR GI Acknowlgdgment First of all I would like to thank Allah for help me and get me a long patient to complete this research Secondly I would like to thank my supervisor Ashraf AL_Attar because he is very challenger to exit this research in this picture I would like to thank my husband Ashraf Salama and my child because they suffered with me so much I would li
51. TOOLS Main EVAL 17 5 gt E settings g gm ABBYY FineReader 11 W Quick Tasks gt File PDF Image to Microsoft Word po Search 0 ABBYY Screenshot Reader Photo to Microsoft Word 2 Scan and Save Image Heb and support G scan to Microsoft Word E Run amp Scan to Searchable PDF g Shut Down rabia Figure A 8 to run ABBYY Quick Tasks ABBYY FineReader 11 3 Welcome to ABBYY FineReader 11 Professional Edition 2 days and 47 pages remaining before the trial period For this copy of ABBYY FineReader 11 expires You can save only 1 page s at a time Please buy and activate a license to start using the product in Full mode Activate License and Switch to Full Mode Serial Number Required Figure A 9 this window appear when you choose Quick Tasks in figure A 6 86 Arabic Typed Text Recognition in Graphics Images ATTR GI Open Image Look in images 0 2 2 EM Show preview P IMG_1232012537524M jpg 1 16 1232012537528132 My Recent jpg Documents untitled2 bmp N untitled3 bmp N untitled4 bmp untitledS bmp jaa N untitled bmp WWW SIMPLESITES PS 0592331117 20970592331117 1 1 File name pal jpg My Network Files of type Common Image Files Places Page range All Pages Detect page orientat
52. abic Text Recognition The Arabic text recognition stage encompasses three steps 1 pre processing 2 text detection stage and 3 post processing These stages are explained as follows Pre processing In this stage the following pre processing operations are performed on the input image to improve the detection efficiency e Extract intensity from color image e Intensity enhancement e Noise removal e Thresholding binarization Text Detection This stage contains the core processing operations that we will develop e The outcome of this stage is to localize the sub regions which contain Arabic text Localization can be given as a bounding box counter or centroid e Our technique s performance will be evaluated visually based on the percentage of text regions it is successfully able to detect Also make a correlation function that checks how much the output result and the exact output are equals 25 Arabic Typed Text Recognition in Graphics Images ATTR GI Post processing e Clean extracted text sub image from any remaining background clutter e Rescale and resize the image to be suitable for OCR segmentation segmenting words into sub words and recognition recognizing each letter separately Stage 3 Evaluate integrating our module into other applications This stage involves two steps first evaluate our model by showing the rate of correct text localization from all the contents of graphics image
53. age e A reusable component that can be used with other applications or devices which depends on extracting textual content from images and we can call it Friendly OCR e Enhance search engine capabilities in dealing with images o Find images which contain specific words o Find images which contain alias words o Index document which contain images e Make dealing with softcopy easier such as reading copy cut delete and print e OCR is already being used widely in the legal profession where searches that once required hours or days can now be accomplished in a few seconds 49 1 9 Research Format The research is organized as follows Chapter one is introduction research problem objectives scope and significant Chapter two is related work Chapter three is technique and implementation Chapter four is evaluation and discussion Chapter five presents conclusions and future work 10 Arabic Typed Text Recognition in Graphics Images ATTR GI Chapter 2 Background and State of the Arts In this section we review a number of research works in OCR OTR We start with non Arabic works in order to cover the techniques used to deal with the general challenges and then move on to some Arabic research works to cover techniques which handle the specific problems related to Arabic text 2 1 Non Arabic OCR OTR research work Kim et al 11 combined two different levels of text features These two levels are Low level
54. aining the rules as much as possible We have some restrictions in our work The first one is the height of sub image does not exceed than seventy five pixels to ignore the largest shape This restriction is used from the beginning before saving the items in an object array to establish new label The second restriction is using fixed number in the conditions These numbers identify by manual experiment After that we minimize as much as possible the amount of region images that not related to text Now we put rules that related to exclude the Arabic region images from the rest of regions as 1 compute the ratio of foreground area to the total area of region to give small semi identical ratio to prevent display graphics on the output result 2 find the position of Base line the position of maximum peak and its value to prevent display English text in the output result The output result is an empty image has the same size of the original image but with white background The final result is formed to be an input to OCR system See algorithm 3 14 Algorithm 3 14 Set of Rules Input 1 objarea is the ratio of foreground area to the region area Input 2 Base the value of base line Input 3 halfline the value of half line Input 4 ratio the range could be found the base line below the half line Input 5 accum two dimension array includes in first column the times of strait line of black pixels in the base line are redundant
55. amining if the item is text or not In spite of this when we use the connected label method without any addition stage that we talked about it gives most of sub items that are located in the entire origin image So we were forced to use it more than one time to get each item alone We have to because this method works only when the background is black So for this reason there are two cases The first one is when we convert the entire image to binary image all pixels in the background is white i e contains ones So we need altered to zeros to able to put labels for each item in that object The second case is after convert to binary image there are variations of background between black and white Therefore the sub object that has white background does not work with the connected label method in a proper manner So we exchange the background to be black then again use connected label This led to effect of the execution speed to become slower To explain the reason for using connected label method twice we will strike an example for more declaration Look at figure 3 6 picture with label 1 is the original image with colored background Picture with label 2 is black and white image Picture with label 3 is a centroid image which illustrates the blue star in the middle This 32 Arabic Typed Text Recognition in Graphics Images ATTR GI meaning all the contents is one item And picture with label 4 is a result of applyi
56. ance value by giving a separate number for each connected item After that we use the labels for each item to take the minimum and maximum coordinates row and column x1 y1 and x2 y2 to identify a rectangle or square shapes around the object As an example see figure 3 2 3 3 3 4 and 3 5 Figure 3 2 example to show how identify a rectangle shape over an items 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Figure 3 3 circle item with label one NINN IN N NINN INN NINN INN NINN IN N NINN IN N NINN IN Figure 3 4 rectangle item with label two 31 Arabic Typed Text Recognition in Graphics Images ATTR GI 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 J3 Figure 3 5 triangle item with label three According try the above stages in that order in shuffle with each other or using connected label method alone The number of segmented items with three stages is better than using connected label alone For example if the image contains a border and the item is very close to that border when we increase the value of distance in a small amount it is effect on separate process and regards them as a single item thus minimize the number of sub images that will enter to final stage which ex
57. and type of text All of these things are obstacles in the identification of the text and added to above challenges as shown depicted in table 1 1 Table 1 1 A comparison between Scand pages and graphics advertisement me pages Graphics ads 2 Start Fonts 1 Type restrictions Yes mainly black Background 101 color 2 Noise a elements in background Arabic Typed Text Recognition in Graphics Images ATTR GI 1 5 Problem Statement The problem of this research is how to increase the performance and accuracy of identifying Arabic text from graphics image The process must be done automatic without select the text in graphics image 1 6 Objectives Our objective is divided to main and specific objectives Specific objectives are much closed to main objective and give more details for our main objective 1 6 1 Main Objective The main goal of this research is to develop a method for filtering detecting and extracting Arabic text from graphics images The output of the method is a white background image that contains black Arabic text 1 6 2 Specific Objectives e Collecting suitable data that contains Arabic and Latin based text with colored background and different shapes to improve our idea and to solve the problem of text extraction from complex image e Select suitable features to be available for maintenance e Develop the method to Extract Arabic text from gr
58. aphics image e Evaluate our module by calculate the rate of text that recognize correctly e Integrate our method in an open source Text Information Extraction system 1 7 Scope and Limitation e Cover images with both clear as well as cluttered backgrounds e Covers horizontal text orientation only e Will not include character recognition only post process in order for OCR algorithms to handle this part e Our module is off line system Arabic Typed Text Recognition in Graphics Images ATTR GI e Will not include handwritten characters e Will not include English e Will not include region that contain more than one statement e Integration will be limited to either OCR system 1 8 Significance of the thesis e The set of features extracted about the text region can be utilized to reconstruct the text for editing purposes e The algorithm could be used as additional properties to the printer as helper to decrease the amount of ink via remove the pictures and print only text also make the font size thinner than in the origin image e The most recent application of Arabic OCR discussed a one problem that is to rescanning the old damaged papers to extract text and save it in new documents e Prove that there are different ways to use OCR techniques for enhancement like extend OCR into graphics e The proposed research can be used in Spam Detection from image In the past the ads are sent as a text document file but now
59. arer eae saan ene 76 S 1 CONCIUSION ics sicesaeees tn n a eet eens 76 eee 77 teeny eh FUCUPE WONK isan 5 2 Bibliography 00 78 52 e 52 teen 6 00 11 Details about ABBYY Finereader aa a 92 vested esasa 006 00 8 lS toesecacteres ves teesecs clitnreastcavses tetesranceeeeaees 92 006 List of Figures Figure 1 1 the architecture of TIE SYSt M ccccccecseeceeceeceeceeceeeeeeseeceeceesaeseeseesaeeaeeeas 5 Figure 2 1 the proposed method for Kim et al in 11 22 22 60000600000 12 Figure 2 2 represent localize text region and text extraction c ceeeeeeeeeeeeeeeeees 13 Figure 3 1 the model for research design ccccccecceeceeceeceeceeceeseeeseeceeseeseeseeseeaeesaes 24 Figure 3 2 example to show how identify a rectangle shape over an items 31 Figure 3 3 circle item with label One 252222 31 Figure 3 4 rectangle item with label two sss 31 Figure 3 5 triangle item with label three
60. as the dot of the i They present the conditional dilation algorithm that depends on Character Connectivity Condition which is An expansion pixel needs to connect to at least one and at most two characters This is because the maximum neighboring characters that any character in a text string can have is two and Character Size Condition which is If an expansion pixel connects to two characters the sizes of the two characters must be similar 2 Dynamically group the extracted text pixels into the text strings which are the major contribution in his paper 3 Detect the orientation of each string and rotate it to horizontal direction for text recognition using commercial OCR product like 13 Arabic Typed Text Recognition in Graphics Images ATTR GI ABBYY 10 and Strabo systems but experiments on Arabic Text are not considered Zhidong et al 35 this paper represented a technique to extract a character text from printed document after scanned the document Used optical character recognition OCR system and hidden Markov modeling HMV technology to model each character they used this technique on English Arabic Chinese and in Fax data The challenge in this paper is combining between more than one language in HMV model Because most previous researches used single language in HMV model The basics system OCR system includes two parts training and recognition system The recognition system used the same preprocessing and featur
61. be used to deal with huge vocabulary such as printed documents library and needs a largest hardware distributed infrastructure Through the experiments of the authors on around 20000 Arabic words was randomly chosen from high and medium quality documents The obtained result shows that the recognition rate average is more than 97 and the segmentation rate average is more than 98 and increases with the size of text font used The difference here is inserts character in different shapes and styles in HMM to use it in classification but we need to try this HMM to insert picture for each character and see the result Mohieddin et al in 21 proposed a novel Farsi text detection from video images by corner detection An edge detector operator in all possible direction vertical horizontal 45 and 135 are extracted To extract text some pre processing is done by dilation and erosion according to the font size Then corners map are extracted from edges cross point They used histogram analysis to prevent non text from appear After that rescale the image to get new corners map Finally to detect candidate text they use empirical rules analysis For experimental results they use precision 72 8 and recall 78 54 on 50 images with resolution 720x576 The precision rate is defined as the ratio of correctly detected words to the sum of correctly detected words plus region which actually are not text but the algorithm detected as text region false posit
62. characters wrote in above and below the base line a Has small characters with G gives long height b Capital and small characters have the same height of a c and d All characters wrote on three lines have the same height 74 Figure A 1 to run ABBYY Screenshot Reader cccccecceeceececseeceeeeeeceeceeseeceesesseesaeeaaes 82 Figure A 2 The ABBYY Screenshot Reader WiINGOW ccceccecceeceeceeceeceeseeceesaeseesaeesees 82 Figure A 3 Example takes by ABBYY Screenshot Reader by click Enter key for English page E E cd E NTE TE EE A e 83 Figure A 4 the result of extracting the text in image was appearing in figure A 1 84 Figure A 5 Example takes by ABBYY Screenshot Reader by click Enter key for Arabic page84 Figure A 6 The result of extracting the text in image was appearing in figure A 3 85 Figure A 7 supported language WINCOW sass 85 30 2115 29 011 assesses 86 Figure A 9 this window appear when you choose Quick Tasks in figure A 6 86 Figure A 10 the first example containing Arabic and English text with white background 87 Figure A 11 conformation message because resolution is less than 400 87 Figure A 12 this window appear to continue the Convetion 0 6655 ssccecceeeeeeeeeeees 88 Figure A 13 the result for first example used in figure A 8 0ccc
63. d Text to Microsoft Word Figure A 2 The ABBYY Screenshot Reader window 82 Arabic Typed Text Recognition in Graphics Images ATTR GI In figure A 2 you must choose Capture to identify the shape of selection area That may take Area window screen Timed Screen Also choose the Language to be used as a data set to compare the text in the image The common languages used are English German French Spanish and Italian also there is another choice to select more Languages Finally select from Send list box the type of the document you want to convert to it There are many types like text word table and image Then click on the button beside the list boxes Then you can see the capture you are select from ABBYY Screenshot Reader To try it on English page choose your preferred part and then click Enter from key board The part you select convert to image as you see in figure A 3 Then after that automatically give a result as a text on a word document as you selected see figure A 4 It gives an excellent result for extracting English text with little mistakes to identify character D We tried the same program on Arabic web site as we see in fig 5 The result represents a garbage characters as we see in figure A 6 because Arabic choice is not in the language list box see figure A 7 Desktop Installation Desktop Desktop Soran Desktop Desktop Per Seat License Freeware Standalone CTE Standalone Stan
64. dalone Product Information More Info More Info More Info More Info More Info Buy OCR Software You can use the buy Now buttons to order online and download your O Did you know Desktop OCR Software ABBYY FINEREADER 11 PROFESSIONAL FineReader Professional is a highly accurate and easy to use OCR software that includes host of features including digital camera OCR intelligent document layouts image enhancement barcode recognition and command line integration FineReader is our pick for OCR software e MEY Fleeheader 11 because its document layout retention will save you much Figure A 3 Example takes by ABBYY Screenshot Reader by click Enter key for English page 83 Arabic Typed Text Recognition in Graphics Images ATTR GI Ar The ABBYY Screenshot Reader window Microsoft Word ETE 6 5 Home Insert Page layout References Mailings Review View Add Ins AaBbC AaBbce AAB 4astce c ganada FET 15 2 2 tc Courier New Je a ay Replace 3 Copy selea Aa A 2 5 3 2 2 lt Heading 2 Title Subtitle Subtle Em B Z U Paste a 2 Clipboard m Font Paragraph Styles J 19110 1LL 112 113 1 14 11S 1 16 117 118 119 120 1 21 installation uesKtop uesxiop Server uesxtop Desktop Rex seat License Freeware Standalone Standalone Standalone Concurrent Product Infor
65. e extraction used in training For feature extraction used feature vectors for each frame each word or sub word and the time is an independent variable The feature vectors are intensity vertical derivative for intensity horizontal derivative for intensity and local slope and correlation across a window of 2 cell square Combined with the different knowledge estimated from training were used to find character sequence They conclude that there are systems could be work on different language by training on new set of data based on the HMM model in the OCR system Also with training the system could achieve a robust performance on degraded data The different the system applied on Arabic English Chinese and Fax text but we focus only in Arabic The steps to solve the problem are semi equal to our steps We can say from the above related work there are two types or levels to verify recognition process One of them based on text recognition and the other one based on character recognition Also they used different features which might be used in our work especially the features mentioned in 11 and 34 14 Arabic Typed Text Recognition in Graphics Images ATTR GI 2 2 Arabic Optical character Recognition Optical Text Recogntion research work Kasmiran et al in 9 said that Arabic Optical Text Recognition AOTR is a branch of OCR The first published work on AOTR is back in 1975 as a Master s thesis that developed a system fo
66. e image tem i ia VI Arabic Typed Text Recognition in Graphics Images ATTR GI Table of Contents Abstract A usec ai scaasasesoanecusecehsace O OS 27 ae VI Chapter d O 010101010101011 Saa 1 Itr oduction ES O E a 1 1 1 Optical Character Recognition 222222226 1 1 2 Optical Text 6 8 007 22222220 aaa asa aa aaa 2 1 3 Text Information Extraction and its Stages ccccccseccseeceeeceeeeeeeceeeseeeeaeeees 4 1 4 Arabic Optical Text Recognition and challenges sass 6 1 5 Problem Statement asas a ohana 8 1 0 066111765 issa a SS a a Sh 8 1 6 1 Main Objective 8 1 6 2 Specific ODjectives asides aaa 8 1 7 Scope anid Limitation aa 8 1 8 Significance Of the thesis ns 9 1 9 6563 6 3 22 naa et 10 3016 7 a esis hak Baan 11 Background and 5 enn 11 2 1 Non Arabic OCR OTR research work sss 11 2 2 Arabic Optical character Recognition Optical Text Recogntion research work
67. e try to separate text recognition step from the rest of steps By making text recognition as a special stage separate and essential to accomplish the rest of the tasks This will cause a paradigm shift so that makes it easier for researchers to focus only on the part of the solution So that development can be easily and become an operations molecules can be linked with any appropriate portion which increases the efficiency and improvement in test results the appropriate section and link it with the rest of the parts So that we will present more definitions and details that are supports and describe our idea In next two sub sections we try to clarify the meaning of Text Information Extraction TIE and Arabic Optical Text Recognition AOTR Also present some challenges in this area 1 3 Text Information Extraction and its stages Kim said in 11 Text extraction research can be divided into two groups graphic text 7 37 and scene text 12 23 and 38 extractions Jain et al 7 extracted graphic texts in various images binary web color and video frames Gray level values and color continuity were used as features The method had good performance for binary web and video images but it had poor results on color images To extract text from an image you need as a first step zoning that defined in 34 as which analyzes the layout of an input image for location and ordering the text blocks Then each text blocks
68. eccseceeceeceeseseeeeaeesees 88 Figure A 14 the second example containing only Arabic text with background color 89 Figure A 15 conformation message represent the program can not detect any object to FECOPMIZE teeter sieeve E a a E a 89 Figure A 16 result for second example take in figure A 12 sss 90 Figure A 17 third example containing Arabic and English with background colotr 90 Figure A 18 conformation message because resolution is less than 400 91 Figure A 19 result for third example take in figure A 15 sass 91 Arabic Typed Text Recognition in Graphics Images ATTR GI List of Tables No table of figures entries found Table 1 1 A comparative between Scand pages and graphics advertisement 7 Table 4 1 variant background level complexity 63 Table 4 2 variant text level complexity 63 Table 4 3 sample results for OCR system 65 Table 4 4 results for all above 90 images and first of 50 images using first evaluation 65 Table 4 5 represents results for each group A B C D and E using first evaluation 66 Table 4 6 represents values for each 50 and 90 item using second evaluation 66 Table 4 7 represent values for each group using second evaluation 67 Table C 1 group A includes different style texts were written on white background 106 Table C 2 group B includes different style texts were written on
69. ect area cliparea dx dy gt width height blackratio 2 3 cliparea countwhiteratioarea countwhite cliparea countblackratioarea countblack cliparea blackoverwhite countblack countwhite if countblackratioarea lt 0 28 if BWarea gt 30 if blackratio BWarea gt 2 5 If blackoverwhite lt 0 251 ixc 1 For hr from minr to maxr 104 Arabic Typed Text Recognition in Graphics Images ATTR GI 15 1 16 For hc from minc to maxc 17 If hr O and hc 0 18 I3 hr hc BW3 ixc iyc 19 End 20 iyc iyc 1 The variable ixc and iyc are using 21 End to make a copy from the region image crop to its position in 2 22 ixc ixc 1 Sulu iE 23 End 24 End 25 End 26 End 27 End Pseudo Code for Algorithm 3 16 Show Post Processing Input I is a colored image Output Img is output image for show I2 convert to binary of I 13 preprocessing I I3 and 14 is binary image 14 postprocessing I3 as Swap between black and white in I4 Now apply thicken skeleton on I4 Swap again between black and white in 14 rgb Convert 14 image to RGB image I2 Convert I2 image to RGB image Change foreground color in I4 to red color 10 Img multiply 12 with rgb RW 105 Arabic Typed Text Recognition in Graphics Images ATTR GI 1
70. ers such as Canny Sobel and Prewitt But any one does not enough to get a suitable amount of unwanted details and to represent the most foregrounds So after searching into many filters we decide to use these two filters wiener2 and rangefilt respectively A graythresh wiener2 and rangefilt are MATLAB build in functions see figure 3 9 27 Arabic Typed Text Recognition in Graphics Images ATTR GI Wiener2 lowpass filters a grayscale image that has been degraded by constant power additive noise wiener2 uses a pixel wise adaptive Wiener method based on statistics estimated from a local neighborhood of each pixel The additive noise Gaussian white noise power is assumed to be noise Rangefilt determines the center element of the neighborhood by floor size NHOOD 1 2 Our decision is coming from two reasons The first reason is to remove some noises like graphics from the origin image by keep it the most amounts of elements in the image The second reason is to perform image with just two colors black background and white foreground to help us to deal with the variance background color in the same origin image This situation is required to apply other techniques like segmentation process and skeleton thinning process After that we use Graythresh It is an automatic threshold to find a suitable threshold that associate for each image separately The graythresh function uses Otsu s method which chooses the thre
71. faces like pen based computers 9 The first published work on AOTR is in 1975 by Nazif in 22 as a master s thesis Developed system for recognizing printed Arabic character based on extracting strokes that he called radicals and their position This kind of research faces many challenges The challenges in AOTR are related on e The different shapes of Arabic characters e The overlapping between characters when it s written as a printed computer character or handwritten character e Diacritics and variety of Arabic font 9 We make a comparison between the standard or traditional image among Scanned pages that used to insert paper into computer to use it as a data set to approve his her technique which is the only way to repair degraded papers Papers with A4 size are easy to know the start point the spaces between lines It can be easily identify the text from other part of the scanned paper because it has a white background and a black foreground Also the text s size type orientation is known Arabic Typed Text Recognition in Graphics Images ATTR GI The other type which is newest one is a graphics image The image is ready in a computer with variant background color also includes different objects with different shapes With or without text If it has a text not necessary to start every time at the same point In addition could be written horizontally vertically or in diagonal Other problems are the size color
72. for each item in original image is done Note the background in part 4 in figures 3 7 and 3 8 are white because we exchange the color between background and entire elements to display the rectangle shape around each item Figure 3 8 example 3 Apply connected label on black background The new idea is depending on connected label only with small distance we assume it equal to one including information about each item This information is maximum row and column and minimum row and column which are preserved in a two dimension array with two additional fields These fields are object number and label An ascending sort is applied to the content of the two dimension array depending on the value of minimum row see algorithm 3 2 Object number holds the value of item order that presents by connected label method Label initialized with zeros and use it as classification field to put a new label for the same item with the nearest minimum row minimum column and maximum row These values depending on how much the content are closed on each image So according to the label value we make groups or regions of items that verify pervious conditions These conditions come from the computation of the difference between the minimum rows The differences saved in one dimension array Items in the same row have minimum difference and if the difference has large value means that is a pointer to start a new row So to know the difference is small o
73. ge Finally multiply 13 by 14 to get results appear in table 4 7 clearGraphic 14 BW I read RGB image 12 preprocessing I 13 postprocessing I2 Convert 3 to RGB red white Convert 4 to RGB black white Multiply 14 by 3 Display result image Figure 3 11 clearGraphic filter that represents the base structure for our work 56 Arabic Typed Text Recognition in Graphics Images ATTR GI Preprocessing Apply algorithm 3 6 Apply algorithm 3 1 give localization process step 4 Apply algorithm 3 11 find threshold for image filter accumulator of adjacent sequence black pixels in base Apply algorithm 3 7 Apply algorithm 3 2 localization process step 1 localization process step 5 Apply algorithm 3 12 find peak and its position Applv algorithm 3 3 i Apply algorithm 3 8 P ee ee ea collect information difference and mean Apply algorithm 3 13 find half line and base line ratio Apply algorithm 3 4 PRP BONES il i alization process step calculation Arabic base line Apply algorithm 3 14 set of rules Apply algorithm 3 5 localization process step 3 Figure 3 12 structure represents the steps of preprossing Postprocessing Apply algorithm 3 15 OCR Apply algorithm 3 16 show Apply algorithm 3 10 find length of adjacent sequence black pixels in base line postprocessing postprocessing Figure 3 13 structure represents the two ways of po
74. graphics documents and preparing them for OCR processing Our detection approach is based on some mathematical measurements to know is it a text or not and to know is it Arabic Based Text or Latin Based These measurements are follows measure the Base Line the line has maximum number of black pixels Also measure Item Area the content of extracted sub images Finally find maximum peak for the adjacent black pixels in Base line and maximum length for sub adjacent black pixels Our experiment results will come in more details We believe our technique will enable OCR systems to overcome their major shortcoming when dealing with text in graphics images This will further enable a variety of OCR based applications to extend their operation to graphics documents such as SPAM detection from image reading advertisement for blind people search and index document which contain image enhancing for printer property black white or color printer and enhancing OCR Keywords Arabic Text Recognition ATR Arabic text information recognition ATIR Arabic Typed Text Recognition in Graphics Image ATTR GI Arabic optical Arabic Typed Text Recognition in Graphics Images ATTR GI character recognition AOCR Text Extraction Optical Character Recognition Text recognition Glossary In our research there are some words we use it in this meaning like EJO tiers ord 7 7 1 This word we use it to refer the elements in entir
75. hancement 4 ClearGraphic filter could be localize all text but after applying text type detection method there are some Arabic character word or text are avoid it 5 2 Future Work e After finish the segmentation process enter the segmented parts into machine learning with extracting feature to make a text classification method which minimize error ratio e Extract diagonal Arabic text We have some attempt depending on computing the diagonal direction and its angle then make a rotation to that angle in clock wise minus 10 We are depending on our work on a ready method done by Gonzalez It works successfully but needs more preprocessing steps to be ready to work in our algorithm e Enhance our results by create a method that can detect South West angle This angle helps us to detect Arabic text in easy way in a paragraph region e Enhance segmentation process such as find a way to display texts that remove by rangefilt like using inverse method Also use filter to remove noise from final output e Add our ClearGraphic filter to MATLAB library as an enhancement filter in image processing e Integrate ClearGraphic filter with OCR system to increase the chance of extracting text from complex image 77 Arabic Typed Text Recognition in Graphics Images ATTR GI Bibliography 1 Abuhaiba S 2006 Segmentation of Discrete Arabic Script Document Images Journal of Al Azhar University Gaza Natural Sciences
76. he most important features of text than graphic Such as computes a ratio of item area number of black pixels over the segment area to take a small factor that separate between text and other graphics The other feature is Arabic text was written just on one line which we call it Base Line where the density of black pixels larger clear rather than English text which was written on four lines Finally compute the maximum peak for adjacent black pixels in the Base Line to know if it is English or Arabic Because we claim that Arabic text has a long horizontal adjacent line rather than English which is written as separate characters So that to solve the problem there are three steps These steps are 1 preprocessing 2 text detection and 3 post processing and their implementation will be discussed in the next sub sections 3 2 Techniques Based on the set of Arabic features different techniques are discussed to select the best one which provides an acceptable performance and results 3 2 1 Preprocessing In this stage we are making some preparations on the original colored image to be ready to text localization stage This preparation revolves around remove some noise from background image Also make the background to have a uniform color Finally convert the original image to black and white image The first preparation is converting colored image to gray level image Then we need to perform an edge map There are many filt
77. ic Typed Text Recognition in Graphics Images ATTR GI 7 The evaluation steps that used in second evaluation technique As mentioned this evaluation has two partitions manual and automatic e Manual partition as follows 1 Take a copy from each image after applying an interior filters and localization step Save these copies in a folder Open Paint program application to erase unwanted data from the copies e Automatic partition as follows 1 Open two images the first one is the colored image and the other one is copied image after erase unwanted data Count all black pixels in the original copied image to use it into final evaluation Our filter call comparison function that compares each black pixel into the origin filter copy is in the same place in the output result or not If the black pixel in the same place we computed as correct result If there is a missing data or unexpected data in the output result we computed as incorrect result Finally to compute accuracy rate for each image divide correct result over all black pixels and to compute error rate divide incorrect result over all black pixels To compute accuracy rate for all images Find the total of correct result for all images and divide it over the total of all black pixels for all images To compute error rate for all images Find the total of incorrect result for all images and divide it over the total of all black pixels for all images see
78. ifferent style texts were written on background filled with variant gradient color 3 6 Yee 4 Jee geass ail sioi pili er hry O 42 3 Siew LG Aig bind gt m ae O correlation correlation correlation correlation n IRO OE in 0 99 out 0 0005 171 078 0171 in 0 86 out 0 13 in 0 83 6 9 ETE 75 1i 5 I mero aL Vif correlation correlation correlation correlation correlation in 0 72 out 0 27 ma oa GEO in 0 95 out 0 04 in 0 92 out 0 073 13 14 15 E 2 a gt correlation
79. ify the area of new label 42 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 7 Localization process step 5 select text region Input 1 start_min_row vector is a vector that contains all minimum rows for after classification labels use algorithm 3 6 Input 2 start_min_col vector is a vector that contains all minimum columns after new classification labels use algorithm 3 6 Input 3 end_max_row vector is a vector that contains all maximum rows after new classification labels use algorithm 3 6 Input 4 end_max_col vector is a vector that contains all maximum columns after new classification labels use algorithm 3 6 Output 1 vector of text region crop Output 2 obj region number Output 3 4 5 6 start_min_row start_min_col end_max_row end_max_col vectors 1 Compute width and height for each new object 2 draw rectangle a round new object 3 crop the new object to check if it is a text or not To do previous steps we need to collect some information The technique we use it is described in next sub section 3 1 2 2 Collect Information This information is necessary to deal with crop of images in a proper way Such as to find the larger row that include maximum black pixels i e Base line To find maximum peak value is calculated the frequency of smaller connected black pixels in the base line Values of peak and base line are using to build a set of formulas and rules to get final goal
80. ine to display different Arabic font size and style Base Line is the horizontal not vertical not diagonal and not curved straight lines which use it for writing on Arabic papers Over this line there are characters are written either above or below the line according to the script of each Arabic character So the intensity of black pixels in this line is very clear and you can call it a winner row In addition Arabic base line has a big amount of density of color after apply thinning over the region of crop_image Intuitively the base line comes in specific place below the half of the region of crop_image not in the above Therefore any row exists in the upper half or any place rather than the winner row is not acceptable Thus we minimize the number of region of crop_images that we will be checked The check process is only restricted for text not for other shapes Sometimes this feature removes some of English text Thirdly to distinct between English and Arabic we use feature of English script on Base line This feature is the English character either small or capital letter wroteon Base line after thinning has maximum peak for one black pixel See algorithm 3 10 3 11 and 3 12 Thus some English text is removed from the output result You can see the results in chapter 4 49 Arabic Typed Text Recognition in Graphics Images ATTR GI In the next sub sections we illustrate the calculation Arabic Base line ratio algorithm This a
81. ion Enable image preprocessing Split facing pages Figure A 10 the first example containing Arabic and English text with white background Open and convert images PDF file Opening the document Opening the document in Microsoft Word A File C Documents and Settings lalsaedi Desktop images images pal ipg page 1 The resolution of the source image is too small and the image has been stretched J Page 1 Image resolution may be incorrect For best results set the resolution to 400 Figure A 11 conformation message because resolution is less than 400 87 Arabic Typed Text Recognition in Graphics Images ATTR GI ABBYY FineReader 11 Px You are using ABBYY FineReader 11 in trial mode Pages left 50 You can save only 1 pages at a time Activate License and Switch to Full Mode Serial Number Required Figure A 12 this window appear to continue the convetion process CENOM Document2 Microsoft Word Lex Home inset Pagelayout References Mailings Review View _ ABBWFineReaderi1 Acrobat 9 hie E E 2 a A Find gt ie ESSE RE B AANA EHS 2 22 21 hasce AaBbC aaBbcc AAB A4aBsce acsbce 2 A esate a 8 Ht No Spad Subtitle Subtle Em Change am 3 Styles gt i Select Editing S 9J Lsl99jd LxO gt xJ19 OULo NI J O UAT o 1 211119 G JRA 3 t QLOQJT 990 ojb WWW SIMPLES
82. ion accurately as depicted in figure 12 Arabic Typed Text Recognition in Graphics Images ATTR GI 2 2 Text extraction consists of adaptive thresholding dam point labeling and inward filling a Result of labeling dam points in gray b Result of inward filling Figure 2 2 represent localize text region and text extraction This work is much related to our work according the solution technique steps but not in how to implement it Also they work on English and Chinese but our work is on Arabic Based only Other difference is they work on video but our work is on static image Yao Yi et al in 34 focus the light on the public problem on English text recognition where previous researchers were not going to solve it and tolerated to find a solution for it which that is Multi oriented multi sized and Curved text They discussed the problem of text that not written in horizontal but written in vertical or on curved line and there techniques does not require training for specific fonts and can be easily integrated with a commercial OCR product They are going to measure the curved angle then put a box on each individual character to make the recognize process easy for him These are major steps for them approach for text recognition 1 Extracts the text pixels from the input document then they have a binary image where each connected component in the foreground is a single character or a part of character such
83. ive Fp known as wrong detection Recall rate is defined as the ratio of correctly detected words to the sum of correctly detected words plus region which actually are text but the algorithm does not detect as text region false negative Fn known as missed detection As a result of some experiments they found the algorithm done on horizontal text not on other direction This work is related to our work because it works on Arabic text but not on different type of text Our algorithm can detect variants font style but not on font with word art Omar et al in 24 used customized technique to recognize Arabic alphabets Also they illustrated the difficulties faced on Arabic alphabets and 18 Arabic Typed Text Recognition in Graphics Images ATTR GI the architecture of system which divided into three parts preprocessing and line extraction segmentation and character recognition using neural network to identify the tall and width of characters and to specify the dots and Hamza Two important features of Arabic Typed text are mentioned The first one is the main line and the other is characters occupies rectangular space in the line Then they displayed the algorithm and experimental results that appear the effectiveness of the system using different font s type and size with average recognition rate of 87 Paul et al in 25 main idea is how to enhance the bi tonal images So they collect group of degraded Arabic documents the
84. ke to thank every people whom supported to me and particularly Dr Ala a El_Halees Dr Eyad EL Agha Dr Rawia Radi Miss Marwa Abo Jalala and Miss Sarah Kuhail I want to thank my father and my sisters Sana a and Shyma a also my brother Abdullah for supported me for all time I don t forget to thank Islamic University and all teachers in faculty of Information Technology specially dean Dr Tawfieq Barhoom Finally I would like to thank all my friends and everyone helps this research to see the sun light Arabic Typed Text Recognition in Graphics Images ATTR GI Abstract While optical character recognition OCR techniques may perform well on standard text documents their performance degrades significantly in graphics images In standard scanned text documents OCR techniques enjoy a number of convenient assumptions such as clear backgrounds standard fonts predefined line orientation page size the start point of written These assumptions are not true in graphics documents such as Arabic advertisements personal cards screenshot Therefore in such types of images greater attention is required in the initial stage of detecting Arabic text regions in order for subsequent character recognition steps to be successful Special features of Arabic alphabet characters introduce additional challenges which are not present in Latin alphabet characters In this research we propose a new technique for automatically detecting text in
85. l Settings Now we will to describe the two types of evaluation techniques to represent accuracy and error rates The first evaluation is Micro precision and Micro recall formulas to evaluate our work All measurements are applied manually and results are summarized in table 4 6 and figure 4 1 The second evaluation is a comparison technique This comparison includes two partitions The first partition is manual and the other partition is automatic This comparison finds the rate of correlation between the origin output image and the output result The definition of Precision is the fraction of correct results where tp is the true positive to summation of correct results and unexpected results as follows 58 Arabic Typed Text Recognition in Graphics Images ATTR GI TP TP FP Precision Equation 3 Recall is the fraction of correct results where TP is the true positive to summation of correct results and missing results where FN is false negative as follows TP TP FN Recall Equation 4 In our work Eq 3 and Eq 4 are used for evaluation To measure Accuracy Eq 5 is used which is the fraction of correct results and correct absence true positive and true negative to the summation of correct results unexpected results missing results and correct absence true positive TP true negative TN false positive FP and false negative FN respectively which are known as binary evaluation
86. lgorithm helps us to display text with different font size and style Set of rules algorithm shows the sequence steps to get result that use as input to OCR system 3 2 2 3 1 Find Base Line Ratio We used the length of height for each image_crop to measure the ratio place for Arabic base line as follows Algorithm 3 13 find half line and Base line ratio Input 1 imx is the height region crop Input 2 Base value for the region crop use algorithm 3 9 Output 1 halfline the value of half line Output 2 ratio range that maybe found place of base line in the region crop 1 halfline is the floor of the middle row of crop image 2 defhalf is the distance between the height of crop image and halfine 3 defBase is the distance between the height of crop image and Base line 4 If defhalf lt defBase then ratio is the floor of defhalf over defBase 5 Else ratio is the floor of defBase over defhalf D Finally ratio is equal ratio multiply by Base 3 2 2 3 2 Set of Rules This section is the last step in our work It is a compilation of previous values We will use the most nearest values in a collection of rules These rules can identify the largest amount of Arabic region of images crop from the set 50 Arabic Typed Text Recognition in Graphics Images ATTR GI region images crop Now as we see from the beginning of this research we want to display only the Arabic text images So in next algorithm we are expl
87. mation More Info More Info More Info More Info More Info Buy OCR Software You can use the BuyNow S buttons to order online and download your O Did you know B Desktop OCR Software ABBYY FINEREADER 11 PROFESSIONAL Buy Now A FineReader Professional is a highly accurate and easy to ugg OCR software that includes host offeatures including digital camera OCR intelligent document layouts image enhancement barcode recognition and command line integration FineReader is our pickfor OCR software because its document layout retention will save you much Page 3of3 Words 121 QS English U S 5 88 4 3 Figure A 4 the result of extracting the text in image was appearing in figure A 1 F Opa gyan CUES oyo m m ren gare 1 11 4 2012 25 4 ies 4 gt 10 4 2012 4 4 Nilesat 11393 V 4 8 Axio 44 9555 nal 10 4 2012 3
88. n be applied for two ways first way is select a specific area from image to extract the text either into word or on a web page etc The second one is choosing a homogenous image with complex or white background to detect and identify the area which contains text then extract the text into word or other types of files The second part is an application or software includes Arabic as a choice and detects the area which contains text but the result is not in an efficient way like OCR TextScan and Readiris On the other hand there is an application based on select a specific region that contains a text to recognize and extract Arabic characters some of them is a free on line application like free Online OCR and the other is a tool comes with windows or office like OneNote 2007 and Microsoft Office Document Imaging MobiReader Biz Business Card OCR Reader is a business card recognition application developed by DIOTEK It has many facilitates one of them is OCR technology The image captured with an iPhone which can be converted into text and stored by using OCR According the virtual business card holder function could make a phone call by selecting a telephone number in the Business Card Holder and send an email message or SMS 50 Another application software is ABBYY FineReader 11 professional Edition refer to Appendix A for exclusive details which export or saving to an extend application via one page at a time 47 This application te
89. nd 128 131 30 Tsoumakas G Katakis amp Vlahavas P 2010 Mining multi label data In O Maimon amp L Rokach Eds In Data Mining and Knowledge Dis covery Handbook pp 667 685 Heidelberg Germany Springer Verlag 2nd ed 31 UP o e 2009 chapter 8 Evaluation in information retrival 79 Arabic Typed Text Recognition in Graphics Images ATTR GI 32 Volker M Haikal El A 2009 Arabic Word and Text Recognition Current Developments Technische Universitaet Braunschweig Institute for Communications Technology IfN Schleinitzstrasse 22 38106 Braunschweig Germany 31 36 33 What Resolution Should Your Images Be 34 Yao Yi C Craig A K 2011 Recognition of Multi oriented Multi Sized and Curved Text 35 Zhidong L Issam B Andras K John M Premkumar N and Richard S n d A Robust Language Independent OCR System 36 Zied T Mohamed and Maher K 2011 Arabic Cursive Characters Distributed Recognition using the DTW Algorithm on BOINC Performance Analysis UACSA International Journal of Advanced Computer Science and Applications Vol 2 No 3 75 79 37 Zhong Y Shang and Jain A K 2000 Automatic Caption localization in Compressed Video IEEE Trans on PAMI Vol 22 No 4 385 392 38 Zhong Y Karu K Jain A K 1995 Locating Text in Complex Images Pattern Recognition Vol 28 No 10 1523 1535 39 Character Recognition by Fea
90. nds to work on English and on Arabic pages It gives a good result in English but in Arabic is misunderstand because does not support Arabic although the description of the software claims it support Hebrew Farsi and Arabic 21 Arabic Typed Text Recognition in Graphics Images ATTR GI Another program is OCR TextScan 2 Word 1 0 Ink program can easily scan paper documents The program now tries to find the text information out of the scanned picture and to save it as word file or text file Still have to correct the texts but you save a lot of time compared with complete retyping of the text Text in the most used fonts can be recognized Can start the OCR process over command line for large amounts of scanned files Only have to provide the program with the name of the picture file and the output RTF DOC file This information written in the user manual comes with the program But when tried this program to open an image and extract the text can t give any clear information neither Arabic nor English Also we try the Readiris Pro 8 0 With Readiris 8 0 the OCR software detects columns in the document and can recreate them in the output file Scan a columned document produces a Word document with editable columns By editing the text the text flows naturally from one column to another Recognizing a color page can now easily take a few seconds less and Readiris was already the fastest OCR package on the market This inf
91. ng a connected label method As a result of this the segmentation process does not work correctly to separate the elements in the original image has colored background 3 Figure 3 6 example 1 Apply connected label on colored background For more declaration look at figure 3 7 We have in part 1 image with two different backgrounds one of them is white Label 2 is a black white image Label 3 contains two centroid stars one of them is exactly in the centered of item but the other is outside the item That means there are two items in the original image one of them is correct and the other is not So when we apply connected label it works only in the part of binary image have black background as you see in part 4 Thus the method recognized only one item For this point we need to exchange the partition which has a white background to be black and again apply connected label method for this part This case exactly needs to apply connected label method twice Figure 3 7 example 2 Apply connected label on colored and white background 33 Arabic Typed Text Recognition in Graphics Images ATTR GI In figure 3 8 we can see the original image in part 1 has a black background and two white items In part 3 we have two correct centroid stars in the middle of each item As a result of this we get a correct final result in part 4 That has two dashed square boxes around the items So the localization process
92. ng to image_crop 4 Farea Find foreground area 5 Imagearea Find image_crop area width height 6 Objarea Farea Imagearea 7 Apply Pseudo Code for algorithm 3 13 to find halfline and ratio 8 Apply Pseudo Code for algorithm 3 11 to find val and 9 Size I2 size I I iS the origin image 10 12 1 i e 12 has white background 11 if Objarea gt 0 01 and Objarea lt 0 16 12 if Base gt halfine and Base lt hafline ratio 13 if accum C 1 gt 1 and val gt 3 14 ixc 1 15 For hr from start_min_row obj to end_max_row ob 16 iyc 1 103 Arabic Typed Text Recognition in Graphics Images ATTR GI 17 18 19 20 21 22 23 24 25 26 27 For hc from start_min_col obj to end_max_col obj If hr O and hc 0 I2 hr hc BW3 ixc iyc End iyc iyc 1 The variable ixc and iyc are using End to make a copy from the region image crop to its position in 2 ixc ixc 1 output image End End End End Pseudo Code for Algorithm 3 15 OCR Post Processing Input 12 output result image this include Arabic text with some unwanted data use Pseudo Code for algorithm 3 13 Output 13 output image to OCR system 1 Apply Pseudo Code for algorithm 3 2and return clip region maxr maxc minr and minc but with change the distance with 4 and without use objectarray and ascending sort dx dy size clip BWarea foreground obj
93. nt amount of research works in the literature which has handled this problem to a great extent but mainly for Latin text but no equally adequate treatment of Arabic text can be found in the literature Arabic text poses an additional set of different challenges to the problem Obviously whether Arabic or any language the challenge is further complicated when the text is handwritten Arabic Typed Text Recognition in Graphics Images ATTR GI In our research we focus on Arabic typed text in still images and limit the processing on the first three stages of the TIE process namely detection localization and extraction We may extend our work in the future to handle the problem in image sequences At this extent presented thus far the problem can be referred to as Arabic Optical Text Recognition AOTR and we discuss it in more detail in the remainder of this introduction 1 4 Arabic Optical Text Recognition and challenges AOTR Arabic Optical Text Recognition is a branch of OCR specified in Arabic language It is include on line and off line reading technology for handwritten text That used normal paper or electronic media The application of optical character recognition was using this technique in many familiar live areas like check sorting in banks zip code reading mail sorting providing assistance to blind people reading of customer filled form automatic office archiving and retrieving text and improving human computer inter
94. nternational Journal of Computer Vision ICV 65 1 2 43 72 21 Mohieddin Moradi Saeed Mozaffari and Ali Asghar Orouji 2010 Farsi Arabic Text Extraction from Video Images by Corner Detection IEEE 22 Nazif A 1975 A System for the Recognition of the Printed Arabic Characters Master thesis Faculty of Engineering Cairo University 23 Ohya J Shio A Akamatsu S 1995 Recognizing Characters in Scene images IEEE 24 Omar Al J Samer Al K Bashar Al G Mohamed F Hani K 2000 A New Algorithm for Arabic Optical Character Recognition 1 14 25 Paul H Ben H Chris J Linda V G Amlan K 1997 Optimizing OCR accuracy for bi tonal noisy scans of degraded The MITRE Corporation 7515 Colshire Drive McLean VA 22102 7508 Proc of SPIE Vol 5817 179 187 26 Prof Mohsen A A Rashwan and Dr Mohamed Attia 2008 Academic history of Arabic OCR of RDI and principal co principal researchers 27 RASHWAN M A A FAKHR M W T ATTIA M EL MAHALLAWY M S M 2007 ARABIC OCR SYSTEM ANALOGOUS TO HMM BASED ASR SYSTEMS IMPLEMENTATION AND EVALUATION 28 Tarek A E Aly A F 2009 English Arabic Cross Language Information Retrieval CLIR for Arabic OCR Degraded Text Communications of the IBIMA Volume 9 ISSN 1943 7765 208 218 29 Tolba M S Wahab and A Salem 1987 A Recognition Algorithm for Printed Arabic Character Proc IASTED inter Symp In applied information Switzerla
95. o Save the actual value into the accum in index one 47 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 12 find peak and its position Input 1 accum x 2 is a two dimension accumulator first column is the length value and the second column is how many times this value redundant in Bline vector use algorithm 3 11 Input 2 bc value where bc is length of accum two dimension array Output 1 maximum peak value Output 2 C the position of maximum peak 1 Initialize peak by 1 2 Initialize C by 1 3 For all elements in accum search for the maximum peak value and save it in peak variable store its position in C variable 48 Arabic Typed Text Recognition in Graphics Images ATTR GI 3 2 2 3 Text Typed Recognition This section is the final step in text detection It is about how we can know if this part of image I mean here the region of images_crop is a text or not Especially here we make a search only for Arabic text For this reason we focus on the most important feature marked Arabic script is Base Line Firstly to distinguish between text and other shapes or figures in the origin image we compute the region area using existence method bwarea This function returns the area of a binary image The area is a measure of the size of the foreground of the image Then divide the bwarea over the region area to give a specific ratio for each region Secondly measure the Base l
96. ocument4 Microsoft Word i Home Insert Page layout References Mailings Review View ABBW FineReader11 Acrobat Format Cut Pes Sass EF TENT AaBbC aaBbce AAB 4aBbce aasbce Paste BZ U ahe x x A Hr EE ding 2 Title Subtitle Subtle Em f select Paragraph 120111101111111 Styles 1 Page 1of1 Words 0 lt J English U S Open Image Look in yj My Recent Documents My Network Places aa Figure A 16 result for second example take in figure A 12 images 1 X 113 Show preview IE IMG_123201253752AM jpg IMG_1232012537524Maa jpa pall jpg pal jpg N untitled2 bmp N untitled3 bmp N untitled bmp S untitledS bmp untitled bmp File name untitled bmp Files of type Common Image Files Page range All Detect page orientation Split facing pages Pages Enable image preprocessing Figure A 17 third example containing Arabic and English with background color 90 Arabic Typed Text Recognition in Graphics Images ATTR GI Open and convert images PDF file Opening the document Opening the document in Microsoft Word A File C Documents and Settings lalsaedi Desktop images images untitled bmp page 1 The resolution of the source image is too small and the image has been stretched A Page 1 Image resolution may be incorrect
97. ombining and Verification IEEE 12 Li C Ding X Wu Y 2001 Automatic Text Location in Natural Scene Images Proc of 6 ICDAR 1069 1073 13 Magazine M E 2007 Essay On Arabic Ocr Packages In The Market_Windows Retrieved from www itp net Arabic 14 Maher K and Abdelfettah B 2009 Towards A Distributed Arabic OCR Based on the DTW Algorithm Performance Analysis The International Arab Journal of Information Technology Vol 6 No 2 15 Mahmoud 5 1994 Arabic Character Recognition using Fourier descriptors and character contour encoding Pattern Recognition 27 6 815 824 78 Arabic Typed Text Recognition in Graphics Images ATTR GI 16 Matas J Chum O Urban M Pqjdlo T 2002 Robust Wide base line Stereo form maximally stable extremal regions In proceeding of British Machine Vision Conference BMVC 384 393 17 Michael Bukland and Fredric Gey 1994 The Relationship between Recall and Precision Journal of the American society for information science 12 19 18 Michael D Horst B Silk W 2008 Using Web Search Engines to Improve Text Recognition IEEE 19 Michael R Lyu F I FEBRUARY 2005 A Comprehensive Method for Multilingual Video Text etection Localization and Extraction VOL 15 NO 2 243 255 20 Mikolajczyk K Tuylelaars T Schmid C Zisserman A Matas J Schaffalitizky F Kadir J and Van G L 2005 A comparison of a affine region detectors I
98. on for discrete Arabic script documents This study based on scanned document as a dataset to experiment his algorithm The success rate of applying the new algorithm is 94 4 and the average time required to segment one character was 62 msec In his experiments Simplified Arabic and Traditional Arabic used as two basis fonts by developing two discrete versions discrete Simplified Arabic and discrete Traditional Arabic There is no similarity between our approach which is text recognition and the Abuhaiba s approach is character segmentation Also the differences is that we don t need to scan the 15 Arabic Typed Text Recognition in Graphics Images ATTR GI paper to insert into a computer and repair the resolution of the text to give best result because the image is ready and stored by default in a computer and has a good resolution Another thing is in our work there is no standard font style or font size and the background contains a various noise Ahmed et al in 2 represent a framework to segment Arabic Text also to extract features font that used in recognition process Their work based on the combination between Line Adjacency Graph LAG and Base line features They test their special algorithm on two sets of text files The first one includes 31 pages but does not contain diacritics while the second set includes 15 pages which contain diacritics signs All of these pages were insert into computer through scanning process with
99. or each new row Input 1 ascending sort object_array by minimum row use algorithm 3 2 Input 2 difference array of min row see algorithm 3 3 Input 3 average of differences of min row see algorithm 3 3 Output 1 ascending sort object_array by minimum column 1 Put 1 into the label of first item in Object_array 2 Check the value in difference_array according the minimum row to be greater than the average of difference_array If the check is true then increment the label value The rest items labeled with zero Give the new label value to the correct check item and so on This operation identify the beginning item in each row Save the place of each new label and use it as a range in sort by min column 3 Sort object array Obj XI min_col _ sort elements in range of each label in ascending order depending on min_column 40 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 5 Localization process step 3 give labels for the rest of items in object_array Input 1 ascending sort object_array by minimum row and column use algorithm 3 4 Input 2 Avgmaxrow is average of max row see algorithm 3 3 Input 3 Avgmincol is average of min column see algorithm 3 3 Output 1 sort object_array by labels in ascending order 1 In this algorithm need to identify if each item belongs to the new label or it should get a new label value according the distance between each item in one row i
100. orithm 3 10 find lengths of adjacent sequence black pixels in Base line Input Base line for BW3 region of Binary image crop use algorithm 3 9 Output descending Bline vector this vector contains length of adjacent sequence black pixels in base line 1 find the size of BW3 2 Initialize Bline vector by zero The length of Bline is equal the width of crop image 3 For all pixels in the crop image a Check if the two adjacent pixels are black then count black pixels in sumb variable b If the two adjacent pixels are different from each other then Save sumb in the first index in Bline Increment the index of Bline by one Reinitialize sumb by one 4 Sort Bline vector in descending order 46 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 11 find accumulator for lengths of adjacent sequence black pixels in Base line Input 1 descending Bline vector this vector contains length of adjacent sequence black pixels in base line use algorithm 3 10 Input 2 x value where x is length of Bline vector Output accum x 2 is a two dimension accumulator first column is the length value and the second column is how many times this value redundant in Bline vector 1 After sort Bline vector all equalize values are come in front of to each other so 2 Initialize accum with zero 3 For all elements in Bline vector do the following Count the equalize values and save it into the accum in index tw
101. ormation written in the user manual comes with the program but it does not provide any result Other free applications in 40 41 and 45 such as Microsoft office OneNote 2007 and Microsoft Office Document Imaging from Microsoft Office Tool these are depend on selecting the part of text then paste it on any document but does not work correct in Arabic From the previous related works we can summarized as follows Most Arabic related works interested in segmentation process to increase the performance and accuracy either in handwritten or in typed characters Languages rather than Arabic solve the problem of extracting text from graphics image either automatic or by selecting the region of text There is an OCR application success to extract automatically English and other languages rather than Arabic from graphics image 22 Arabic Typed Text Recognition in Graphics Images ATTR GI Arabic research late one step from other languages in this area of interest The next chapter presents the proposed solution to solve the problem of extracting text from graphics image and removing other things This solution works as filter that could be used as previous step for any work uses text processing 23 Arabic Typed Text Recognition in Graphics Images ATTR GI Chapter Three Research Methodology and Techniques In this chapter we will present two sections The first one is our technique which presents
102. plex image that contains horizontal Arabic and English text to extract Arabic text Our approach summarized in three steps 1 Pre processing 2 Text detection and 3 Post processing Use Wiener2 and rangefilt filters and automatic thresholding in pre processing step Also prepare an output result with white background and red marking color for text detection also a copy for OCR system using set of rules These rules depending on 1 compute the ratio of item area over segment area 2 identify base line and 3 find maximum peak In post processing step we enhance the results by removing unwanted items from the results come from text detection step We evaluate the proposed approach on a comprehensive dataset including variety of image samples gathered by our self Using 90 samples and evaluated by two types of evaluation The first evaluation by information extraction retrieval is overall precision 86 96 recall 87 23 and an error rate is 8 86 The second evaluation by correlation is overall accuracy 91 24 and error ratio 8 78 is obtained The ClearGraphic filter is applied on different style of images with variance of difficulties but there are some limitations as follows 1 ClearGraphic filter does not work on diagonal orientations 2 Does not distinguish between English and Arabic in a proper way 3 There is still little 76 Arabic Typed Text Recognition in Graphics Images ATTR GI noise in final results that needs en
103. r big we 34 Arabic Typed Text Recognition in Graphics Images ATTR GI compute the mean of the differences values The mean value is used as a threshold to identify the items in the same row and new labels are provided if the difference is larger than the mean value see algorithm 3 3 and 3 4 But we face a problem in some items The problem is some items in the same minimum row should be in the same region but perform more than one region or group This problem was appearing because items appear in random order So for this reason we apply another ascending sort but this time on the value of minimum column This sort applies on the items belong to the same label Also compute the differences between the maximum columns and minimum columns then find the mean to know if the item is near enough to the previous item or not So that we decide to give a new label for this item or give it the same label for the previous one After that sort labels in ascending order to find coordinates in an easy way see algorithm 3 5 Finally take the minimum and maximum row and column for all items have the same label to perform new selected region Each value stores in one dimension array Then take these regions to know is it text or not see algorithm 3 6 and 3 7 To show the results see figure 3 9 and 3 10 35 Arabic Typed Text Recognition in Graphics Images ATTR GI Wiener2 Rangefilt S
104. r recognizing printed Arabic characters based on extracting strokes and their position by 22 In last three decades most of researches were focused on how to enhance the module of Arabic Character Recognition from three sides 1 accuracy 2 performance and 3 decreasing the time of character detection Start with handwritten detection and finish with reading text file to rewrite make a copy etc Therefore all enhancement techniques have a similarity in a base strategy which summarized as follows 1 Pre processing 2 Segmentation 3 Feature extraction 4 Classification and 5 Post processing Step 4 and 5 were applied according the kind of technique The Arabic OCR for printed text was a research topic started in the 1990s 29 In the next few lines we illustrate some researcher s work used the above steps to recognize Arabic typed characters in A4 scanned image file with white background and black foreground Furthermore they didn t need to detect if it has a text or not because they sure it contains text and start the seek process from right to left But they face some problems to define Arabic features extraction likes Base line is defined in 9 as the line on which all letters lie which use it to determine the space between lines and the distance between characters in the word In addition the orientation of text is needed to be specified Abuhaiba in 1 addressed the problem of line segmentation and character separati
105. racy for this system is 90 with a 20 char s recognition rate The similarity between this work and our work is the steps that must be done in an image to get a single character 16 Arabic Typed Text Recognition in Graphics Images ATTR GI But also it differs because these steps are applied on scanned document with white background and black foreground Kasmiran et al 9 a survey presented states a different Arabic optical recognition AOTR used as off line and focus the light over the characteristics of Arabic writing They presented different stages for preprocessing segmentation feature extraction classification post preprocessing and evaluation methods for different system In addition illustrate the popular steps in preprocessing like 1 Binarization 2 Filtering and smoothing 3 Thinning 4 Normalization 5 Slant correction and 6 Base line and Skew Detection The feature extraction here is for character not for text because authors know there is a text in the image and they jump this step to segmentation step These features are different from one author to another In 15 used these features 1 global transformation 2 Structural features 3 statistical features and 4 template matching and correlation In 19 used template matching between template of the radicals and character image In 23 match the histogram of the input characters to those of the templates The comparative was made between various
106. rocess step 2 give label for each new row Input 1 ascending sort object_array by minimum row use algorithm 3 2 Input 2 difference array of min row see Pseudo Code for algorithm 3 3 Input 3 average of differences of min row see Pseudo Code for algorithm 3 3 Output 1 ascending sort object_array by minimum column 1 Object_array 1 1 1 2 Label 1 3 Obj z 0 4 XI 0 5 For ph from 2 toz gt z iS the number of items 6 If Dif ference_minrow ph gt Avg_minrow 7 Label label 1 8 Object_array ph 1 label E PETT 9 XI XI 1 each new label and 10 Obj XI ph use it as a range in 11 End sort by min column 12 End 13 Sort object array Obj Xl min_col gt sort elements in range of each label in ascending order depending on min_column 94 Arabic Typed Text Recognition in Graphics Images ATTR GI Pseudo Code for Algorithm 3 5 Localization process step 3 give labels for the rest of items in object_array Input 1 ascending sort object_array by minimum row and column use Pseudo Code for algorithm 3 4 Input 2 Avgmaxrow is average of max row see Pseudo Code for algorithm 3 3 Input 3 Avgmincol is average of min column see Pseudo Code for algorithm 3 3 Output 1 sort object_array by labels in ascending order 1 For fg from 2102 2 is the number of items 2 Fe fg 1 3 While object_array fg 1 0 4 If object_array fg maxrow object_array fc maxrow lt
107. s ATTR GI and D see table C 1 C 2 C 3 and C4 respectively Results are depicted in table 4 3 and C 5 Group A contains images with variance text written on white background Group B contains images with variance text written on background filled with one color rather than white Group C contains images with variance text written on background filled with gradient variant color Group D contains images with variance text written on picture background 64 Arabic Typed Text Recognition in Graphics Images ATTR GI Sample of final output results that should be given to OCR system are shown in the next table Table 4 3 sample results for OCR system 1 2 3 4 11 12 After displaying the data set group and result for each image We put now the aggregated results for all 90 images and for each group separately We use four kinds of data representation to represent the results First one is Table shape The second is Doughnut chart Third is Pie chart Last one is Column chart Table 4 4 results for all above 90 images and first of 50 images using first evaluation Macro Macro Macro Micro ac 2 acro acio cro Micro recall precision recall accuracy precision 86 08 76 14 90 53 86 96 87 23 87 94 92 08 92 26 86 90 89 46 65 Arabic Typed Text Recognition in Graphics Images ATTR GI 90 images 50 images E Macro accuresy Merror E Macro acc
108. s like compare sign on a paper or to 19 Arabic Typed Text Recognition in Graphics Images ATTR GI recognize the character But it does not work on other document that contains variant colors font size different orientation and font style In the other hand there are many English applications as we can see can be applied on colored background with different shapes for English character and extract English Text from anywhere in the image and printed on a word file But when we apply it on Arabic image the result either to be unclear or give a result as a picture Point out that all related works found in the literature are based on the assumption that the image is known to contain Arabic text and therefore the purpose of these works is to recognize Arabic characters from the input text Our focus is on the steps before that identify or existence text in the image However most of the techniques used in these approaches can be utilized in our work 20 Arabic Typed Text Recognition in Graphics Images ATTR GI 2 3 Optical Character Recognition Optical Text Recognition Applications There are many applications depend on OCR techniques and text detection to extract text from complex image which are divided into two parts the first one work with English text and the other work with Arabic text Some applications work with English i e MobiReader and ABBYY is needed to install in computer and the other work free on a web It ca
109. sent as images via e mail Companies used this technique to send thousands of messages as a shape of advertisements This way maybe closed the personal mail box as a reason for a lot of messages received in mail box and calls it a Daniel of Service which makes a problem for users that prevent him to receive or send any more or important mail messages In fact there is a SPAM filter to prevent these messages One of the most common ways to prevent Spam is to use data mining method to classify the message contain as Spam or not Spam To overcome these techniques spammer sends the Spam messages as images SPAM filter can t detect the SPAM message that send as an image message which is a new way used by companies to send it s advertisements through electronic mail So SPAM filter may prevent all SPAM image messages depending on the conditions and rules that use it to identify if the message is spam or not SPAM image message is an image contains background color graphic shapes Arabic Typed Text Recognition in Graphics Images ATTR GI photos and Arabic English text with various formats This kind of image used in advertisement and call it Poster or Business card When we need to use it in an Internet must be have 72dpi as a resolution about 400 600 pixels wide for large image 100 200 for thumbnail image Preferred file format is JPEG and the approximate file size 20 200 KB 33 e Enable people with limited vision to read Arabic text from im
110. shold to minimize the intraclass variance of the black and white pixels Finally the combination between filter results and automatic threshold as describe in algorithm 3 1 is to create black and white image Binary image Provided that the background should be white and the foreground should be black to look like an A4 document 28 Arabic Typed Text Recognition in Graphics Images ATTR GI Algorithm 3 1 threshold for edge map Input colored image I Output binary image 1 Convert I to grayscale image 2 K image after applying filter to give a uniform background and foreground color and to remove noise 3 Level the value of threshold for K BW convert grayscale image to black white K Level The segmentation process with edge detection is very comfortable because it makes the background for any colored image to be black and the foreground to be white as a uniform color Then one technique can be used to make the process of segmentation for any RGB image regardless of its content with high speed But without edge detection the RGB image should be converted to binary image through the conversion to grayscale directly which may produces a different variation of background and foreground between black and white So we need to apply different techniques based on exchange between black and white to allow us to use a segmentation process which decreases the execution speed In the next section we propose the adopted
111. stprossing From our experience in this area work on a features of font to find the text in the image is better than comparing each character because it is reduce the seek time process Thus this algorithm could be used as a pre step for the OCR system The next chapter is evaluation and discussion For evaluation we will use two types one of them is manual and the other is automatic on set of 90 images Finally we will list our conclusion and future work 57 Arabic Typed Text Recognition in Graphics Images ATTR GI Chapter Four Evaluation and Discussion To evaluate our work we used two types of evaluation First one is information retrieval context and the second type is finding the rate of correlation between origin output and result output 4 1 Evaluation This section is divided into two subsections First one is experimental settings In this section we descript the formula of information retrieval context and also descript the correlation steps The second is experimental results This section describes the data set classified in groups and present result for each image on alone Also to presents the results for all data set in a table and for each group in another table But I would like to note that there are two tables for each result One of the tables is for first evaluation by using the information retrieval context The second table is for second evaluation by using correlation method 4 1 1 Experimenta
112. table 4 1 and 4 2 62 Arabic Typed Text Recognition in Graphics Images ATTR GI 4 1 2 Experimental results In this section we will show our work results on different levels of difficulties These difficulties come from combination between background levels and text levels see table 4 1 and 4 2 Such as image has variant background color with shapes and pictures also includes text with any font color size style and in different horizontal positions Also includes both Arabic and English languages This example is the most difficult one Table 4 1 variant background level complexity Background Background Include Include Level Color Picture Shape One White Gradient Variant Table 4 2 variant text level complexity Text font size Text Text Include less than font style wrap position Arabic English 75 1 Same Same Any Without pictures Arabic English both Variant Variant 3 With pictures in left Same Same Any right above belowor Arabic English both Variant Variant inside So the rate of found of those factors with each other are affecting on the ratio of final evaluation Our ClearGraphic filter is applied on about 90 images The types of these images are JPEG PNG and BMP Each image that has own nature and there is no similarity between of them So we are classifying them into four groups A B C 63 Arabic Typed Text Recognition in Graphics Image
113. techniques to give a conclusion that the best algorithm used for binary image is MB2 thinning algorithm The best author s knowledge fast and accurate algorithm is the one designed by Amin et al in 4 Also they used decision tree with different masks that give the best result for reliable classification From this work we can decide or choose the best algorithm and best technique that must be used in AOTR Maher et al in 14 used Dynamic Time Wrapping algorithm which considered as one of the strong approaches to several studies have shown that the OCR based on DTW algorithm provides a very interesting recognition rate especially for large and huge vocabularies The attractive sides of DTW algorithm is ability recognize properly connected or cursive characters words or sub words without prior segmentation Also performs the recognition process from within a reference library of isolated characters and owns a very good immunity against noises But the execution time is very slow and restricts its utilization because there is a big amount of computing during the recognition process There are two algorithms used for Arabic recognition 17 Arabic Typed Text Recognition in Graphics Images ATTR GI Hidden Markov Model HMM and the Dynamic Time Wrapping algorithm DTW There is a comparative study between these two algorithms The recommendation is to use HMM to recognize small size vocabulary otherwise the DTW is one recommended to
114. the word is not found in the data set the model does not recognize the word So for this reason they need a huge data to make the search process more accurate Michael et al 18 worked on English language are depend on search process by connected to Google and Yahoo search engine on two levels The first one is word level and the other is statement level Because these search engine contains a huge number of words to improve the recognition process The meaning of Text Recognition in 18 passes through the comparison process between stored words that used for recognition and they decided if this part of image is a text or not So text recognition going through the following phases 1 Text Localization 2 Segmentation 3 Character Recognition and 4 Contextual post processing The researcher should go through these all steps when the researcher needs to make text recognition But in our work we need to solve a big challenge which that how to recognize the set of pixels grouped as a region if it is a text or not i e only focus Arabic Typed Text Recognition in Graphics Images ATTR GI on preprocessing text extraction post processing to be ready to use in other application by search for a good features and test it to perform a rule that suit for any Arabic Text To make Text recognition phase as an individual phase it must be done to comfort the researcher before that to complete next phases In other words we can say w
115. the system i e some of them needs Hardware and Software and the other needs only Software system Another definition defined by AIM in 3 as OCR is the acronym for Optical Character Recognition This technology allows a machine to automatically recognize characters through an optical mechanism Also it defined by Abuhaiba in 1 as is the process of converting a raster image representation of a document into a format that a computer can process It involves many subscriptions of computer science including image processing artificial intelligence and data base systems So we can define OCR as a technique used as part of recent series of operations to recognized handwritten or printed typed characters This needs learn the system how to recognize different characters And possible to make it part of machine learning Arabic Typed Text Recognition in Graphics Images ATTR GI Optical character recognition OCR has been a topic of interest since possibly the late 1940 s when Jacob Rabinow started his work in the field 38 The OCR research started in 1960 s and addressed at first for reading Latin characters After ten years the first paper on Arabic word recognition was published 32 The OCR system can broadly categorize into two categories on line and off line OCR systems 24 The on line OCR defined by 24 as the recognition is performed at the time of writing as it is the case in PADs needs some information like
116. the text and other parts of the image has low intensity So the edge map is perfect solution for complex background and to extract most amount of text automatically from complex image From figure 4 8 and 4 9 by first evaluation You can see the amount of unwanted item in each chart is very high This means our algorithm it gives high accuracy And it is work will to identify unwanted items and prevent it to display into final output result In addition the amount of missing and unexpected items is very low If the image has text region rather than word character region i e long string line this gives a strong factor to success appear Arabic text For example in large font size the character Alef may has character segmentation So when apply thinning method we take a thin vertical line that may to be a garbage line because it does not has base line Thus for that the character was removed From our experiments in this area the reason that displays some English text on the final results We observed that for example the English font with size twenty two and font style is Times New Roman has different heights approximately between twenty one and twenty seven as you see in Figure 4 10 But the heights in Arabic with the same format are between twenty three and twenty nine as you see in Figure 4 11 So there is some interlacing between of them in the conditions It is difficult to separate Arabic from English So we think that
117. there is a sensitive 73 Arabic Typed Text Recognition in Graphics Images ATTR GI factor could be used to remove the interlacing between Arabic and English at all This sensitive factor needs more image analysis X 47 Y 29 Index 1 RGB 1 1 1 a Index 1 ROB 1 b ac c X 67 3 Index 1 RGB 1 1 1 Figure 4 10 Arabic text with parts of characters wrote in above and below the base line a Has characters Geem gives long height b All characters under the base line have the same height c All characters wrote on the base line eghim X 76 Y 27 Index 1 RGB 1 1 1 Kg a b end EHL X 45 Y 21 hee 2 Index 1 RGB 1 1 1 RGB 1 1 1 c d Figure 4 11 English text with parts of characters wrote in above and below the base line a Has small characters with G gives long height b Capital and small characters have the same height of a c and d All characters wrote on three lines have the same height 15 We test our algorithm on Scand image like personal card We are not sure it gives a result But the surprise is our algorithm gives acceptable results And make another test on image that takes via print screen button from desktop and web site And also works well too Finally from our research exactly on how could be evaluate our work We see that each work evaluate his work in variance ways depending on his job
118. ture Point Extraction http www ccs neu edu home feneric charrec html 2 5 2012 40 Free OCR http www free ocr com 2 5 2012 41 Free Online OCR Convert JPEG PNG GIF BMP TIFF PDF DjVu to Text http www newocr com 2 5 2012 42 GOCR http jocr sourceforge net 2 5 2012 43 Google Operating System Open Source OCR Software Sponsored by Google http googlesystem blogspot com 2007 04 open source ocr software sponsored by html 2 5 2012 44 lifehacher five best text recognition tools http lifehacker com 5624781 five best text recognition tools 2 5 2012 45 MakeUseOf http www makeuseof com tag top 5 free ocr software tools to convert your images into text nb 2 5 2012 46 RDI OCR Arabic omni font written OCR http www rdi eg com projects OCR htm 20 3 2012 47 SimpleSoftware http www simpleocr com OCR Software Guide asp 5 4 2012 48 tjansson dk OCRopus open source text recognition http www tjansson dk p 1616 2 5 2012 80 Arabic Typed Text Recognition in Graphics Images ATTR GI 49 Webopedia optical character recognition http www webopedia com TERM O optical character _recognition html 2 5 2012 50 Yshowtopapp http Business Card OCR Reader 5 4 2012 www yshow net show all all all MobiReader Biz 2B Korean 26amp 3B English 0 368180313 html 81 Arabic Typed Text Recognition in Graphics Images ATTR GI Appendix A Details abo
119. urecy W error Figure 4 1 comparison between 50 and 90 image according accuracy and error ratio using first evaluation Table 4 5 represents results for each group A B C D and E using first evaluation Group A Group B Group C Group D 19 15 22 34 Macro precision 84 90 91 60 87 35 83 48 Macro recall 91 34 94 28 83 75 80 95 Macro accuracy 90 86 88 46 90 18 Error 8 06 10 13 9 77 Micro precision 86 57 93 94 89 15 81 23 Micro recall 91 22 96 88 88 63 78 40 Table 4 6 represents values for each 50 and 90 item using second evaluation All black pixel All black in All black out Accuracy Error ratio 297346 271312 26109 91 24 8 78 207712 190264 17511 91 59 8 43 66 Arabic Typed Text Recognition in Graphics Images ATTR GI 350000 300000 250000 200000 E 50 item E 90 item 150000 100000 50000 all black out all black in all black pixel Figure 4 2 comparison between 50 and 90 image according number of black pixels all in and out using second evaluation 90 item 50 item E accuracy W error ratio E accuracy W error ratio Figure 4 3 comparison between 50 and 90 image according accuracy and error ratio using second evaluation Table 4 7 represent values for each group using second evaluation 65195 43730 64683 60374 42569 58177 Accuracy 92 61 9261 973556 89 05 89 94 Error 7 3 11 10
120. use Pseudo Code for 97 Arabic Typed Text Recognition in Graphics Images ATTR GI algorithm 3 6 Output 1 vector of text region crop Output 2 obj region number Output 3 4 5 6 start_min_row start_min_col end_max_row end_max_col vectors For obj from 1 to label widthRegion end_max_col obj start_min_col obj heightRegion end_max_row obj start_min_row obj if widthRegion gt 0 and heightRegion gt 0 draw rectangle a round start_min_col obj start_min_row obj widthRegion heightRegion ImageCrop crop region on black and white image with the same parameters of rectangle shape End End Pseudo Code for Algorithm 3 8 collect information Input BW3 region of Binary image crop Output full info matrix is a Tow dimension array EOE ODN 0 N zx zy size BW3 Full Info matrix zx 3 0 img row 1 For i 1 to zx B count 0 W count 0 For j 1 to zy Full Info matrix j 1 img row If BW3 i j 0 B count B count 1 Else W count W count 1 98 Arabic Typed Text Recognition in Graphics Images ATTR GI 13 End 14 Full Info matrix j 2 B count 15 full Info matrix j 3 W count 16 Img row Img row 1 17 End Pseudo Code for Algorithm 3 9 Calculation Arabic Base Line Input two dimension array full info img use Pseudo Code for algorithm 3 8 Output The Base Line Number maxm full info img 1 3
121. ut I2 output result image this include Arabic text with some unwanted data use algorithm 3 13 Output 13 output image to OCR system 1 oo SS VN FP FP Fe Be on R O 13 14 15 16 Apply algorithm 3 2 and return clip region maxr maxc minr and minc but with change the distance with 4 and without use objectarray and ascending sort dx dy size clip BWarea foreground object area cliparea dx dy width height blackratio 2 3 cliparea countwhiteratioarea countwhite cliparea countblackratioarea countblack cliparea blackoverwhite countblack countwhite if countblackratioarea lt 0 28 if BWarea gt 30 if blackratio BWarea gt 2 5 If blackoverwhite lt 0 251 Copy the segmentation part from the region image crop to its position in 12 output image End End End End 54 Arabic Typed Text Recognition in Graphics Images ATTR GI 3 2 3 2 Result Image for Display This result is just used as a displaying image that selects text region on original image Therefore the appropriate way is take a copy from RGB image and converted to black and white image This helps to present red color clearly Then multiply red white image with black white image Algorithm 3 16 Show Post Processing Input I is a colored image Output Img is output image for show 1 I2 convert to binary of I 13 preprocessing I 3 I3 and 14 is binary image 14 postprocessing 1
122. ut ABBYY Finereader 11 Exclusive details about ABBYY FineReader 11 are given here How to use it and what are the results if we try to test Arabic graphics image eS Set Program Access and Defaults eS Windows Catalog Windows Update 5 Programs Accessories I Microsoft Office Fe Documents gt 9 eDocFile gt 3 LEADTOOLS Main EVAL 17 5 gt fe Settings gt aa ABBYY FineReader 11 Quick Tasks Search gt A ABBYY Screenshot Reader ABBYY Screenshot Reader captures a rectangular area on o Help and Support your screen and coverts it into a text table or image 7 Run 0 Shut Down Figure A 1 to run ABBYY Screenshot Reader After install ABBYY FineReader 11 you found it in programs window And to try it you can choose ABBYY ScreenShot Reader or Quick Tasks depends on the kind of use you need This program you can apply it by take a screenshot which like as print screen function to get an image from any part you need on your screen as you see in figure A 3 and A 5 or you can browse the image place and select the image file as you see in figure A 10 A 14 and A 16 At first we select an ABBYY ScreenShot Reader to apply it on two web sites one of them is English and the other is Arabic to see if it is work or not and how will the results form To run this program see figure A 1 ABBYY Screenshot Reader Capture ij Area Language English Sen
123. y are depend for document collection to Linguistic Data Consortium LDC these documents contains Romanize style They tried to find a manual treatment solution to develop a corpus of bi tonal images Volker et al in 32 discussed the printed and handwritten Arabic words Evaluation methods to select a best optical character recognition OCR depends on the best quality They made a comparison between different published recognition systems for Arabic handwritten There are different measurements to evaluate the recognition systems such as testing depends on a large dataset and complexity that solving diverse tests Second measurement the recognition rate is a global parameter hardly significant for system computer development Finally the quality is not enough measure based on the output recognizer but the quality of zoning and segmentation into words or characters represents an important feature of recognition system We take these measurements into account to help us forgive a decision about accuracy and performance for our approach All of the above mentioned related work focused on scanned documents with high resolution more than or equal to 200 dpi and used variant techniques and methods to extract the text from white background with black text and they give a good result The aim of the above work is either to create a new document from the degraded document or to measure the similarity of handwritten to achieve certain symptom

Arabic Typed Text Recognition in Graphics Images (ATTR-GI)

Contents

Download Pdf Manuals

Related Search

Related Contents