MDP Feature Extraction Technique for Offline Handwritten ... - Khilare

Smart Computing Review, vol. 3, no. 6, December 2013

397

Smart Computing Review

MDP Feature Extraction Technique for Offline Handwritten Gurmukhi Character Recognition Munish Kumar1, M. K. Jindal2, and R. K. Sharma3 1

Computer Science Department, Panjab University Constituent College / Muktsar, Punjab / [email protected]

2

Department of Computer Science and Applications, Panjab University Regional Centre / Muktsar, Punjab

3

School of Mathematics & Computer Applications, Thapar University / Patiala, Punjab

* Corresponding Author: Munish Kumar

Received August 5, 2013; Revised October 25, 2013; Accepted November 1, 2013; Published December 19, 2013

Abstract: Character recognition is intricate work because of the various writing styles of different individuals. Most of the published work on handwritten character recognition problems deals with statistical features, and a few works deal with structural features, in general, and Gurmukhi script, in particular. In the present work, we propose a methodology for offline handwritten Gurmukhi character recognition by using a modified division points (MDP) feature extraction technique. We also compare this technique with other recently used feature extraction techniques, namely zoning features, diagonal features, directional features, intersection and open end points features, and transition features. To select a representative set of features is the most significant task for a character recognition system. After feature extraction, the classification stage makes use of the features extracted in the previous stage to recognize the character. In this work, we used linearsupport vector machines (linear-SVM), k-nearest neighbor (k-NN), and multilayer perceptron (MLP) classifiers for recognition. For experimental analysis, we used 10,500 samples of the isolated, offline, handwritten, basic 35 akhars of Gurmukhi script. The proposed system achieved a maximum recognition accuracy of 84.57%, 85.85% and 89.20% with linear-SVM, MLP and k-NN classifiers, respectively, with a five-fold cross validation technique. Keywords: Handwritten character recognition, feature extraction, classification, k-NN, SVM, MLP

DOI: 10.6029/smartcr.2013.06.001

398

Kumar et al.: MDP Feature Extraction Technique for Offline Handwritten Gurmukhi Character Recognition

Introduction

R

ecognition of handwritten characters is a complex task. The main difficulty encountered is the extreme cursiveness of written characters. Handwritten character recognition is divided into two categories: online and offline. In the online handwritten recognition category, data is captured throughout the course of writing with the aid of a special pen and an electronic surface. Offline documents are scanned images of a prewritten text, usually from a sheet of paper. An offline handwritten character recognition system consists of many stages, such as digitization, preprocessing, feature extraction and classification. Feature extraction is the most significant phase in a character recognition system and is used to measure the relevant shape contained in the character by extracting its features. The performance of a character recognition system depends on the features being extracted. Many researchers have reported work on handwritten character recognition problems over the last couple of years. Compared to non-Indic scripts, research on Indic scripts has not achieved excellence yet. There are many systems available for handwritten character recognition for non-Indic scripts, but little work has been done on handwritten character recognition systems for any Indic script, in general. A few attempts have been carried out on recognition of Devanagari, Bangla, Tamil and Oriya, which are discussed in the next section. This paper is organized into eight sections. In Section 2, we present related work. An introduction to Gurmukhi script and data collection for this work are described in Section 3. Section 4 presents the digitization and pre-processing stages of handwritten character recognition systems. The proposed feature extraction technique is illustrated in Section 5. Section 6 describes classification techniques. Section 7 shows some experimental results to prove the usefulness of this approach. Conclusions are included in Section 8.

Related Work Almuallim and Yamaguchi (1987) proposed a technique for cursively handwritten Arabic script recognition. In this technique, words are segmented into strokes, and these strokes are further classified by using the geometric and topological features. Dutta and Chaudhury (1993) presented a system for isolated Bangla alphabet and numeral recognition using curvature features. They used a feed-forward neural network for classification and achieved a recognition accuracy of about 90.0% and 85.0% for numerals and alphabets, respectively. Bhattacharya et al. (2002) presented a hybrid scheme for handprinted numeral recognition based on a self-organizing network and multilayer perceptron (MLP) classification techniques. Roy et al. (2004) presented a system for Indian postal automation to sort postal documents written in Arabic script. They extracted features based on black-pixel densities and the number of components inside a block. They obtained a maximum recognition accuracy of 92.1% for handwritten numeral recognition with a two-stage MLP classifier. Bhattacharya et al. (2006) proposed a scheme for Bangla character recognition. They obtained features by computing local chain code histograms of input character shapes. They achieved recognition accuracy of about 94.6% and 92.1% for training and testing, respectively. Sharma et al. (2006) proposed a scheme for offline handwritten Devanagari character recognition based on a quadratic classifier. They achieved an accuracy of about 98.9% and 80.4% for Devanagari numerals and characters, respectively, with a five-fold cross validation technique. The work done by Pal et al. (2007a) deals with recognition of offline handwritten Bangla compound characters using gradient features. They obtained a recognition accuracy of about 89.9% with a five-fold cross validation technique. Pal et al. (2007b) presented a system for offline handwritten Devanagari character recognition. They extracted directional information obtained from the arc tangent of the gradient. They achieved a recognition accuracy of about 94.2% with a five-fold cross validation test. Pal et al. (2007c) proposed an offline handwritten Oriya script recognition system. They extracted curvature features and obtained accuracy of about 94.6% from a few offline handwritten Oriya samples. Rajashekararadhya et al. (2008) extracted zoning features for offline handwritten numeral recognition of four widely used Indian scripts. They got recognition accuracy of about 98.6% for Kannada numerals with a SVM classifier. Alaei et al. (2010) proposed a two-stage scheme for isolated handwritten Persian character recognition. They extracted features based on modified chain code directional frequencies and employed an SVM for classification. They obtained 98.1% and 96.6% recognition accuracy with 8-class and 32-class problems, respectively. Desai (2010) presented a technique for Gujarati handwritten numeral recognition. In this work, the author used features abstracted from four different profiles of digits with a multilayered feed forward neural network, and achieved an approximate 82% recognition accuracy for Gujarati handwritten digit identification. Rampalli and Ramakrishnan (2011) discussed online and offline strategies for recognition of handwritten Kannada characters. Bhattacharya et al. (2012) presented an efficient two-stage approach for handwritten Bangla character recognition. A small number of handwritten character recognition systems are available for handwritten Gurmukhi script recognition. One of them is an online handwritten Gurmukhi script recognition system proposed by Sharma et al. (2008). Sharma and Jhajj (2010) extracted zoning features for handwritten Gurmukhi character recognition. They employed two classifiers, namely k-NN and SVM. They achieved a maximum recognition accuracy of about 72.5% and 72.0% with k-NN and SVM,


399

respectively. Kumar et al. (2011a) extracted intersection and open end points features for offline handwritten Gurmukhi character recognition. They used SVM for classification, taking 90% of the dataset as a training set and 10% of the dataset as a testing set. They achieved a maximum recognition accuracy of about 94.3%. They also extracted curvature features for offline handwritten Gurmukhi character recognition (2011b).

Gurmukhi Script and Data Collection Gurmukhi script is the most commonly used script for writing the Punjabi language and is derived from the old Punjabi term Guramukhi, which means “from the mouth of the Guru.” Gurmukhi script is the 12th most widely used script in the world, with a writing style from top to bottom and left to right that is not case sensitive. In Gurmukhi script, most of the characters have a horizontal line in the upper part, called a headline, and the characters are connected to each other with this line. For the present work, we collected 3,500 samples of isolated, offline handwritten Gurmukhi characters from 100 different writers. These writers were asked to write 32 consonants and 3 vowel bearers from the Gurmukhi character set.

Digitization and Preprocessing Digitization is the process of converting paper-based handwritten Gurmukhi script documents into electronic form. Here, each document included only one Gurmukhi handwritten character. The electronic conversion was accomplished by using a procedure whereby the document is scanned and an electronic representation of the original document in a tagged image file format was produced. Digitization produces the digital image, which is fed to the pre-processing phase. In this phase, the gray-level character image is normalized into a window 100×100. After normalization, we produced a bitmap image of the normalized image.

Proposed Methodology In this stage, the features of the input characters are extracted. The performance of a handwritten character recognition system primarily depends on features that are extracted. The extracted features should be able to uniquely classify a character. We propose a new feature extraction technique (a modified division points feature extraction technique) based on sub-division points of the character image. We also compare this technique with other recently used feature extraction techniques for character recognition, namely zoning features (Kumar et al., 2011c), diagonal features (Kumar et al., 2011a), directional features (Kumar et al., 2011c), intersection and open end points features (Kumar et al., 2011a), and transition features (Kumar et al., 2011d).To select a representative set of features is the most important task for a character recognition system. In the proposed technique, initially, we divide the character image into n (=100) zones, each of size 10×10 pixels. Let Img be the character image having 1’s for foreground pixels and 0’s for background pixels. The proposed methodology is based on subdivisions of the character image so that the resulting sub-images have balanced numbers of foreground pixels. Let be the vertical projection and be the horizontal projection of the particular zone as shown in Figure 1. = = Here, in , the division point of the array is 5 (the fifth element), because the sum of the left sub-array elements and the sum of the right sub-array elements is balanced, as much as possible, if we consider the fifth element as part of the left sub-array. Similarly, we calculate the division point of as 4 (the fourth element). The values of division points of each zone are stored as features in the feature vector. This will give us 2n features for a character image. The steps used to extract these features are given below. Step I: Step II: Step III: Step IV: Step V:

Divide the bitmap image into n (=100) zones, each 10×10 pixels. Find the horizontal projection profile and vertical projection profile in each zone of the bitmap image. Store the horizontal projection profile values in array H and vertical projection profile values in array V. After that, calculate the value of the division point of array H and the division point of array V based on subdivisions of the arrays so that the resulting sub-arrays have balanced numbers of foreground pixels. Consider the values of and in the left sub-array to make a possible balance between the left subarray and the right sub-array.

400


Step VI: Calculate the values of and for each zone and place them in the corresponding zone as its feature. Step VII: Corresponding to the zones that do not have a foreground pixel, the feature value is taken to be zero. Step VIII: Normalize the values of the feature vector by dividing each element of the feature vector by the largest value in the feature vector. These steps will give a feature set with 2n elements.

Figure 1. Bitmap of zone

Classification Classification is the decision-making phase of an offline handwritten character recognition engine. This phase makes use of the features extracted in the previous stage to decide class membership. In the present work, we used linear-SVM, k-NN and MLP classifiers because these classifiers are included in the top 10 classification techniques. SVM is an awfully helpful technique for data classification. SVM is a learning machine, which has been generally applied in pattern recognition. SVMs are based on statistical learning theory that uses supervised learning. In this work, the C-support vector classification type classifier in the Lib-SVM tool was used for SVM classification purpose. In the k-NN classifier, Euclidean distances from the candidate vector to the stored vector were computed. The value of k is taken to be 3. Considering learning and generalization abilities of MLP, MLP is also used in the present work for classification. A back propagation learning algorithm with learning rate (y) = 0.3 and momentum term (a) = 0.2 is used here for training of the MLP-based classifier. For developing a training set and a testing set for each of the classifiers employed in this work, the relevant dataset was segmented into a ratio of 4:1.

Evaluation and Experimental Results In this section, we describe recognition results of the proposed system for offline handwritten Gurmukhi character recognition. In this work, we used 3,500 samples of isolated offline handwritten Gurmukhi characters. The experimental results are based on different feature extraction techniques, namely zoning features (F1), diagonal features (F2), directional features (F3), intersection and open end points features (F4), transition features (F5), and modified division point features (F6). As mentioned above, for classification, we used three different classifiers, namely linear-SVM, k-NN and MLP classifiers. Here, we used five-fold cross validation for obtaining recognition accuracy. In general, r-fold cross validation divides the complete data set of each category into r equal subsets. Then, one subset is taken as testing data and the remaining r-1 subsets are taken as training data. From cross validation, each sample of the training data is predicted, and it gives the percentage of the correctly recognized testing dataset. We present classifier recognition results in sub-sections 7.1 to 7.3.

■ Recognition Results Based on Linear-SVM In this sub-section, recognition results of the linear-SVM classifier are presented. Using this classifier, we achieved an average recognition accuracy of 84.57% with the proposed feature extraction technique. As such, we determined that our proposed feature extraction technique achieves better recognition accuracy than other recently proposed feature extraction techniques. The recognition results of different features considered under this work are given in Table 1.


401

Table 1. Recognition results based on the linear-SVM classifier Linear-SVM Classifier Feature

Fold 1

Fold 2

Fold 3

Fold 4

Fold 5

Average

F1

63.57%

69.71%

66.14%

60.42%

70.00%

65.99%

F2

77.85%

78.86%

83.86%

72.00%

81.00%

78.71%

F3

53.28%

55.28%

56.00%

51.00%

60.00%

55.11%

F4

61.86%

67.71%

64.71%

58.14%

70.28%

64.54%

F5

54.00%

56.57%

55.71%

60.21%

62.00%

57.70%

F6

84.28%

85.14%

86.42%

79.86%

87.14%

84.57%

■ Recognition Results Based on k-NN Classifier In this sub-section, experimental results based on the k-NN classifier are presented. We see that the proposed features with the k-NN classifier achieved an average recognition accuracy of 89.20%. Recognition results based on the k-NN classifier are shown in Table 2. Table 2. Recognition results based on k-NN k-NN Classifier Feature

Fold 1

Fold 2

Fold 3

Fold 4

Fold 5

Average

F1

70.85%

69.42%

76.14%

63.71%

71.00%

70.22%

F2

86.71%

86.14%

88.00%

79.28%

88.14%

85.65%

F3

70.71%

74.57%

77.00%

67.71%

72.57%

72.51%

F4

77.14%

88.28%

79.42%

82.86%

72.14%

79.97%

F5

89.28%

83.14%

89.00%

78.14%

82.42%

84.40%

F6

91.14%

88.71%

94.71%

81.14%

90.28%

89.20%

■ Recognition Results Based on MLP In this sub-section, we present recognition results of different features considered in this work based on the MLP classifier. Using this classifier, we achieved an average recognition accuracy of 85.85% with the proposed technique (Table 3). Table 3. Recognition results based on MLP MLP Classifier Feature

Fold 1

Fold 2

Fold 3

Fold 4

Fold 5

Average

F1

86.00%

84.71%

85.14%

84.14%

59.57%

79.91%

F2

81.14%

82.00%

80.00%

80.86%

80.28%

80.86%

F3

53.14%

56.57%

56.14%

52.42%

56.28%

54.91%

F4

76.14%

63.71%

71.00%

77.85%

67.71%

71.28%

F5

79.57%

67.71%

80.85%

65.14%

78.71%

74.40%

F6

86.14%

86.85%

88.14%

84.71%

83.42%

85.85%

402


Conclusion In this paper, a new technique is presented for an offline handwritten Gurmukhi character recognition system that relies on structural features based on subdivisions of the image. The classifiers employed in the present work are k-NN, linear-SVM and MLP. We used 3,500 samples of isolated, offline handwritten Gurmukhi characters in this study and achieved a fivefold cross validation accuracy of 84.57%, 85.85% and 89.20% with linear-SVM, MLP and k-NN classifiers, respectively, with the proposed feature extraction technique. We saw that the proposed feature extraction technique performs better than other recently proposed techniques, as shown in Figure 2. We also determined that the k-NN classifier is the preeminent classifier in this study, as shown in Figure 3. Our future research will focus on applying the proposed features to word recognition, as well as combining them with other feature extraction schemes in order to further improve recognition performance. This work can be extended to offline handwritten character recognition of other Indian scripts. This technique can also be explored for other Indian scripts that are structurally similar to Gurmukhi script.

Figure 2. Recognition accuracy obtained by different features with various classifiers

Figure 3. Classifier recognition accuracy

References [1] H. Almuallim, S. Yamaguchi, “A method of recognition of Arabic cursive handwriting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 9, no. 5, pp. 715-722, 1987. Article (CrossRef Link) [2] A. Alaei, P. Nagabhushan, U. Pal, “A new two-stage scheme for the recognition of Persian handwritten characters,” in Proc. of 12th ICFHR, pp.130-135, 2010. Article (CrossRef Link) [3] U. Bhattacharya, T. K. Das, A. Datta, S. K. Parui, B. B. Chaudhuri, “A hybrid scheme for handprinted numeral recognition based on a self-organizing network and MLP classifiers,” International Journal for Pattern Recognition and Artificial Intelligence, vol. 16, no. 7, pp. 845-864, 2002. Article (CrossRef Link) [4] U. Bhattacharya, M. Shridhar, S. K. Parui, “On recognition of handwritten Bangla characters,” in Proc. of ICVGIP, pp. 817-828, 2006. [5] U. Bhattacharya, B. B. Chaudhuri, “Handwritten numeral databases of Indian scripts and multistage recognition of


[6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23]

403

mixed numerals,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 31, no. 3, pp. 444-457, 2009. Article (CrossRef Link) U. Bhattacharya, M. Shridhar, S. K. Parui, P. K. Sen, B. B. Chaudhuri, “Offline recognition of handwritten Bangla characters: an efficient two-stage approach,” Pattern Analysis and Applications, vol. 15, no. 4, pp. 445-458, 2012. Article (CrossRef Link) A. A. Desai, “Gujarati handwritten numeral optical character reorganization through neural network,” Pattern Recognition, vol. 43, no. 7, pp. 2582-2589, July 2010. Article (CrossRef Link) A. Dutta, S. Chaudhury, “Bengali alpha-numeric character recognition using curvature features,” Pattern Recognition, vol. 26, no. 12, pp. 1757-1770, 1993. Article (CrossRef Link) M. Kumar, R. K. Sharma, M. K. Jindal, “SVM based offline handwritten Gurmukhi Character Recognition,” in Proc. of SCAKD, pp. 52-63, 2011. M. Kumar, M. K. Jindal, R. K. Sharma, “Offline handwritten Gurmukhi character recognition using curvature feature,” in Proc. of AMOC, pp. 981-989, 2011. M. Kumar, M. K. Jindal, R. K. Sharma, “Classification of characters and grading writers in offline handwritten Gurmukhi Script,” in Proc. of ICIIP, pp. 1-4, 2011. Article (CrossRef Link) M. Kumar, M. K. Jindal, R. K. Sharma, “k-NN based offline handwritten Gurmukhi character recognition,” in Proc. of ICIIP, pp. 1-4, 2011. M. Kumar, M. K. Jindal, R. K. Sharma, “Offline handwritten Gurmukhi character recognition: study of different feature-classifier combinations,” in Proc. of IWDAR, 94-99, 2012. Article (CrossRef Link) U. Pal, T. Wakabayashi, F. Kimura, “Handwritten Bangla compound character recognition using gradient feature,” in Proc. of 10th ICIT, pp. 208-213, 2007. Article (CrossRef Link) U. Pal, N. Sharma, T. Wakabayashi, F. Kimura, “Off-line handwritten character recognition of Devanagari script,” in Proc. of 9th ICDAR, pp. 446-450, 2007. Article (CrossRef Link) U. Pal, T. Wakabayashi, F. Kimura, “A system for off-line Oriya handwritten character recognition using curvature feature,” in Proc. of 10th ICIT, pp. 227-229, 2007. Article (CrossRef Link) S. V. Rajashekararadhya, P. V. Ranjan, “Efficient zone based feature extraction algorithm for handwritten numeral recognition of four popular south Indian scripts,” Journal of Theoretical and Applied Information Technology, vol. 4 no. 12, pp. 1171-1181, 2008. R. Rampalli, A. G. Ramakrishnan, “Fusion of complementary online and offline strategies for recognition of handwritten Kannada characters,” Journal of Universal Computer Science, vol. 17, pp. 81-93, 2011. Article (CrossRef Link) K. Roy, S. Vajda, U. Pal, B. B. Chaudhuri, “A system towards Indian postal automation,” in Proc. of 9th IWFHR, pp. 361-367, 2004. Article (CrossRef Link) N. Sharma, U. Pal, F. Kimura, S. Pal, “Recognition of off-line handwritten Devanagari characters using quadratic classifier,” in Proc. of ICVGIP, pp. 805-816, 2006. A. Sharma, R. Kumar, R. K. Sharma, “Online handwritten Gurmukhi character recognition using elastic matching,” in Proc. of Congress on Image and Signal Processing, pp. 391-396, 2008. Article (CrossRef Link) D. V. Sharma, P. Jhajj, “Recognition of isolated handwritten characters in Gurmukhi script,” International Journal of Computer Applications, vol. 4, no. 8, pp. 9-17, 2010. Article (CrossRef Link) D. V. Sharma, U. Jain, “Recognition of isolated handwritten characters of Gurumukhi script using Neocognitron,” International Journal of Computer Applications, vol. 10, no. 8, pp. 10-16, 2010. Article (CrossRef Link)

Munish Kumar received his Masters degree in Computer Science & Engineering from Thapar University, Patiala, India in 2008. He started his career as an Assistant Professor in computer application at Jaito Centre of Punjabi University, Patiala. He is working as Assistant Professor in the Computer Science Department, Panjab University Rural Centre, Kauni, Muktsar, Punjab, India. He is currently pursuing his Ph.D. degree from Thapar University, Patiala, Punjab, India. His research interests include Character Recognition.

404

Kumar et al.: MDP Feature Extraction Technique for Offline Handwritten Gurmukhi Character Recognition M. K. Jindal received his Bachelors degree in science in 1996 and Post Graduate degree in Computer Applications from Punjabi University, Patiala, India in 1999. He holds a Gold Medal in his post graduation. He received his Ph.D. degree in Computer Science & Engineering from Thapar University, Patiala, India in 2008. He is working as Associate Professor in Panjab University Regional Centre, Muktsar, Punjab, India. His research interests include Character Recognition and Pattern Recognition.

R. K. Sharma received his Ph.D. degree in Mathematics from the University of Roorkee (Now, IIT Roorkee), India in 1993. He is currently working as Professor at Thapar University, Patiala, India, where he teaches, among other things, statistical models and their usage in computer science. He has been involved in the organization of a number of conferences and other courses at Thapar University, Patiala. His main research interests are statistical models in computer science, Neural Networks, and Pattern Recognition.

Copyright © 2013 KAIS