Offline Handwritten Gurmukhi Character Recognition: Study ... - Khilare

0 downloads 0 Views 393KB Size Report
handwritten Bangla script recognition. Bansal and Sinha (2000) have proposed a technique for Devanagari script recognition. In this technique, they have ...
Offline Handwritten Gurmukhi Character Recognition: Study of Different Feature-Classifier Combinations Munish Kumar

R. K. Sharma

Computer Science Department School of Mathematics and Computer Panjab University Constituent College, Applications Muktsar, Punjab, India Thapar University Patiala, Punjab, India +919465409800, +919872319157 +919872202705 [email protected]

[email protected]

ABSTRACT Offline handwritten character recognition (OHCR) is the method of converting handwritten text into machine processable layout. Since late sixties, efforts have been made for offline handwritten character recognition throughout the world. Principal Component Analysis (PCA) has also been used for extracting representative features for character recognition. In order to assess the prominence of features in offline handwritten Gurmukhi character recognition, we have recognized offline handwritten Gurmukhi characters with different combinations of features and classifiers. The recognition system first sets up a skeleton of the character so that significant feature information about the character can be extracted. For the purpose of classification, we have used k-NN, Linear-SVM, Polynomial-SVM and RBF-SVM based approaches. In present work, we have collected 7,000 samples of isolated offline handwritten Gurmukhi characters from 200 different writers. The set of basic 35 akhars of Gurmukhi has been considered here. A partitioning policy for selecting the training and testing patterns has also been experimented in present work. We have used zoning feature; diagonal feature; directional feature; intersection and open end points feature; transition feature; parabola curve fitting based feature and power curve fitting based feature extraction technique in order to find the feature set for a given character. The proposed system achieves a recognition accuracy of 94.8% when PCA is not applied and a recognition accuracy of 97.7% when PCA is applied.

Keywords OHCR, Feature extraction, Classification, PCA, k-NN, SVM

1. INTRODUCTION These days, we are being greatly influenced by computers and almost all the vital processing is done automatically. Keeping this in mind, it becomes essential that transfer of data between the human beings and computers should be easy and quick. Handwriting recognition offers a methodology for improving the interface between human beings and computers as it permits computers to read and process handwritten documents. This recognition gives a significant benefit to bridge the Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. DAR '12, December 16 2012, Mumbai, IN, India Copyright 2012 ACM 978-1-4503-1797-9/12/12 $15.00.

M. K. Jindal Department of Computer Science and Applications Panjab University Regional Centre, Muktsar, Punjab, India +919814637188, +919779351188

[email protected]

communication gap between man and machine. Handwritten Character Recognition (HCR) is a fast growing field in the pattern recognition research. HCR can either be online or offline. In online handwriting recognition, data is captured during the writing process with the help of a special pen and an electronic surface. In offline handwriting recognition, documents are scanned images of prewritten text, generally, on a sheet of paper. Offline handwriting recognition is different from online handwriting recognition, because here, stroke information is not accessible. The research work is continuing in character recognition field since late sixties. It is still an active area of research as the problem concerned is complex in nature. Researchers have proposed different solutions for recognition of characters for Indian scripts also. It is worth mentioning that most of the published work on Indian scripts recognition deals with printed documents and there are a few articles that deal with handwritten character recognition problem for Indian scripts. To mention a few, Pal and Chaudhury (1994) have presented an OCR for handwritten Bangla script recognition. Bansal and Sinha (2000) have proposed a technique for Devanagari script recognition. In this technique, they have recognized the characters in two steps. In the first step, they recognize the strokes and in the second step they recognize the character based on strokes recognized in the first step. Garain et al. (2002) have achieved an accuracy of 96.3% for Bangla handwritten character recognition. Roy et al. (2004) have offered a system towards Indian postal automation for sorting of postal documents. They have employed a two-stage Multi Layer Perceptron (MLP) based classifier to recognize Bangla and Arabic numerals. Joshi et al. (2004) have presented a Tamil handwritten recognition system. They have employed subspace and DTW based classifiers for recognition. Deepu et al. (2004) have presented a system based on principal component analysis for online handwritten character recognition. Bhattacharya et al. (2006) have proposed a scheme for Bangla character recognition. They have obtained features by computing local chain code histograms of input character shape. They have achieved recognition accuracy of about 94.6% and 92.1% for training and testing, respectively. Bhowmik et al. (2006) have proposed a novel HMM for handwritten Oriya numerals recognition. Pal et al. (2007) have proposed an offline handwritten Oriya script recognition system. They have extracted curvature features for recognition and obtained an accuracy of about 94.6% from a few offline handwritten Oriya samples. Sundaram and Ramakarishnan (2008) have presented 2D-PCA for online Tamil character recognition. Basu et al. (2009) have proposed a hierarchical approach for handwritten Bangla characters recognition. Pal et al. (2010) have presented a scheme for lexicon-

driven bi-lingual (English and Bangla) city name recognition for Indian postal automation. They have obtained 92.2% recognition accuracy when tested on 11,875 samples of the city names. Arora and Namboodiri (2010) have proposed a system for online handwritten Malayalam character recognition. Their system achieves a stroke level accuracy of about 97.8%. Rampalli and Ramakrishnan (2011) have discussed online and offline strategies for recognition of handwritten Kannada characters. Bhattacharya et al. (2012) have presented an efficient two stage approach for handwritten Bangla characters recognition. Lehal and Singh (2000) have developed a complete recognition system for printed Gurmukhi script, where connected components are initially segmented using thinning based approach. Jindal et al. (2009) have provided a complete recognition system for recognition of degraded printed Gurmukhi script documents. Some authors have also worked on handwritten Gurmukhi script recognition. Sharma et al. (2008) have presented an online handwritten Gurmukhi script recognition system. They have used elastic matching technique in which character is recognized in two stages. In the first stage, they recognize the strokes and in the second stage character is constructed on the basis of recognized strokes. Sharma and Jhajj (2010) have used zoning density based features for isolated handwritten Gurmukhi character recognition. They have used SVM and k-NN classifiers for classification and achieved a maximum recognition accuracy of 72.83% with RBF kernel of SVM classifier. Kumar et al. (2011) achieved a recognition accuracy of 94.29% while using intersection and open end points as features and SVM with polynomial kernel as classifier for offline handwritten Gurmukhi character recognition. This work is an attempt towards the recognition of offline handwritten Gurmukhi characters with different combinations of features and classifiers. This work can be extended to other Indian scripts that are structurally akin to Gurmukhi script.

2. GURMUKHI SCRIPT AND DATA COLLECTION Gurmukhi script is the script for Punjabi language. The word Gurmukhi has been derived from the Punjabi term “Guramukhi”, which means “from the mouth of the Guru”. Gurmukhi script is the twelfth most widely used script in the world. The character set of the Gurmukhi script is given in Figure 1.

signs and three half characters. The writing style of the Gurmukhi script is from top to bottom and left to right. In Gurmukhi script, there is no case sensitivity and most of the characters have a horizontal line at the upper part called headline and characters are connected with one another through this line. For the present work, we have collected 7,000 samples of isolated offline handwritten Gurmukhi characters from 200 different writers. These writers were requested to write thirty two consonants and three vowel bearers form the Gurmukhi character set. A sample of handwritten characters by 5 different writers (W1, W2, …, W5) is given in Figure 2.

Script Character

W1

W2

W3

W4

W5

ੳ ਅ ੲ ਸ ਹ Figure 2. Samples of Handwritten Gurmukhi characters.

3. HANDWRITTEN CHARACTER RECOGNITION SYSTEM Handwritten character recognition system consists of phases, namely, digitization, pre-processing, feature extraction, and classification. Block diagram of handwritten character recognition system is given in Figure 3. Handwritten Character

The Consonants

ਸਹਕਖਗਘਙਚਛਜਝਞਟਠਡਢਣਤਥਦਧਨਪਫਬਭ

Digitization

ਮਯਰਲਵੜ The Vowel Bearers Preprocessing

ੳਅੲ The Additional Consonants (Multi Component Characters)









ਗ਼ ◌਼ਲ

Feature extraction

The Vowel Modifiers

◌ੋ ◌ੌ ◌ੇ

◌ੈ ਿ◌ ◌ੀ

◌ਾ ◌ੁ ◌ੂ

Auxiliary Signs

◌ੱ

Classification

◌ੰ ◌ਂ

The Half Characters

◌੍ਹ ◌੍ਰ ◌੍ਵ Figure 1. Gurmukhi Script Character Set. Gurmukhi script has three vowel bearers, thirty two consonants, six additional consonants, nine vowel modifiers, three auxiliary

Recognized Character Figure 3. Block Diagram of Handwritten Character Recognition System.

3.1 Digitization Digitization is the process of converting paper based handwritten Gurmukhi script document into electronic form. Here, each document includes only one Gurmukhi character. The electronic conversion is accomplished by using a procedure whereby a document is scanned and an electronic representation of the original document, in the form of a TIFF image, is produced. Digitization produces the digital image which is fed to the preprocessing phase.

3.2 Preprocessing In this phase, the gray level character image is normalized into a window of size 100×100. After normalization, we produce bitmap image of the normalized image. Now, the bitmap image is transformed into a thinned image using parallel thinning algorithm (Zhang and Wang, 1984).

3.3 Feature Extraction In this phase, features from input characters are extracted. The performance of handwritten character recognition system, primarily, depends on the features, that are being extracted. The extracted features should be able to classify a character in a unique way. In this work, we have proposed two efficient feature extraction techniques, namely, parabola curve fitting based features and power curve fitting based features for offline handwritten Gurmukhi character recognition. These techniques have been compared with other recently proposed feature extraction techniques, namely, zoning features, diagonal features, directional features, transition features, and intersection and open end points features.

3.3.1 Parabola Curve Fitting Based Features Thinned image of a character is divided into n (= 100) zones. A parabola is then fitted to the series of foreground pixels in each zone, using the least square method. A parabola ‫ ܽ = ݕ‬+ ܾ‫ ݔ‬+ ܿ‫ ݔ‬ଶ is uniquely defined by three parameters a, b and c. As such, this will give 3n features for a given bitmap.

The steps that have been used to extract these features are given below. Step I: Divide the thinned image into n (= 100) number of equal sized zones. Step II: In each zone, fit a power curve using least square method and calculate the values of a and b. Step III: Corresponding to the zones that do not have a foreground pixel, set the value of a and b as zero.

4. PRINCIPAL COMPONENT ANALYSIS Principal component analysis (PCA) is a mathematical procedure. It uses the transformation to convert a set of observations of possibly correlated features into a set of values of uncorrelated features called principal components. PCA is a well-established technique for extracting representative features for character recognition and is used to reduce the dimension of the data. The technique is useful when a large number of variables do not include effective interpretation of the relationships between different features. By reducing the number of variables, one can interpret from a few features rather than a large number of features. The number of principal components is less than or equal to the number of original variables. By selecting top j eigen vectors with larger eigen values for subspace approximation, PCA can provide a lower dimensional representation to expose the underlying structures of the complex data sets. Let us consider that there are P features for handwritten character recognition system. In the next step, the symmetric matrix S of covariance between these features is calculated. Now, the eigen vectors ܷ௜ (݅ = 1, 2, … , ܲ) and the corresponding eigen values ∆௜ (݅ = 1, 2, … , ܲ) are calculated. From these P eigen vectors only j eigen vectors are chosen corresponding to the larger eigen values (also called as principal components). An eigen vector corresponding to higher eigen value describes more characteristic features of a character. Using these j eigen vectors, feature extraction is done by using PCA. In the present work, seven features for a Gurmukhi character have been considered and the experiments have been conducted by taking 2, 3, 4, 5, 6 and 7 principal components.

5. CLASSIFICATION

Figure 4. Parabola Curve Fitting Based Feature Extraction. The steps that have been used to extract these features are given below. Step I: Divide the thinned image into n (= 100) number of equal sized zones. Step II: For each zone, fit a parabola using least square method and calculate the values of a, b and c (Figure 4). Step III: Corresponding to the zones that do not have a foreground pixel, set the values of a, b and c as zero.

3.3.2 Power Curve Fitting Based Features The thinned image of a character is again divided into n (= 100) zones. A power curve is fitted to the series of foreground pixels in every zone using the least square method. A power curve of the form ‫ ݔܽ = ݕ‬௕ is uniquely defined by two parameters a and b. These parameters will give 2n features for a given bitmap.

Classification phase is the decision making phase of an offline handwritten character recognition engine. This phase uses the features extracted in the previous stage for making the class membership. In the present work, we have used k-NN and SVM classifiers for character recognition. In the k-NN classifier, Euclidean distances from the candidate vector to the stored vector are computed. Here, we have considered the value of k is 3.The Euclidean distance between a candidate vector and a stored vector is given by, ே

݀ = ඩ ෍(‫ݔ‬௞ − ‫ݕ‬௞ )ଶ ௞ୀଵ

Here, N is the total number of features in the feature set, ‫ݔ‬௞ is the library stored feature value and ‫ݕ‬௞ is the candidate feature value. Other classifier, SVM used in this work is also a very useful technique for data classification. The SVM is a learning machine, which has been widely applied in pattern recognition. It is based on supervised learning. In supervised learning, a machine is trained instead of programmed to perform a given task on a number of input/output pairs. SVM classifier has also been considered with three different kernels, namely, linear kernel, polynomial kernel, and RBF kernel in the present work. In this

work, C-SVC type classifier in Lib-SVM tool has been used for classification purpose.

6. EXPERIMENTAL DISCUSSION

RESULTS

AND

In this section, the results of different combinations of features and classifiers based recognition systems for offline handwritten Gurmukhi characters are presented. As stated earlier, we have also experimented a few partitioning strategies while using a classifier. We have divided the complete data set using five partitioning strategies. In the first strategy (strategy a), we have taken 50% data in training set and other 50% data in the testing set. In the second strategy (strategy b), we have considered 60% data in training set and remaining 40% data in the testing set. Strategy c has 70% data in training set and 30% data in testing set. Similarly, strategy d has 80% data in training set and 20% in testing set. Strategy e is formulated by taking 90% data in training set and remaining 10% data in testing set. Category-wise experimental results for Non-PCA and PCA based recognition system are presented in the following sub-sections.

Figure 6. Non-PCA Experimental Results Based on SVM with Polynomial Kernel Classifier.

6.1.3 Recognition classifier

results

based

on

RBF-SVM

In this sub-section, experimental results of partitioning strategies (a, b,…, e), based on SVM with the RBF kernel classifier are illustrated (Figure 7). One can note that power curve fitting based features and SVM with the RBF kernel, achieved a maximum recognition accuracy of 84.2% when we used strategy c.

6.1 Non-PCA Based Experimental Results In this section, we have considered each feature individually. The features that have been considered here are the zoning features, diagonal features, directional features, transition features, intersection and open end points features, straight line fitting based features, parabola curve fitting based features, and power curve fitting based features. Classifier wise experimental results are presented in the following sub-sections.

6.1.1 Recognition results based on Linear-SVM classifier In this sub-section, experimental results of partitioning strategies (a, b,…, e), based on SVM with linear kernel classifier, are presented (Figure 5). One can see that a maximum recognition accuracy of about 88.5% has been achieved with parabola curve fitting based features when we used strategy e and SVM with the linear kernel.

Figure 7. Non-PCA Experimental Results Based on SVM with RBF Kernel Classifier.

6.1.4 Recognition results based on k-NN classifier Experimental results of partitioning strategies (a, b,…, e), based on the k-NN classifier are presented (Figure 8). It has been noted that power curve fitting based features, with k-NN classifier, achieved the maximum recognition accuracy of 94.8% when we used strategy a.

Figure 5. Non-PCA Experimental Results Based on SVM with Linear Kernel Classifier.

6.1.2 Recognition results based on Polynomial-SVM classifier In this sub-section, experimental results of partitioning strategies (a, b,…, e), based on SVM with polynomial kernel classifier, are demonstrated (Figure 6). It has been observed that parabola curve fitting based features make it possible to achieve a recognition accuracy of 81.1% when we used strategy e and SVM with the polynomial kernel.

Figure 8. Non-PCA Experimental Results Based on k-NN Classifier.

6.2 PCA Based Experimental Results In further experimentation, PCA has been applied on the set of features extracted that are considered in this work. The results of this study are presented in this section. Features considered in section 3.3 have been explored in this section using PCA. For the sake of comparisons between the performance of principal components, two principal components (2-PC), three principal components (3-PC), …, seven principal components (7-PC) have been considered to be taken as input to the classifiers. Classifier wise experimental results are presented in the following subsections.

6.2.1 Recognition results based on Linear-SVM classifier In this sub-section, number of principal components wise recognition results of partitioning strategies (a, b,…, e), based on SVM with linear kernel classifier have been presented. One can see that a maximum recognition accuracy of about 94.5% has been achieved with 2-principal components when we use strategy e and SVM with the linear kernel. These results are shown in Figure 9.

6.2.4 Recognition results based on k-NN classifier Here, results of partitioning strategies (a, b,…, e), based on the kNN classifier with different principal components are presented (Figure 12). It has been again seen that 2-principal component with k-NN classifier achieved the maximum recognition accuracy of 97.7% when we used strategy e.

Figure 12. PCA Experimental Results Based on k-NN Classifier.

Figure 9. PCA Experimental Results Based on SVM with Linear Kernel Classifier.

6.2.2 Recognition results based on Polynomial-SVM classifier In this sub-section, experimental results of partitioning strategies (a, b,…, e), based on SVM with polynomial kernel classifier have been illustrated. It has been observed that 3principal component makes it possible to achieve a recognition accuracy of 94.5% when we use strategy e and SVM with the polynomial kernel. These results are depicted in Figure 10.

Figure 10. PCA Experimental Results Based on SVM with Polynomial Kernel Classifier.

6.2.3 Recognition classifier

results

based

on

RBF-SVM

Experimental results of partitioning strategies (a, b,…, e), based on SVM with RBF kernel classifier have been presented (Figure 11). It has been seen that 2-principal components and SVM with the RBF kernel, achieved a maximum recognition accuracy of 87.8% when we used strategy e.

7. CONCLUSION AND FUTURE SCOPE The work presented in this paper analyses the performance of offline handwritten Gurmukhi character recognition system. This system is experimented by applying the PCA on extracted features and by considering the features as such, without applying PCA. The features of a character that have been considered in this work include zoning features, diagonal features, directional features, transition features, straight line fitting based features, intersection and open end points features. In present work, we have also proposed two efficient feature extraction techniques for offline handwritten Gurmukhi character recognition. The classifiers that have been employed in this work are k-NN, Linear-SVM, Polynomial-SVM and RBF-SVM. We have analyzed that PCA based offline handwritten Gurmukhi character recognition system performs better than non-PCA based offline handwritten Gurmukhi character recognition system as it has been noted that without using PCA, the maximum recognition accuracy of 94.85% could be achieved and with the use of PCA, the maximum recognition accuracy of 97.71% was achieved. This accuracy can possibly be increased by considering a combination of classifiers and by considering a larger data set while training the classifiers. This work can also be extended for offline handwritten character recognition of other Indian scripts. We can also used other techniques like Multi-factor dimensionality reduction, Multilinear subspace learning, Non-linear dimensionality reduction for reduce the dimension of data and to extract representative features for character recognition.

8. ACKNOWLEDGEMENTS Authors are highly thankful to the reviewers for their valuable suggestions that helped in improving this manuscript.

9. REFERENCES

Figure 11. PCA Experimental Results Based on SVM with RBF Kernel Classifier.

[1]

Almuallim, H. and Yamaguchi, S. 1987. A method of recognition of Arabic cursive handwriting. IEEE Transactions on Pattern Analysis and Machine Intelligence. 9, 5, 715-722.

[2]

Arora, A. and Namboodiri, A. M. 2010. A hybrid model for recognition of online handwriting in Indian scripts. In Proceedings of the 12th International Conference on Frontiers in Handwriting Recognition (ICFHR). 433-438.

[3]

Bansal, V. and Sinha, R. M. K. 2000. Integrating knowledge sources in Devanagari text recognition system. IEEE

Transaction on Systems, Man and Cybernetics. 30, 4, 500505. [4]

[5]

[6]

[7]

[8]

[9]

Basu, S., Das, N., Sarkar, R., Kundu, M., Nasipuri, M., and Base, D. K. 2009. A hierarchical approach to recognition of handwritten Bangla characters. Pattern Recognition, 42, 7, 1467-1484. Bhattacharya, U., Shridhar, M., Parui, S. K., Sen, P. K., and Chaudhuri, B. B. 2012. Offline recognition of handwritten Bangla characters: an efficient two-stage approach. Pattern Analysis and Applications. 15, 4, 445-458. Bhattacharya, U. and Chaudhuri, B. B. 2009. Handwritten numeral databases of Indian scripts and multistage recognition of mixed numerals. IEEE Transactions on Pattern Analysis and Machine Intelligence. 31, 3, 444-457. Bhattacharya, U., Shridhar, M., and Parui, S. K. 2006. On recognition of handwritten Bangla characters. In Proceedings of ICVGIP. 817-828. Bhowmik, T. K., Parui, S. K., Bhattacharya, U. and Shaw, B. 2006. An HMM based recognition scheme for handwritten Oriya numerals. In Proceedings of 9th International Conference on Information Technology (ICIT). pp. 105-110. Deepu, V., Madhvanath., S., and Ramakrishnan, A. G. 2004. Principal Component Analysis for online handwritten character recognition. In Proceedings of 17th International Conference Pattern Recognition. 2, 327-330.

[10] Dutta, A. and Chaudhury, S. 1993. Bengali alpha-numeric character recognition using curvature features. Pattern Recognition. 26, 12, 1757-1770. [11] Garain, U., Chaudhuri, B. B., and Pal, T. T. 2002. Online handwritten Indian script recognition: A human motor function based framework. In Proceedings of 16th International Conference on Pattern Recognition (ICPR). 3, 164-167. [12] Jindal, M. K., Lehal, G. S., and Sharma, R. K. 2009. On segmentation of touching characters and overlapping lines in degraded printed Gurmukhi script. International Journal of Image and Graphics (IJIG). 9, 3, 321-353. [13] Joshi, N., Sita, G., Ramakrishnan, A. G., and Madhvanath., S. 2004. Tamil handwriting recognition using subspace and DTW based classifiers. In Proceedings of ICONIP. 806813. [14] Kumar, M., Sharma, R. K., and Jindal, M. K. 2011. SVM based offline handwritten Gurmukhi character recognition. In Proceedings of SCAKD. 758, 51-62. [15] Lehal, G. S. and Singh, C. 2000. A Gurmukhi script recognition system. In Proceedings of 15th International Conference on Pattern Recognition (ICPR). 2, 557-560.

[16] Pal, U. and Chaudhuri, B. B. 1994. OCR in Bangla: an Indo-Bangladeshi language. In Proceedings of 12th International Conference on Pattern Recognition (ICPR). 269-273. [17] Pal, U., Wakabayashi, T. and Kimura, F., 2007. A system for off-line Oriya handwritten character recognition using curvature feature. In Proceedings of 10th International Conference on Information Technology (ICIT). 227-229. [18] Pal, U., Roy, R. K., and Kimura, F. 2010. Bangla and English city name recognition for Indian postal automation. In Proceedings of 20th International Conference on Pattern Recognition (ICPR). 1985-1988. [19] Rampalli, R. and Ramakrishnan, A. G. 2011. Fusion of Complementary Online and Offline Strategies for Recognition of Handwritten Kannada Characters. Journal of Universal Computer Science. 17, 81-93. [20] Roy, K., Vajda, S., Pal, U., and Chaudhuri, B. B. 2004. A system towards Indian postal automation. In Proceedings of 9th International Workshop on Frontiers in Handwriting Recognition (IWFHR). 361-367.

[21] Sharma, A., Kumar, R., and Sharma, R. K. 2008. Online handwritten Gurmukhi character recognition using elastic matching. In Proceedings of Congress on Image and Signal Processing. 391-396.

[22] Sharma, D. V. and Jhajj, P. 2010. Recognition of isolated handwritten characters in Gurmukhi script. International Journal of Computer Applications. 4, 8, 9-17. [23] Singh, B., Mittal, A. and Ghosh, D. 2011. An Evaluation of Different feature extractors and classifiers for offline handwritten Devanagri character recognition. Journal of Pattern Recognition Research. 2, 269-277. [24] Sundaram, S. and Ramakrishnan, A. G. 2008. Two Dimensional Principal Component Analysis for Online Tamil Character Recognition. In Proceedings of 11th International Conference Frontiers in Handwriting Recognition. 88-94. [25] VijayKumar, B. and Ramakrishnan, A. G. 2004. Radial basis function and subspace approach for printed Kannada text recognition. In Proceedings of IEEE ICASSP. 5, 321324. [26] Zhang, T. Y. and Suen, C. Y. 1984. A fast parallel algorithm for thinning digital patterns. Communications of the ACM. 27, 3, 236-239.