Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
www.ijcsit-apm.com International Journal of Computer Science &Information Technology 1
IJCSIT, Vol. 1, Issue 2 (April 2014...
International Journal of Computer Science & Information Technology 2 www.ijcsit-apm.com
A. Handwritten Manipuri digits
Fro...
www.ijcsit-apm.com International Journal of Computer Science &Information Technology 3
Figure 4. Finding decision hyperlan...
International Journal of Computer Science & Information Technology 4 www.ijcsit-apm.com
Real parts of filters Magnitudes o...
www.ijcsit-apm.com International Journal of Computer Science &Information Technology 5
6 120 103 17 85.83
7 120 116 4 96.6...
International Journal of Computer Science & Information Technology 6 www.ijcsit-apm.com
Support Vector Machine, Decision-T...
Upcoming SlideShare
Loading in …5
×

RECOGNITION OF CHEISING IYEK/EEYEK-MANIPURI DIGITS USING SUPPORT VECTOR MACHINES

536 views

Published on

This paper describes the recognition of Cheising Iyek-Manipuri digits; handwritten as well as printed and comparison of recognition accuracy using Support Vector Machines (SVM). The paper also presents the steps starting right from binarization of scanned images in pre-processing till recognition of the digits using a trained model.

Published in: Engineering, Technology
  • Be the first to comment

  • Be the first to like this

RECOGNITION OF CHEISING IYEK/EEYEK-MANIPURI DIGITS USING SUPPORT VECTOR MACHINES

  1. 1. www.ijcsit-apm.com International Journal of Computer Science &Information Technology 1 IJCSIT, Vol. 1, Issue 2 (April 2014) e-ISSN: 1694-2329 | p-ISSN: 1694-2345 RECOGNITION OF CHEISING IYEK/EEYEK-MANIPURI DIGITS USING SUPPORT VECTOR MACHINES Kansham Angphun Maring1 , Dr. Renu Dhir2 1, 2 Department of Computer Science and Engineering National Institute of Technology, Jalandhar, Punjab, India 1 kam10maring@gmail.com, 2 dhirr@nitj.ac.in Abstract-The development of handwriting recognition systems began in the 1950s when there were human operators whose job was to convert data from various documents into electronic format, making the process quite long and often affected by errors. Automatic text recognition aims at limiting these errors by using image pre-processing techniques that increases the speed and precision to the entire recognition process. This paper describes the recognition of Cheising Iyek-Manipuri digits; handwritten as well as printed and comparison of recognition accuracy using Support Vector Machines (SVM). The paper also presents the steps starting right from binarization of scanned images in pre-processing till recognition of the digits using a trained model. Gabor filter-based technique is used for feature extraction. The experiment is carried out with image size 14x10 using MATLAB. An overall accuracy of 89.58% and 98.45% is achieved for handwritten and printed respectively. Index Terms- Meitei/Meetei Mayek, Cheising Iyek/Eeyek, Binarisation, Gabor filter, Feature Extraction, Support Vector Machines (SVM), Support Vectors. I. INTRODUCTION During the last few decades, pattern recognition has got lot of attentions from the researchers with its vast practical applications [1] in science and technology viz; Number Plate Recognition, Bank Cheque Processing, Postal Automation Service, Conversation of Ancient Manuscripts etc. Meitei/Meetei Mayek is the script used in Manipuri language which is primarily written and spoken by the valley people of Manipur, India and the language so spoken is called Meiteilon [2, 3], Meiteiron and Meithei [4]. The exact origin of the Meetei Mayek is shrouded in mystery by the destruction of Pre-Hindu places and the burning of all the historical documents called the „Puya Mei Thaba‟ [5]. The current Manipuri script is a reconstruction of the ancient Manipuri script. Meitei Mayek is a member of the Tibeto-Burman branch of the Sino-Tibetan language family and is also spoken in Bangladesh and Mynmar[5]. This script contains Iyek Ipee/Mapung Iyek, which have 27 alphabets (18 original plus 9 derived letters called Lom Iyek), Lonsum Iyek (8 letters), Cheitek Iyek (8 symbols), Khudam Iyek (3 symbols), Cheising Iyek (10 numeral figures) [6]. All the Meetei Mayek numerals are derived from the embryonic development stages of a human foetus [5]. The script was officially approved by theGovernment of Manipur on April 22, 1980 [6] but it got its place in academics only in 2005 replacing Bengali script. This is the reason that research in Manipuri script recognition has not yet been widely introduced to the research community while much research on other scripts of different languages has been published and introduced internationally. Meetei Mayek has now been included in the Unicode Standard, Version 5.2 which was released on 1st October 2009 [7]. The range is ABC0- ABFF. The rest of the paper is organized with literature survey in Section II followed by details of image acquisition in Section III. Then image pre-processing steps in Section IV and Support Vector Machine in Section V. Section VI describes the feature extraction technique using Gabor filter and the recognition system in Section VII. Experimental results are shown in Section VIII and conclusion is drawn in Section IX. Figure 1. Cheising Iyek (Manipuri digits) and English digits II. LITERATURE SURVEY Digit recognition systems have been developed for different languages in the world. In context to Indian languages, one can find recognition of off-line handwritten Gujarati Digits using Neural Network Approach in [8]. In [9] handwritten Devnagari Digit Recognition: Benchmarking on New Dataset is proposed using Neural Network. Then in [10] and [11] Handwritten Gurmukhi Numeral Recognition using Zone-basedHybrid Feature Extraction Techniques and recognition system for handwritten Assamese numerals using mathematical morphology respectively is discussed. When it comes to Manipuri digits, one can find limited papers. In [12] simulation and modelling of handwritten MeiteiMayek Digits using Neural Network Approach is discussed. They have achieved 85% accuracy rate and mentioned that the accuracy can be improved by using different feature sets and classification algorithms like support vector machines (SVM). III. IMAGE ACQUISITION The image acquisition for handwritten as well as printed are described below:
  2. 2. International Journal of Computer Science & Information Technology 2 www.ijcsit-apm.com A. Handwritten Manipuri digits From different age groups and genders, around 800 samples for each digits are collected on A4 size paper.These samples are thenscanned using a scanner at 250dpi and save in .PNG format. Figure 2 shows a sample of the scanned copy which are then feed for pre-processing steps. Figure 2. A sample image of the scannedhandwritten Manipuri digits (0-9) B. Printed Manipuri digits Two different fonts of Meetei/Meetei Mayek namely “epaomayek” and “Rathayek” are used. The digits are written on MS Word with font size of 10 points. These digits are then captured using snipping tool and save in .PNG format. (a) (b) Figure 3. Cheising Iyek in different fonts containing digits from 0 to 9: (a) epaomayek (b) Rathayek. IV. IMAGE PRE-PROCESSING The image pre-processing can significantly increase the reliability of an optical inspection. Image pre- processing is a series of operations performed on the scanned input image that essentially enhances the image rendering it suitable for segmentation. Several filter operations which intensify or reduce certain image details enable an easier or faster evaluation. So, pre-processing steps differ by requirements. Image binarization (thresholdingis the conversion ofa gray-scale image into a binary image. Global thresholding picks one threshold value for the entiredocument image based on an estimation of the background level (I) Handwritten Manipuri digit „3‟. (II) Printed Manipuri digit „5‟. Figure 3. Pre-processing of Manipuri digits: (a). Segmented digit (b). Skeleton (c). Dilated Image (d). Smoothed Image. from theintensity histogram of the image. In this paper global thresholding technique using Otsu‟s method is applied. Since the digits are unconnected, explicit character segmentation technique is applied. Disk structuring element of radius 2 is applied for dilating the images. Before dilating, the images arethinned to infinity to find out the skeletons of the images. This is done so that when the images are dilated they will have equal thickness of strokes. V. SUPPORT VECTOR MACHINES The idea of Support Vector Machines (SVM) were first shared by Vapnik [13]. Support Vector Machine classifier is an algorithm which maximises the margin between the classes and tries to minimises the classification error. SVM used to identify a set of linearly separable hyperplanes which are linear functions of the feature space. Among the separable hyperplanes only one hyperplane is chosen and placed such that such that the distance between the classes is maximum. SVM algorithm for linearly non-separable classes is discussed here. Let us consider {Xi, Yi} for i=1,2,……..,N denoting the training dataset where Yi is the target output for training data Xi.The aim of SVM is to maximize the objective function L(𝛼) given by: L(𝛼)= 𝛼𝑖 - 1 2 𝛼𝑖 𝛼𝑗 𝑁 𝑗 =1 𝑁 𝑖=1 YiYjΦ(Xi)T Φ(Xj) (1) subject to constraints αi N j=1 Yj=0, 0≤ 𝛼𝑖 ≤C; ∀ i where C is the cost parameter that determines the cost caused by constraint violation, 𝛼𝑖 is the hyperparameter and Φ(·) is the feature mapping function. Asking for the maximum-margin linear separator in equation (1) leads to standard Quadratic Programming(QP) problems.With the mentioned constraints, the QP solution leads to the following classification function for support vector machines. Y=sgn(W. Φ(Z)+b) Y=sgn 𝛼𝑖 𝑞 𝑖=1 𝑌𝑖 𝑋𝑖 𝑍 + 𝑏 (2) where𝛼𝑖 is the lagrance multiplier assigned to each training data whose value depend on the role of training the data in the classifier system.The non-zero values of 𝛼𝑖 correspond to the support vectors that are used to construct the classifier in equation (2), „q‟denotes the number of support vectors. If the feature functions Φ(·) are chosen with care one can calculate the scalarproducts without actually computing all features, therefore greatly reducing
  3. 3. www.ijcsit-apm.com International Journal of Computer Science &Information Technology 3 Figure 4. Finding decision hyperlane function in SVM. the computationalcomplexity. In SVM the learning algorithms that only require dot products between the vectors in the originalinput space, and chooses the mapping such that these high-dimensional dot productscan be computed within the original space, by means of a kernel function is called “kernel trick”. K(x, xi) = 𝜑(x). 𝜑(xi) (3) In this experiment, linear kernel used. Linear kernel is given by K(x, y) = x · y; where no mapping is performed and only the optimal hyperplane is calculated. VI. FEATURE EXTRACTION USING GABOR FILTER It is well known that the performance of a digit- recognition system depends significantly on the features used. Selection of a feature extraction method is probably the single most important factor in achieving high recognition rate in digit recognition. The extracted features should be able to identify each digits uniquely and there should be large variations in the features for different digits. Gabor filters [14] have been used extensively in computer vision [15] and texture analysis for their excellent properties: optimal joint spatial/spatial frequency localization and ability to simulate the receptive fields of simple cells in the visual cortex [16]. These characteristics suggest that the Gabor-filter based feature seems to be similar to features extracted by humans and thus, may be effective in classifying digits. Shustorovich [17]applied Gabor filters to handwritten numerals and showed that Gabor-filter-based features are superior to pixel-based features.A two-dimensional Gabor filter is defined by: f(x,y,θk,λ,σx,σy)=exp − 1 2 𝑅1 2 𝜎 𝑥 2 + 𝑅2 2 𝜎 𝑦 2 ∗ 𝑒𝑥𝑝 𝑖 2𝜋𝑅1 𝜆 (4) whereR1= x cos 𝜃 𝑘+ ysin 𝜃 𝑘 and R2= -x sin 𝜃 𝑘 + ycos 𝜃 𝑘. λand𝜃 𝑘are the wavelength and orientation of the sinusoidal plane wave respectively. The values of R1 andR2 allows the Gabor to stretch in any direction defined by 𝜃. In order to prevent the occurrence of undesired effects at the image borders, the wavelength value should be smaller than on fifthof the input image size.σxandσy are the standard deviations of the Gaussian envelope along the x-axis and y-axis directions respectively which gives the spread in the axes of two-dimensional Gaussian. A rotation of the x-y plane by an angle 𝜃 𝑘will result in a Gabor filter at orientation 𝜃 𝑘. The value of 𝜃 𝑘is given by 𝜃 𝑘 = π (k-1) / m, k = 1,……,m, where m denotes the number of orientations. For instance when m = 4, the orientation θused are: 0° , 45° , 90° and 135° and are shown in figure 5 (a). A set of Gabor filters with 5 spatial frequencies and 8 distinct orientations making 40 different Gabor filters that is used in the experiment is presented in Figure. 6. The Gabor feature can be viewed as the response of the Gabor filter located at a sampling point. The response is obtained by convolving the filter with an image. Gabor filters extract the orientation-dependent frequency contents, i.e., edge like features, from as small an area as possible. Figure 5 shows how Gabor features are extracted using by varying orientations, with the value of λ fixed.For each sampling point, m Gabor features can be obtained for morientations. The feature vector is a 2-D matrix with features as rows and levels of the digits in first column. When Gabor filters are applied to each pixels of the image, the dimension of the filtered vector can be very large (proportional to the image dimension). So, it will lead to expensive computation and storage cost. To alleviate such problem and make the algorithm robust, Gabor features are collected at regular intervals of pixels. This technique is very useful for face recognition system where Gabor features are obtained only at the fiducial points [18]. (a) (b) Figure 5. Extracting Gabor features for handwritten Manipuri digit: (a) 6 (six) and (b) 7 (seven).
  4. 4. International Journal of Computer Science & Information Technology 4 www.ijcsit-apm.com Real parts of filters Magnitudes of filters Figure 6. 40 Gabor filters with 5 spatial frequencies and 8 different orientations used to extract features of handwritten as well as printed Manipuri digits. In order to recognize handwritten Chinese characters, eight orientations were needed [19]. But this is not true for all the scripts due to the differences in required number of strokes in writing a character, structure of the characters and writing styles. VII. RECOGNITION OF CHEISING IYEK/EEYEK For recognition of the Manipuri digits, SVM classifier is used. As we all know that SVM is binary classifier but it can be used as multiclass classifier in different ways. In this paper all-at-once multiclass classifier support vector machines is used. In all-at-once support vector machines for an n-class problem, the decision function for class iis defined by: Di(x) = 𝑊𝑖 𝑇 Φ(x) + bi , (5) whereWiis the weight vector for class i in the feature space, Φ(x)is the mapping function, and bi is the bias term. For class i data x to be correctly classified, Di(x) needs to be the largest among Dj(x) (j = 1,………..,n), namely, the following inequalities must hold: 𝑊𝑖 𝑇 Φ(x) + bi >𝑊𝑗 𝑇 Φ(x) + bj for j≠i,j= 1,………..,n. (6) All the digits are labelled with labels 0, 1, 2,…..,9. The feature vector for training and testing are created separately using the feature extraction technique discussed in section VI. For handwritten 600 and 120 samples for each digit are selected for training and testing respectively. In case of printed, 1000 samples are taken for training and 200 samples for testing. Approximately same number of datasets from the two fonts are used for training and testing. A classifier model is created using linear kernel and feature vector of training dataset as input. The testing feature vector is applied to the trained model and recognition is done finding out overall percentage accuracy, predicted labels, andprobability values. From these values experimental results are drawn. VIII. EXPERIMENTAL RESULTS The experiment was carried for image size 14x10. When a set of Gabor filters with 5 spatial frequencies and 8 distinct orientations (40 Gabor filters), the accuracy is at its highest point. Varying the number of filters, accuracies are compared. The experimental results are sectioned as follows: A. Handwritten Manipuri digits The Table I shows the confusion matrix. The diagonal element shows the number of digits that are correctly recognized. While the values in the rest of the rows shows the number of digits that were misclassified against the digit in column number 1. Consider digit 9(nine), 82 digits out of 120 were correctly recognized and misclassified seven times as 2, twice as 3, sixteen times as 6, once as 7, and twelve times as 8. TABLE I CONFUSION MATRIX OF HANDWRITTEN MANIPURI DIGITS 0(ZERO)- 9(NINE). 0 1 2 3 4 5 6 7 8 9 0 118 2 1 117 1 2 2 1 98 4 1 8 8 3 7 112 1 4 2 1 116 1 5 113 2 5 6 3 2 4 103 8 7 3 116 1 8 2 1 6 7 100 4 9 7 2 16 1 12 82 TABLE II PERCENTAGE ACCURACY OF RECOGNITION OF HANDWRITTEN MANIPURI DIGITS 0(ZERO)-9(NINE). Digits Attempts CR FR Accuracy (%) 0 120 118 2 98.33 1 120 117 3 97.50 2 120 98 22 81.67 3 120 112 8 93.33 4 120 116 4 96.67 5 120 113 7 94.16
  5. 5. www.ijcsit-apm.com International Journal of Computer Science &Information Technology 5 6 120 103 17 85.83 7 120 116 4 96.66 8 120 100 20 83.33 9 120 82 38 68.33 Table II shows the performance of the SVM model for all the testing digits. For each digits, 120 images are given for recognition and result is shown below along with accuracy for each digits. FR means false recognition and CR means correct recognition. From the table it is known that digit 9 has the lowest accuracy, 68.33% with 35 false recognitions and recognizing 14 times as 6. This drags down the overall accuracy of the recognition system to 89.58%. The graphical representation between correct classification (CR) and false recognition (FR) is shown in figure 7. When more Gabor filters are applied beyond require, the performance improves little at the cost of dramatically increasing computations. This is shown Table III below: With the increase in number of Gabor filters used the accuracy increases a little bit but at some point it saturates and drops down. This data may be different for different scripts because of difference in their strokes and styles. Figure 7. Graphical representation of CR and FR for handwritten Manipuri digits. TABLE III EFFECT OF NUMBER OF GABOR FILTERS IN ACCURACY AND EXECUTION TIME. No. of Gabor Filters Execution time (Min.) Accuracy of the recognition system (%) Training(600 samples) Testing(120 samples) 20 28.21 5.80 87.16 30 42.72 8.17 88.91 40 53.04 12.62 89.58 50 73.62 15.75 89.25 B. Printed Manipuri digits The experimental results for printed Manipuri digits are analysed in figure 8. For each digits 1000 samples, mixed digits of both fonts are taken for training and 200 samples are considered for testing. Each printed digits has higher accuracy than its counter digits in the handwritten. Digits „0‟, „1‟, „4‟, and„7‟has 100% accuracy rates. As in handwritten, digit „9‟got lowest accuracy at 89.5 %. Figure 8. Graphical representation of CR and FR for printed Manipuri digits. C. Comparison between Handwritten and Printed The accuracy for handwritten and printed digits are compared in the figure 9. Digit „0‟ has comparatively equal accuracy for handwritten as well as printed. Huge increased in recognition rate is observed for digits „2‟(two), „6‟(six), „8‟(eight) and „9‟(nine). Figure 9. Accuracy comparison between handwritten and printed Manipuri digits. IX. CONCLUSIONS In this paper Gabor filter-based feature extraction method is used for the recognition of handwritten Manipuri digits. The overall performance of the system for handwritten is 89.58% and that of printed is 98.45%. Some of the digits are very much similar to each other. The presence of blob and little change in stroke angle differentiates them. So, the recognition system can be improved by fusing blob detection technique to the proposed system. As per the printed digits is concerned, taking more fonts and trying to classify the digits of different fonts which are not in training will be interesting. In future Pairwise Support Vector Machines will be used for same feature extraction technique. To avoid unclassifiable regions in Pairwise
  6. 6. International Journal of Computer Science & Information Technology 6 www.ijcsit-apm.com Support Vector Machine, Decision-Tree-Based Support Vector Machine may be considered. REFERENCES [1]. CIA/DOE Partnership Program Proposal for FY99 (Sandia NationalLaboratories Proposal), 1998. [2]. Wangkhemcha Chingtamlen, A short history of Kangleipak (Manipur)part-II, Kangleipak Historical & Cultural Research Centre,Sagolband Thangjam Leirak,Imphal,2007. [3]. Ng.Kangjia Mangang, Revival of a closed account, a brief history of kanglei script and the Birth of phoon (zero) in the world of arithmetic and astrology, SanmahiLaining Amasung Punshiron Khupham (SalaiPunshipham),Lamshang,Imphal,2003. [4]. T.C.Hodson, The Meitheis, Low price publications, Delhi,1908. [5]. Neelakash Kshetrimayum,Meitei Mayek: The Ignored face, Diploma Project for Graduate Diploma Programme in Design, National Institute of Design, Ahmedabad, 2006. [6]. Government of Manipur, APPROVED MEITEI/MEETEI MAYEK, Approved script vide " Manipur Gazzette No 33 dated April 22, 1980 Annexure 1 to 5 (1/2/78-SS/E)". [7]. Unicode 5.2.0 [Online]. Available:{http://unicode.org/versions/Unicode5.2.0/} [8]. Avani R. Vasant, Sandeep R. Vasant, Dr. G. R. Kulkarani, Performance Evaluation of Different Image Sizes for Recognizing Offline Handwritten Gujarati Digits using Neural Network Approach, International Conference on Communication Systems and Network Technologies, 270-273 (2012). [9]. Rajiv Kumar, Kiran Kumar Ravulakollu, Handwritten Devnagari Digit Recognition: Benchmarking on new dataset, Journal of Theoritical and Applied Information Technology, Vol. 60 No.3 543- 555 (2014). [10]. Gita Sinha, Rajneesh Rani, Renu Dhir, Handwritten Gurmukhi Numeral Recognition using Zonebased Hybrid Feature Extraction Techniques, International Journal of Computer Applications, Vol. 47 No. 21 24-29 (2012). [11]. Medhi, K, Kalita, S.K., Recognition of assamese handwritten numerals using mathematical morphology, Advance Computing Conference (IACC), IEEE International, 1076-1080 (2014). [12]. Romesh Laishram, Angom Umakanta Singh, N.Chandrakumar Singh, A.Suresh Singh, H.James, Simulation and Modeling of Handwritten Meitei Mayek Digits using Neural Network Approach, Proc. of the Intl. Conf. on Advances in Electronics, Electrical and Computer Science Engineering — EEC, 2012. [13]. Support Vector Machine [Online]. Available: {http://en.wikipedia.org/wiki/Support_vector_machine#History/Sup port_Vector_Machine}. [14]. D. Gabor, Theory of communication, J. Inst. Electr. Engng. 93, 429- 459 (1946). [15]. M. Porat and Y. Y. Zeevi, The generalized Gabor scheme of image representation in biological and machine vision, IEEE Trans. Pattern Analysis Mach. Intell. 10, 452-468 (1998). [16]. J. G. Daugman, Uncertainity relation for resolution in space, spatial frequency and orientation optimized by two-dimensional visual cortical filters, J. Opt. Soc. Am. A 2 (7) (1985) 1. [17]. A. Shustorovich, A subspace projection approach to feature extraction: two-dimensional Gabor transform character recognition, Neural Networks 7(8), 1295-1301 (1994). [18]. Yousara BEN JEMAA, Sana KHANFIR, Automatic local Gabor feature extraction for face recognition, (IJCSIS) International Journal of Computer Science and Information Security, Vol. 3, No. 1, 2009. [19]. Xuewen Wang, Xiaoqing Ding, Chansong Liu, Gabor filters-based feature extraction for character recognition, Pattern Recognition, 38 (2005) 369-379. AUTHORS First Author – Kansham Angphun Maring: Received his bachelor‟s degree in Computer Science and Engineering from North Eastern Regional Institute of Science and Technology (NERIST), Arunachal Pradesh, India. He is currently pursuing his master‟s degree in Computer Science and Engineering at Dr. B.R Ambedkar National Institute of Technology, Jalandhar, India. His area of interest are Digital Image Processing, Natural Language Processing, Computer Networks and Wireless Sensor Networks. Second Author–Dr. RenuDhir: Associate Professor in Department of Computer Science and Engineering at Dr. B.R Ambedkar National Institute of Technology, Jalandhar, India. Her area of interest includes Image Processing, Pattern Recognition, Natural Language Processing and Machine Learning. .

×