Name card image enhancement and restoration for text recognition

NAME CARD IMAGE ENHANCEMENT AND RESTORATION FOR TEXT RECOGNITION ZHANG JIE NATIONAL UNIVERSITY OF SINGAPORE 2003 NAME CARD IMAGE ENHANCEMENT AND RESTORATION FOR TEXT RECOGNITION ZHANG JIE (B.Sc.(Hons.), NUS) A THESIS SUBMITTED FOR THE DEGREE OF MASTER OF SCIENCE DEPARTMENT OF COMPUTER SCIENCE NATIONAL UNIVERSITY OF SINGAPORE 2003 Name: Zhang Jie Degree: Master of Science Department: Computer Science Thesis Title: Name card image enhancement & restoration for text recognition Abstract Photo taking has become a new input means for text recognition due to its high efficiency and little training required However, photos are not directly ready for OCR software to recognize because three problems are seriously affecting the quality of photos; they are complex background, blur and motion blur The research project focuses on the development of methodologies to preprocess photos such that OCR performance will be improved Various methods are proposed to enhance or restore photos; they are proven to provide promising results Keywords: Background Removal, Sharpening, Motion Deblurring, Enhancement, Image Restoration, Text Recognition Image Acknowledgement I would like to thank first, and foremost, A/P Tan Chew Lim This project would never have been completed if not for his time and efforts in guiding me in research and thesis writing I would like to thank Dr Ioannis Rekleitis, who very kindly share with me his program code so that I can smoothly start my own research work based on his master thesis Special thanks to Ms Wang Ying for her great assistance in the Fourier transform and highpass filters -i- Table of Contents ACKNOWLEDGEMENT i TABLE OF CONTENTS ii SUMMARY iv LIST OF FIGURES v LIST OF TABLES vii CHAPTER INTRODUCTION 1.1 1.2 1.3 1.4 Introduction to Hotcard ……………………………………………………… Motivations of the Current Research ………………………………………… Thesis Structure ……………………………………………………………… Definition of Terms ………………………………………………………… CHAPTER BACKGROUND STUDY & LITERATURE REVIEW …………… 10 2.1 Background Removal ………………………………………………………… 2.2 Image Enhancement – Sharpening …………………………………………… 2.2.1 Image Enhancement 2.2.2 Sharpening in Spatial Domain ……………………………………… 2.2.3 Sharpening in Frequency Domain …………………………………… 2.3 Motion Deblurring …………………………………………………………… 2.3.1 Motion Blur 2.3.2 Existing Works of Motion Deblurring for a Single Image …………… 10 13 13 15 17 20 21 26 CHAPTER ALGORITHM DESIGN & IMPLEMENTATION ……………… 34 3.1 Background Removal ……………………………………………………… 3.1.1 Analysis of Name Card Photos ……………………… …… 3.1.2 Design of Algorithm 3.2 Image Sharpening 3.2.1 Spatial Domain Sharpening 3.2.2 Frequency Domain Sharpening 3.3 Precise Motion Blur Estimation 3.3.1 Algorithm Expectations 3.3.2 Angle Estimation 3.3.3 New Finding in Angle Estimation 3.3.3 Distance Estimation - ii - 34 34 37 40 41 42 44 45 50 53 57 3.3.4 Angle Re-estimation 64 3.3.5 Distance Re-estimation and Iteration 66 CHAPTER EXPERIMENTAL RESULTS & EVALUATION ………………… 69 4.1 Background Removal 4.1.1 Experiment Design 4.1.2 Experimental Results & Discussions 4.2 Image Sharpening 4.2.1 Experiment Design 4.2.2 Experimental Results & Discussions 4.3 Motion Deblurring 4.3.1 Performance of Distance Estimation Algorithm 4.3.2 Performance of Motion Deblurring Algorithm 4.3.3 Motion Blur Estimation for Real-Life Images 69 69 70 72 72 73 77 78 80 83 CHAPTER CONCLUSION ……………………………………………………… 5.1 Research Conclusions 5.2 Summary of Contributions 5.3 Future Works 88 88 89 90 REFERENCES 93 - iii - Summary Name card photos are not directly ready for text recognition as they are affected by three problems, namely, complex background, blur and motion blur This research work aims to preprocess name card photos in such a way that the OCR software is able to recognize them better Techniques of document image analysis are widely studied and explored Methods are proposed to tackle each problem Background removal algorithm is based on the pixel value difference between the card and the background Name card photos are studied A card surface normally looks to be of a uniform color, which has a short range in pixel values Because the card occupies most of the space in a photo, we can identify this short range by looking at the image histogram Four borders of a name card are therefore located by comparing pixel values Two categories of filters can be used to sharpen images, i.e., spatial domain filters and frequency domain filters The objective is to identify the optimal one that offers the best OCR performance A variety of filters (with parameters) are applied to sharpen a large set of photos OCR software is used to recognize filtered photos; filters’ performances are compared using precision and recall Our objective is to find the filter with the highest precision and recall Motion blur is characterized by motion angle and distance If they are estimated precisely, Wiener filter can perfectly restore images We apply Rekleitis’ angle estimation algorithm and develop a new distance estimation algorithm As large errors are present in the angle estimation, the detected angle and distance have to be re-estimated in order to get precise estimations The number of iterations is found, which is where performance no longer improves even if more iterations are added In addition, we have found a persistent relationship between correct angles and angles reported by Rekleitis’ algorithm This relationship is used to improve the accuracy rate Research problems are analyzed, and methods are presented The experimental results confirm the effectiveness of the proposed methods, and the research objective has been achieved Recommendations are given for future works - iv - List of Figures Figure 1.1 Scan pen (left) and name card scanner (right) Figure 1.2 A name card photo with complex background Figure 1.3 A motion-blurred name card photo Figure 1.4 Restored results from precise estimation of PSF (60 degrees and 10 pixels) Figure 1.5 Ringing effects when a wrong PSF is applied to deblur a photo Figure 2.1 Extract complex objects, such as people, from background (Gordon et al., 1999) Figure 2.2 An original check image (left) and its filtered version (right) Figure 2.3 Basic steps for filtering in the frequency domain (Gonzalez & Woods , 2002, page 159) Figure 2.4 Two spatial domain sharpening filters Figure 2.5 The masks to implement high-boost filter, A ≥ (Gonzalez & Woods , 2002, page 133) Figure 2.6 Examples of definitions of motion angle and distance Figure 2.7 Image before restoration (left) and after restoration (right) (Yitzhaky and Kopeika, 1997) Figure 2.8 The outline of Rekleitis’ algorithm (Rekleitis, 1995, page 22) Figure 3.1 The structure of a name card photo Figure 3.2 (a) A name card photo; (b) Histogram of (a) Figure 3.3 (a) the original image; (b) the outcome of the background removal algorithm Figure 3.4 (a) A clear image Figure 3.4 (b) The synthetic blurred image Figure 3.4 (c) The deblurred photo using the correct motion angle and distance Figure 3.4 (d) Ringing effects caused by one pixel’s error in distance -v- Figure 3.4 (e) Ringing effects caused by four degrees’ error in angle Figure 3.4 (f) Ringing effects caused by hybrid errors in both angle and distance Figure 3.5 An image that is blurred by pixels Figure 3.6 An image that is blurred by 20 pixels Figure 3.7 Relationship between correct angles and responses from Rekleitis’ algorithm Figure 3.8 (a) The deblurred photo using degree and pixels Figure 3.8 (b) The deblurred photo using degree and pixels Figure 3.8 (c) The deblurred photo using degree and pixels (correct) Figure 3.8 (d) The deblurred photo using degree and pixels Figure 3.8 (e) The deblurred photo using degree and 10 pixels Figure 3.9 (a) the edge of the correctly deblurred image – Figure 3.8 (c) Figure 3.9 (b) the edge of the wrongly deblurred image – Figure 3.8 (e) Figure 3.10 An example of distance estimation Figure 3.11 An example of angle re-estimation, where the global minimum is at m=-2 Figure 4.1 (a) the original name card: color of desk is very similar to that of card surface; (b) the algorithm outcome: desk is recognized as part of name card Figure 4.2 Lights reflect from desk to card surface Figure 4.3 Image sharpened by the filter in Figure 2.4 (b) Figure 4.4 Image sharpened by high-boost filter using the filter in Figure 2.5 (a) with A=1.7 Figure 4.6 (a) the original real-life motion-blurred image; (b) the deblurred image (c) the sharpened deblurred image Figure 4.5 Image sharpened by ideal highpass filter with D0=80 Figure 5.1 The overall image of the system - vi - List of Tables Table 2.1 Experimental results of Rekleitis’ algorithm (1995, page 62) Table 4.1 Background estimation results Table 4.2 OCR performance for various filters (with parameters) Table 4.3 Comparison of distance estimation accuracy rates, given correct angle estimation Table 4.4 Comparison of distance estimation accuracy rates, without given correct angle estimation Table 4.5 Accuracy rates of the overall algorithm after nth iteration Table 4.6 OCR performance comparisons for real-life images - vii - Experimental Results & Evaluations three iterations take about 120 seconds; four iterations take about 150 seconds; five iterations take about 180 seconds If we are more concerned about the accuracy rate, we should stop at the 4th iteration, which offers the best accuracy rate, 84.42%, with a processing time of 150 seconds On the other hand, if the accuracy is not the only concern and we also care about the time taken by the algorithm, we should stop at the 2nd or 3rd iteration, whose accuracy rate is still reasonable, but with less processing time 4.3.3 Motion Blur Estimation for Real-Life Images The experiments in Section 4.3.2 are to blur a photo with a known PSF and to recover this PSF Given a precisely-estimated PSF, Wiener filter can perfectly deblur the photo; OCR software is able to recognize the deblurred photo very well However, we want to know how this algorithm performs on real-life images, whose PSF is unknown to us As PSF is unknown, we cannot directly check whether estimated angle and distance are correct or not Instead, OCR results are used to measure the effectiveness of algorithm For deblurred photos, we further apply the high-boost filter that is found in Section 4.2 To be most accurate, we run to the 4th iteration Procedures are listed below: Use OCR software to recognize an original blurred photo M1; record down its precision p1 and recall r1 Apply high-boost filter on M1, and resulting photo is M2 Use OCR software to recognize M2; record down its precision p2 and recall r2 Apply our algorithm on M1 and estimate motion angle a4 and distance d4 - 83 - Experimental Results & Evaluations Apply Wiener filter with a4 and d4 to deblur M1, and resulting photo is M3 Use OCR software to recognize M3; record down its precision p3 and recall r3 Apply high-boost filter on M3, and resulting photo is M4 Use OCR software to recognize M4; recode down its precision p4 and recall r4 The usefulness of algorithm is measured by the enhancements of precision and recall Six real-life name card photos are tested The second and third columns of Table 4.6 are precision and recall of original photos The next two columns are results when the original photos directly go through by the high-boost filter without being deblurred The fifth and sixth columns are precision and recall for photos that are deblurred by the overall motion deblurring algorithm Deblurred photos are further filtered by the highboost filter found in Section 4.2.2, and their precision and recall are shown in the last two columns Photo No Original Original Precision Recall p1 r1 Average 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% Filtered Filtered Precision Recall p2 r2 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% Deblurred Deblurred Deblurred Precision Recall & Filtered p3 r3 Precision p4 46.67% 52.18% 66.67% 54.74% 65.98% 69.29% 59.26% 5.79% 38.40% 42.86% 36.11% 49.23% 41.90% 35.72% 57.53% 84.17% 71.68% 84.92% 61.91% 73.08% 72.21% Table 4.6 OCR performance comparisons for real-life images - 84 - Deblurred & Filtered Recall r4 34.71% 80.80% 52.60% 71.21% 51.54% 55.24% 57.68% Experimental Results & Evaluations (a) (b) - 85 - Experimental Results & Evaluations (c) Figure 4.6 (a) the original real-life motion-blurred image; (b) the deblurred image; (c) the sharpened deblurred image The original photos lead to recognition failure as certain characteristics of motion blur are not tolerated by the OCR software at all Even if they are sharpened by the high-boost filter, the precision and recall are still 0% Once they are deblurred using the estimated results from our algorithm, the average precision increases from 0% to 59.25%, and the average recall increases from 0% to 35.72% Sharpening with the high-boost filter further enhances the OCR performance, which is shown in the last two columns As we can see, real-life photos are not deblurred as good as synthetic ones One possible reason is that velocity changes during exposure time Previously when we apply a PSF to blur a photo, it is assumed that motion is caused by a uniform motion velocity with a constant speed and direction But this may not always be true for real-life photos - 86 - Experimental Results & Evaluations Accelerated motion is a special kind of motion blur, depending on initial velocity, acceleration and exposure time Som (1971) explains this using LSF (line spread function) LSF ( x ) = t e ( v + 2ax ) / 2 (4-1) where a is acceleration, v0 is initial velocity, and te is exposure time Extensive work needs to be carried out to investigate this issue, which is beyond the scope of the present thesis It is included as one of our future research directions - 87 - Conclusions Conclusions 5.1 Research Conclusions Cameras are being used more and more widely as a new input mechanism for text recognition, but three problems are seriously downgrading OCR performance We propose a few algorithms to preprocess photos so as to make them more suitable for OCR software Algorithms have been implemented, and their performances have been measured Experimental results confirm the validity and effectiveness of the proposed algorithms A new method has been developed to estimate name card background areas by identifying card’s surface color range Various sharpening filters have also been applied to a large number of name card photos; the filter, which produces the best OCR outcomes, has been found Finally, we have developed a method to precisely estimate motion blur parameters with a high accuracy rate Photos can then be satisfactorily restored; OCR performance is significantly improved The algorithms preprocess the images for OCR, and they are integrated into the overall system as: - 88 - Conclusions Identify the name card boundary in the name card image and remove the background from the images Estimate motion blur angle and distance, and use Wiener filter to restore images Apply the sharpening filter to sharpen images By this step, all preprocessing of input images are done, and images are ready for OCR operations This process is presented in Figure 5.1 below: Background removal Motion deblurring Sharpening Input images OCR software Figure 5.1 The overall structure of the system 5.2 Summary of Contributions Previously, algorithms that are used to estimate motion blur can only estimate two motion parameters imprecisely Therefore, photos are only deblurred to some extent, and deblurred ones are not good enough for OCR software The distinguishing feature of our algorithm is that it can estimate motion blur parameters precisely with a high accuracy rate Once motion blur is estimated - 89 - Conclusions precisely, Wiener filter can perfectly restore photos, and OCR results are almost as good as scanner-quality document images OCR software is normally designed for scanned images Camera-taken photos are not sharp enough, especially around the area of character edges Spatial domain and frequency domain sharpening filters are established techniques to sharpen images But which filter is best for sharpening photos for OCR software is unknown We have applied a variety of filters to sharpen photos, then these photos are sent to OCR for recognition The filter providing the best recognition outcomes has been identified, which is the high-boost filter implemented by applying the filter in Figure 2.5 (a) with A=1.7 Complex background, such as pictures and texts at the back, should be removed before photos are sent for recognition No similar algorithms are seen to fulfill the same requirements Our algorithm first finds out the color range of the card surface and next uses this range to identify the background areas The algorithm has been proven to remove the background effectively in most cases 5.3 Future Works There are a lot of directions for future research works The first direction is to reduce the errors in angle estimation, which seriously affect distance estimation right now The relationship between the actual motion angles and the estimated angles (Section 3.3.3) could be further studied To reduce the complexity of the angle estimation algorithm is - 90 - Conclusions also worthwhile studying There are several complex issues in real-life motion, which has lots of rooms for research Currently when we apply the first part of Rekleitis’ algorithm to estimate motion angles, errors are sometimes as big as 15 degrees These errors are propagating into the distance estimation and sometimes lead to distance errors That is why we need to re-estimate motion angles and distances for a few iterations, which is both inefficient and prone to error If the angle estimation algorithm could be greatly enhanced to reach the stage of precise estimation, re-estimations procedures in the current algorithm would not be necessary because our distance algorithm can estimate very well when the correct is known Not only will this save processing time, but also this will increase the accuracy rate Further, the interesting relationship between the actual motion angle a and estimated angle a1 can be utilized as well to enhance the angle estimation accuracy The cause of this relationship might be the source of angle estimation errors, and so it is worthwhile studying the reason of its existence The algorithm complexity is another field where we could improve Currently angle estimation algorithm is of O(n Log n) complexity, where n is the number of pixels in the image Because name cards usually have a large size, e.g 640*480, angle estimation is time consuming Possible research directions include reducing the complexity and - 91 - Conclusions finding out whether using a part of a photo can be used to estimate angles as effectively as using the whole photo Motion blur in real-life images can be further studied Currently, we only assume a linear motion with a constant velocity for motion deblurring Our work has been shown to work generally well for real-life images as most users would usually try to take pictures with steady hands However, for inexperienced users or for photos taken in a hurry, severe motion deblurring may result They are a lot of unknown factors in severe motions, such as constant or variable acceleration, or complex motion paths Detailed studies of such factors may be carried out in future to find improved methods to deal with such motion blur - 92 - References References Aggarwal, J K and Nandhakumar, N “On the Computation of Motion from Sequences of Images – a Review.” Proceedings of the IEEE, 76(8):917 – 935, August 1988 Bascle, B., Blake, A and Zisserman, A “Motion Deblurring and Super-Resolution from an Image Sequence.” Proceedings of European Conference on Computer Vision, vol 2, pg 573-582, 1996 Cambridge, UK Black, M J and Anandan, P “A Framework for the Robust Estimation of Optical Flow.” IEEE (1993), 231 – 236 Bretzner, L “Multi-Scale Feature Tracking and Motion Estimation.” Ph.D Diss Stockholms University, 1999 Cannon, M “Blind Deconvolution of Spatially Invariance Image Blurs with Phase.” IEEE Trans Acoust Speech Signal Processing Vol 24, 1976 Capel, D and Zisserman, A “Super-Resolution Enhancement of Text Image Sequences.” International Conference on Pattern Recognition, Vol 1, Sep 2000 Barcelona, Spain Fabian, R and Malah, D “Robust Identification of Motion and Out-of-focus Blur Parameters from Blurred and Noisy Images.” CVGIP: Graphical Models and Image Processing, 53(5):403 – 412, Sep 1991 - 93 - References Franks, K and Koppen, M “Towards an Universal Approach to Background Removal in Images of Bankchecks.” Proceedings of 6th International Workshop on Frontiers in Handwriting Recognition, Aug 1998 Taejon, Korea Freeman, W T and Adelson, E H “The Design and Use of Steerable Filters.” IEEE Transaction on Pattern Analysis and Machine Intelligence, 13(9):891 – 906, Sep 1991 Gonzalez, R C and Woods, R E Digital Image Processing: 2nd edition Prentice-Hall, Inc., 2002 Gordon, G., Darrell, T., Harville, M., and Woodfill, J “Background Estimation and Removal based on Range and Color.” Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, June 1999 Fort Collins, Colorado Haralick, R and Shapiro, L Computer and Robot Vision, Addison-Wesley Publishing Company, 1992 Horn, B K P and Schunck, B G “Determining Optical Flow.” Technical report Massachusetts Institute of Technology, 1980 - 94 - References Hotcard Technology Pte Ltd, 2004 Jurong East Street 21, #05-30 IMM Building Singapore 609601 Available from: www.hotcardtech.com [14 Mar 2004] Jahne, B Digital Image Processing, 5th edition Springer-Verlag Berlin Heidelberg New York, 2002 Junker, M., Hoch, R and Dengel, A “On the Evaluation and Document Analysis Components by Recall, Precision, and Accuracy.” Proceedings of 5th International Conference on Document Analysis and Recognition, 713-716, 1999 India Lu, Y and Guo, H “Background Removal in Image Indexing and Retrieval.” Proceedings of the IEEE Computer Society Conference on Image Analysis and Processing, 1999 MathWorks Pte Ltd, 2004 Image Processing Toolbox Online Documentation Available from: http://www.mathworks.com/access/helpdesk/help/toolbox/images/images.html [31 Mar 2004] Petrou, M and Bosdogianni, P Image Processing: The Fundamentals John Wiley & Sons Ltd., 1999 Pratt, W K Digital Image Processing, 3rd edition John Wiley & Sons, Inc., 2001 - 95 - References Proakis, J G and Manolakis, D G Digital Image Processing: 2nd edition Macmillan Publishing Inc., 1992 Rav-Acha, A “Restoration of Multiple Images with Varying Motion Blur.” Diss Hebrew University of Jerusalem, 2001 Rekleitis, I M “Optical Flow Recognition from the Power Spectrum of a Single Blurred Image.” Proceedings of International Conference on Image Processing, Sep 1996 Lausanne, Switzerland Rekleitis, I M “Steerable Filters and Cepstral Analysis for Optical Flow Calculation from a Single Blurred Image.” Proceedings of Vision Interface, May 1996 Toronto 159 – 166 Rekleitis, I M “Visual Motion Estimation based on Motion Blur Interpretation.” Master Diss McGill University, 1995 Russ, J C The Image Processing Handbook CRC Press, Inc., 1992 Seeger, M and Dance, C “Binarising Camera Images for OCR.” IEEE (2001) Seul, M., O’Gorman, L., and Sammon, M J Practical Algorithms for Image Analysis: Description, Examples, and Code New York: Cambridge University Press, 2000 - 96 - References Som, S C “Analysis of the Effect of Linear Smear on Photographic Images.” J Opt Soc Am Vol 61, pg 859-864, 1971 Yitzhaky, Y and Kopeika, N S “Identification of Blur Parameters from Motion Blurred Images.” Proceedings of CVGIP: Graphical Models and Image Processing, Sep 1997 Denver, Colorado Yitzhaky, Y., Mor I., Lantzman, A., and Kopeika, N S “Direct Method for Restoration of Motion-Blurred Images.” J Opt Soc Am A/Vol 15, No 6, Jun 1998 - 97 - [...]... Hotcard Hotcard (Hotcard Technology Pte Ltd, 2004) is a pioneering local company in the development of compact sized, multilingual and high accuracy OCR for name card recognition Their cutting-edge products, Scan pen and name card scanner (Figure 1.1), are widely used by business people who collect a large amount of name cards every day They scan name cards, perform OCR operation, and transfer images... resolution is good enough -1- Introduction for text recognition; price is affordable for most people These enhancements in digital cameras offers an attractive input means – photo taking Instead of scanning name cards, users can use a camera to take photos of name cards and transfer them to OCR software for recognition In addition, many high-end PDAs, Pocket PCs, and mobile phones have a built-in digital... scanner for capturing images, and sophisticated software for analyzing images This software is named “OCR software” in the subsequent chapters Photo / Image: Both terms refer to a name card photo that is taken using a digital camera Fine Tune: It is to adjust parameters in a mathematical formula in order to achieve desired effects Precision and Recall: They are two common criteria to measure the performance... demonstrated by the comparisons of two images in Figure 2.2 - 11 - Background Study and Literature Review Figure 2.2 An original check image (left) and its filtered version (right) Lu and Guo (1999) design an algorithm to remove backgrounds for digital image libraries Previously, image features used for indexing and retrieval are global, and features are computed over the entire image The major drawback is that... removed before it is sent for image indexing and retrieval procedures This algorithm consists of three major steps: fuzzy clustering, color image segmentation, and background analysis It is tested on a variety of image collections: cars, crashed cars, face, and earth While appealing as they are, the above approaches cannot be directly applied to identify backgrounds for our problem Name card images have... of image enhancement is to modify images in a way that results are more suitable than original ones for a specific application As various applications have their own characteristics, techniques used to enhance images are different and application-oriented For example, an algorithm that is useful for enhancing medical images may not be able to enhance images that are taken by a space probe Image enhancement. .. read and process the background area Next, the OCR software might recognize the characters in a wordy background and organize them together with the card s contents Due to these two reasons, the background area needs to be removed before photos are sent for text recognition -3- Introduction Figure 1.2 A name card photo with complex background The second problem is that edges of character on name card. .. g '( x, y ) is the enhanced image that we obtain from highpass filters - 19 - (2-14) Background Study and Literature Review 2.3 Motion Deblurring Motion deblurring is a significant field of image restoration Image restoration (Gonzalez and Woods, 2002; Petrou and Bosdogianni, 1999) is the process in which images are improved using objective criteria and prior knowledge and changed to what they should... there are overlapping areas between image enhancement and restoration, the former improves images using subjective criteria in order to satisfy the human visual system or enhance machine testing results While image restoration is attempting to identify degradation functions and applying the inverse process to obtain the original image The degradation process (Gonzalez and Woods, 2002) is modeled in the... easy to differentiate a card from its background by simply not scanning its background But when a user takes a photo of a name card, he/she usually captures the name card along with some background in order not to miss out important information Figure 1.2 shows a name card with a complex background, which includes colorful and wordy areas Undesirable background leads to two issues for the OCR software .. .NAME CARD IMAGE ENHANCEMENT AND RESTORATION FOR TEXT RECOGNITION ZHANG JIE (B.Sc.(Hons.), NUS) A THESIS SUBMITTED FOR THE DEGREE OF MASTER OF SCIENCE DEPARTMENT... collect a large amount of name cards every day They scan name cards, perform OCR operation, and transfer images into editable text Figure 1.1 Scan pen (left) and name card scanner (right) These products... show a pair of images, before and after restoration (Figure 2.7) The original image is somewhat restored, but not completely Figure 2.7 Image before restoration (left) and after restoration (right)