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.

40120140504011

209 views

Published on

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

  • Be the first to like this

40120140504011

  1. 1. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 80 ODIA HANDWRITTEN DIGIT RECOGNITION USING SINGLE LAYER PERCEPTRON Jagyanseni Panda1 , Manaswinee M. Panda1 , Aryapriyanka Samal1 , Niva Das1 1 Department of Electronics and Communication Engineering ITER, Sikha ‘O‘ Anusandhan University, Odisha, India. ABSTRACT The goal of handwriting recognition is to interpret the contents of the data and to generate a description of that interpretation in the desired format. The paper develops an efficient way for recognition of odia numerals by using a non-linear classifier. Here the gradient and curvature feature of odia numerals calculated. Then this processed through a non linear classifier for classification. The classification results using gradient and curvature are better than other method. This paper presents recognition of odia numeral using single layer perceptron where complexity is reduced and by using gradient method accuracy is more. Experimental results demonstrate that on a database of 100 digit patterns written by 100 different people perceptron based classifier which exhibits 85% accuracy. Keywords: Curvature Feature, Gradient Feature, Numeral Recognition, Neural Network, Principal Component Analysis. I. INTRODUCTION Over the years computerization has taken over large number of numeral operations, one such example is offline handwritten numeral recognition. Automatic handwritten odia digit recognition still remains a challenging task for a computer although a lot of research has been done on this topic. Recognition of hand written numeral or character is difficult because different people have different writing style and some characters closely resemble each other. Electronics media has recently gaining popularity because it replaces the paper and is fast to access [1] .So offline character recognition is an active research area. Automatic digit recognition system usually follows two steps: feature analysis and pattern classification. Feature extraction is that of extracting from the raw data the information which is most relevant for classification purposes, in the sense of minimizing the within-class pattern variability while enhancing the between class pattern variability. Considerable research work has been carried INTERNATIONAL JOURNAL OF ELECTRONICS AND COMMUNICATION ENGINEERING & TECHNOLOGY (IJECET) ISSN 0976 – 6464(Print) ISSN 0976 – 6472(Online) Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME: www.iaeme.com/ijecet.asp Journal Impact Factor (2014): 7.2836 (Calculated by GISI) www.jifactor.com IJECET © I A E M E
  2. 2. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 81 out in this area and various methods have been proposed for the classification of handwritten digits. The methods include principal component analysis (PCA) for feature reduction. India is a multi- lingual multi-script country and there are twenty two languages. Among Indian scripts, Devnagri, Tamil, Odia and Bangla have started to receive attention for OCR related research in the recent years [2, 3]. The basic features of offline character recognition are: • Flexibility: it should recognize a large number of character patterns. • Efficiency: it should be efficient. • Automated Learning: it should have automatic learning capability. Online Adaptability: it should have the capability to gather new knowledge of different writer-specific handwritten patterns as it operates. In 2009 Subhadip Basu et al. proposed a hierarchical approach to recognition of handwritten Bangla character. Though the bangla characters contain the matras. The matra hierarchy has three zones; i. e. the upper zone, middle zone and lower zone. The character segments obtained from the said three portions are recognized separately through appropriate pattern classifiers. Literature reports some work related to recognition of handwritten digits of Indian scripts [4],[5],[6],[7]. Technique for recognition of handwritten Hindi numerals based on modified exponential membership function fitted to fuzzy sets is presented in [8]. In this paper we have proposed single layer perceptron for classification. In the first stage of recognition the feature selection is done. Gradient features and curvature feature have been considered for this purpose. Then these feature are used to train the single layer perceptron network. The paper is organised in 5 Sections, Section 2 presents a brief overview of data collection and feature extraction method. Section 3 deals with the classification task using two single layer ANN structures. The simulation results are given in section 4. Section 5 presents the conclusion. II. DATA SET AND FEATURE EXTRACTION Data set of odia handwritten numerals 0 and 2 is created by collecting the handwritten documents from different people. Data collection is done on a sheet specially designed for data collection. Fig.1. General Architecture of handwritten character recognition system
  3. 3. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 82 The feature extraction process consists of procedures for gradient calculation, curvature feature calculation, feature vector generation and dimension reduction of the feature vector. Each procedure is described in succeeding subsections. 2.1 Calculation of Gradient A gray scale image is generated from an input binary image, and the gradient is calculated as described below. (1) Size normalization is applied to a binary character image so that the image has standard width and height. (2) Mean filter of size 3×3 is repeatedly applied r times to obtain a gray scale image. (3) The gray scale image is normalized so that the mean and the maximum of the gray scale are 0 and 1, respectively. (4) Roberts filter [9], [10] given by Eq. (1) and Eq.(2) is applied to each pixel g(i; j) of the normalized image to calculate the gradient. ∆‫ݑ‬ ൌ ݃ሺ݅ ൅ 1, ݆ ൅ 1ሻ െ ݃ሺ݅, ݆ሻ ∆‫ݒ‬ ൌ ݃ሺ݅ ൅ 1, ݆ሻ െ ݃ሺ݅, ݆ ൅ 1ሻ Direction:ߠሺ݅, ݆ሻ ൌ ‫݊ܽݐ‬ିଵ ሺ ∆௩ ∆௨ ) (1) Strength: ݂ሺ݅, ݆ሻ ൌ √∆uଶ ൅ ∆vଶ (2) 2.2 Calculation of Curvature The procedure for curvature calculation, based on bi-quadratic interpolation is described below. Fig. 2 neighbourhood of a pixel x0. The curvature c at x0 in a gray scale image is defined by; ( )32 '1 '' y y c + = (3) Where y=g(x) is equi-gray scale curve passing through x0, (x, y) is the spatial coordinates of x0, y’ and y’’ are the first and the second order derivatives of y, respectively. The derivatives y’ and y’’ are derived from bi-quadratic interpolating surface for the gray scale values in the 8- neighborhood of x0. The bi-quadratic surface is given by; x4 x3 x2 x5 x0 x1 x6 x7 x8 j i
  4. 4. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 83 [ ]                     = 2 222120 121110 020100 2 1 1 y y aaa aaa aaa xxz Then the equi-gray scale curve passing through x0 is given by ( ) ( ) 000010 2 20 0111 2 21 2 02 2 22 =−++ +++++ faxaxa yaxaxayaxa (4) Differentiation of both sides of Eq. (4) by x leads to ( ) ( ){ } ( ) 0111 2 210212 2 22 10201120 2 22 2 222 ' axaxaaxaxay axayaxayxa y +++++ +++++− = (5) Substituting the coordinates (0, 0) of x0 to (5), the value of y’ at x0 is given by 0110 /' aay −= (6) Similarly, the value of y’’ at x0 is given by 01 3 2001 2 1110010210 2 /)(2'' aaaaaaaay +−−= (7) Solving the simultaneous linear equations (3) holding for 8-neighbor of x0, the coefficients of the bi- quadratic surface are given by ܽଵ଴ ൌ ሺ݂ଵ െ ݂ହሻ/2 ܽଶ଴ ൌ ሺ݂ଵ ൅ ݂ହ െ 2݂଴ሻ/2 ܽ଴ଶ ൌ ሺ݂ଷ ൅ ݂଻ െ 2݂଴ሻ/2 ܽ଴ଵ ൌ ሺ݂ଷ െ ݂଻ሻ/2 ܽଵଵ ൌ ሺ݂ଶ െ ଼݂ሻ െ ሺ݂ସ െ ݂଺ሻ/4 (8) The coefficients a10 and a20 are, respectively, the first and the second order partial derivatives of f(x, y) regarding x; a01 and a02 are similar partial derivatives regarding y, and a11 is the one regarding to x and y. Substituting Eqs. (6) And (7) to Eq. (2), the curvature is given by ‫ܥ‬ ൌ െ2ሺܽଶ ଴ଵ ܽ଴ଶ െ ܽ଴ଵܽଵ଴ܽଵଵ ൅ ܽଶ ଴ଵܽ଴ଶሻ/ሺܽଶ ଵ଴ ൅ ܽଶ ଴ଵሻଷ/ଶ (9) 2.3 Generation of Feature Vector A feature vector is composed of the strength of gradient accumulated separately in different directions as described below; (1) The direction of gradient detected is quantized to 32 levels with గ ଵ଺ interval. (2) The normalized character image is divided into 81 (9 horizontal × 9 vertical) blocks. (3) The strength of the gradient is accumulated separately in each of 32 directions, in each block, to produce 81 local spectra of direction.
  5. 5. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 84 (4)The variable transformation (y = x^0.4) is applied to make the distribution of the features Gaussian-like [11], [12]. 2.4 Dimension Reduction The required processing time and storage can be reduced by the dimension reduction employing the principal component analysis (KL transform). The principal component analysis is a typical dimension reduction procedure based on the orthonormal transformation which maximizes the total variances, and minimizes the mean square error due to the dimension reduction. It is shown that the dimensionality can be reduced to without sacrificing the recognition accuracy in handwritten numeral recognition employing the feature vector of size 400 detected from the gradient of the gray scale[13],[14]. III. CLASSIFICATION ANN has been considered as one of the nonlinear classifiers for digit recognition. The most common architecture of ANN is the multi layer feed forward network. But due to its multi layered structure, the complexity is more as compared to other single layer feed forward networks. Hence in this paper we have proposed a single layer perceptron based scheme for the recognition task. 3.1 Single Layer Perceptron Structure The basic structure of a single layer perceptron is given in Fig.3. It can be viewed as a single "neuron" with multiple inputs that generates an output signal. Fig.3: Single layer perceptron network (10) The value of this output depends on the relative strengths of weighted input signals. The perceptron output can be expressed as: (11) where, is the adaptive weight vector is the input signal vector, and b is the bias term. The most commonly used activation functions are sigmoid & hard limiter. The perceptron weights are updated according to:
  6. 6. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 85 (12) Where, η is the learning rate parameter less than 1 d(n) is the desired output or target. IV. SIMULATION RESULT To test the performance of the proposed single layer perceptron based scheme simulation was carried out on MATLAB 2007b platform. 100 isolated handwritten odia digits were collected for the purpose of training and testing. First, the gradient feature was calculated then the curvature feature was calculated. Using the procedure in Section 2. For each data set a feature vector consisting of 2592 features was generated. Then using PCA the gradient and cuvature features were reduced to 95, 97 numbers. The feature vector comprising of 95, 97 features was used to train both the ANN based structures. Out of 100 isholated hand written odia digits 50 numbers of 0 and 50 numbers of 2 were collected. For the single layer perceptron, each training set comprises of 95 reduced gradient features as input and appropriate target. The 95 number of weights are initialized with small random values. Sigmoidal activation function is used as the non-linearity. Training was carried out for 500 iterations and learning curve for the same was plotted in the Fig. 9.And the same procedure is carried out for 97 reduced curvature features was plotted in Fig.10. After training, the error reduces to marginal value and convergence is achieved. 20 data sets were chosen for the purpose of testing out of which 10 data was from odia digit 0 and 10 data was from odia digit 2. It was observed that out of 20 isolated digits only digits have been recognized correctly showing an accuracy of 80% only. The testing results for the perceptron based scheme are shown in Table I. Fig.4: Data Set after Normalization Fig.5: Gray scale image Fig.6: Direction Of Gradient Fig.7: Strength Of Gradient
  7. 7. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 86 Fig.8: Curvature Fig.4: Training curve of Single layer perceptron for Curvature feature Fig.5: Training curve of Single layer perceptron for Gradient feature Table I: Testing Results Collected odia digits Pattern used in testing Results for gradient feature Results for curvature feature classified Not classified classified Not classified 0 10 9 1 8 2 2 10 8 2 6 4 0 50 100 150 200 250 300 350 400 450 500 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 iteration-> mse-> 0 50 100 150 200 250 300 350 400 450 500 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 iteration-> mse->
  8. 8. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 87 Confusion matrix (CT) of gradient in testing 9 2 CT = 1 8 Confusion matrix (CT) of curvature in testing 8 4 CT = 2 6 V. CONCLUSION In this paper we have considered the gradient features and curvature feature are used these for the recognition of isolated handwritten odia digits. These features are then reduced using the Principal Component Analysis. The reduced features are then separately used for training the single layer perceptron network adopted for the classification task. The network is trained as in the single layer perceptron network and the weights are updated using lms algorithm. Different sets of data were prepared for training and testing. The gradient feature are tested for 10 numbers of odia digit 0 and 10 numbers of odia digit 2.Out of 10 number of odia 0 digits only 9 number of digits are classified and out of 10 number of odia 2 digits only 8 number of digits are classified properly. So the Acurracy of classification using gradient feature is 85%, where as in curvature feature out of 10 number of odia 0 digits only 8 number of digits are classified and out of 10 number of odia 2 digits only 6 number of digits are classified properly. So the classification accuracy for the curvature feature feature is 70% only. REFERENCES [1] A. Amin, Off-Line Arabic Character Recognition System, State of the Art, Pattern Recognition, 31(5), 1998, 517-530. [2] Subhadip Basu, Nibaran Das, Ram Sarkar, Mahantapas Kundu, Mita Nasipuri and Dipak Kumar Basu, A hierarchical approach to recognition of handwritten Bangla characters, Pattern Recognition, Elsevier, 42, 2009,1467-1484. [3] J. Park, V. Govindaraju, S.N. Shrihari, OCR in hierarchical feature space, IEEE Transactions on Pattern Analysis and Machine Intelligence,22(24), 2000, 400–408. [4] U. Pal, Chaudhuri, B.B, Automatic Recognition of Unconstrained Off-line Bangla Hand- written Numerals, Proc. Advances in Multimodal Interfaces, Springer Verlag Lecture Notes on Computer Science, LNCS-1948, 2000,371-378. [5] N.Tripathy, M.Panda, U.Pal, A system for Oriya Handwritten Numeral Recognition, SPIE Proceeding, 5296, Eds. Barney Smith,E.,H., Hu, J., Allan, J., 174-181. [6] Y. Wen, Y.Lu, P.Shi, Handwritten Bangla Numeral Recognition System and its Application to Postal Automation, Pattern recognition, 40(1), 2007, 99-107. [7] Rajput,G.G., Hangarge, Mallikarjun, Recognition of Isolated Handwritten Kannada Numerals Based on Image Fusion Method, in Pattern Recognition and Machine Intelligence, LNCS.4815, 2007,153-160. [8] Bajaj, Reena, Dey, Leepika, Choudhuri, Santanu, Marathi Numeral Recognition by Combining Decision of Multiple Connectionist Classifiers,27(1), 2002,59-72.
  9. 9. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online), Volume 5, Issue 4, April (2014), pp. 80-88 © IAEME 88 [9] Roberts, L.G, Machine Perception of Three-Dimensional Solids, in, Tippett (Ed.), J.T, Optical Electro-Optical Processing of Information, MIT Press, Cambridge, MA,pp. 1965, 159–197. [10] Russ,J.C, The Image Processing Handbook, 2nd Edition, CRC Press, Boca Raton, FL.1995. [11] Wakabayashi,T., Tsuruoka,S., Kimura,F., Miyake,Y, Increasing the Feature Size in Handwritten Numeral Recognition to Improve Accuracy, Systems and Computers in Japan (English Edition), Scripta Technical, 26.(8),1995,35–44. [12] Fukunaga,K., in: Rheinboldt,W., Siewiorek (Eds.),D, Introduction to Statistical Pattern Recognition, 2nd Edition, Academic Press, New York,1990. [13] Kimura,F., Wakabayashi,T., Miyake,Y, On Feature Extraction for Limited Class Problem, Proceedings of the 13th ICPR, II, 1996,191–194. [14] Shi,M., Fujisawa,Y., Wakabayashi,T., Kimura,F, Handwritten Numeral Recognition using Gradient and Curvature of Gray Scale Image, Pattern Recognition 35, 2002, 2051 – 2059. [15] Primekumar K.P and Sumam Mary Idicula, “Performance of On-Line Malayalam Handwritten character Recognition using Hmm and Sfam”, International Journal of Computer Engineering & Technology (IJCET), Volume 3, Issue 1, 2012, pp. 115 - 125, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375. [16] Dr. Vangala Padmaja, “A Brief Review on Hand Written Character Recognition”, International Journal of Advanced Research in Engineering & Technology (IJARET), Volume 5, Issue 2, 2014, pp. 70 - 78, ISSN Print: 0976-6480, ISSN Online: 0976-6499. [17] Gunjan Singh, Avinash Pokhriyal and Sushma Lehri, “Fuzzy Rule Based Classification And Recognition Of Handwritten Hindi Curve Script”, International Journal of Computer Engineering & Technology (IJCET), Volume 4, Issue 1, 2013, pp. 337 - 357, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375. [18] Chaitali K.Dhande and P.M.Mahajan, “Evaluating Neural Network and Hidden Markov Model Classifiers for Handwritten Word Recognition”, International Journal of Electronics and Communication Engineering & Technology (IJECET), Volume 4, Issue 6, 2013, pp. 117 - 123, ISSN Print: 0976- 6464, ISSN Online: 0976 –6472.

×