Offline Handwritten Gurmukhi Character Recognition: A Review

Transcription

1 , pp Offline Handwritten Gurmukhi Character Recognition: A Review Neeraj Kumar 1 and Sheifali Gupta 2 1 Electronics & Communication Department Chitkara School of Engineering and Technology Chitkara University, Himachal Pradesh 2 Electronics & Communication Department Chitkara institute of Engineering and Technology Chitkara University, Punjab Neeraj.kumar@chitkarauniversity.edu.in, Sheifali.gupta@chitkara.edu.in Abstract All over India more than 12 crore people utilize Gurumukhi script for speaking, documenting & other purposes. A considerable advancement in the work associated with the recognition of handwritten and printed Gurmukhi text has been reported in last few years. From the last few decades offline handwritten character recognition has gained a lot of interest of researchers. It is well known that each individual has some different writing style, so it is very difficult to identify or recognize the handwritten characters. Researchers have worked in this field using various scripts like Hindi, English but a very little work has been done in Gurmukhi script point of view. Based on data acquirement process a concise classification of recognition system has been discussed in this article. Various feature mining techniques & classifiers like power arc fitting,parabola arc fitting,,diagonal feature extraction, transition feature extraction, K-NN classifier (Knearest neighbor) & SVM classifier (Support vector machine) are also illustrated in this paper. The methodology for word recognition has also been discussed in this paper. Keywords: Handwritten character recognition (HCR), K-NN classifier SVM classifier, feature extraction 1. Introduction Printed & HCR are the main categories of a character recognition system. In printed recognition system, a printed document is initially scanned and transformed into a machine processable form. The machine processable metaphors are pre-processed and segmented to character level to extract features from it. The term Handwritten Character Recognition is the process of conversion of text written in handwritten form into machine process able form.hcr is categorized into two parts i.e. offline and online. Online handwritten recognition of characters is very much dissimilar from offline recognition of handwritten characters as in offline handwriting recognition, the information related to strokes sis not present. Generally a character recognition system involves a number of actions such as Digitization, Pre-processing, Segmentation, Features extraction etc.reading cheques, postcode recognition and verification of signatures are the major applications of OHCR. Many researchers have worked on the problems of recognizing the offline characters.munish Kumar et al [1] showed that training set greatly affects the efficiency of the character recognition system.r.k Sharma et al [2] projected a scheme to recognize Gurmukhi characters which is based on transition & diagonal features with the help of K-NN classifier. To find the K-nearest neighbors, the authors calculated the Euclidian distance b/w test point and reference point. Rajiv Kumar et al [3] showed how segmentation can be done in character recognition of Gurmukhi script. In this work the ISSN: IJSEIA Copyright c 2016 SERSC

2 authors gave emphasis on the segmentation of words and lines in Handwritten Gurmukhi text.r.k Sharma et al [4] have reported that by using various features like diagonal, directional & zoning & K-NN, Bayesian classifiers we can compare the handwriting of writers.vipin Narang projected a approach based on parts or technique suitable for scene images. The author has taken the corner points which are served as parts and these parts are found to make a part based model [14].Adwait Dixit et al proposed a feature extraction & classification approach named as wavelet transform. The proposed technique achieve a accuracy of 70%.Wavelet transform is used in order to extract wavelet features. Generally two wavelet coefficients are determined named as approximation and directional coefficients. Now for each and every wavelet coefficient, various parameters like mean, deviation will be found. Author has used artificial neural networks (ANN) for the classification stage [15].Nitin Kaliraman [16] projected a recognition system for Devanagari character recognition.the author has used ANN and with the proposed system accuracy of 75.6% is achieved. 2. Gurmukhi Script Punjabi language is written using Gurmukhi script and it has been derived from the Punjabi word Gurmukhi. In this script there are a number of consonants, auxiliary signs, vowel bearers, vowel modifiers & half characters. The components used in Gurmukhi script are shown below in Figure 1. ਸ ਹ ਕ ਖ ਗ ਘ ਙ ਚ ਛ ਜ ਝ ਞ ਟ ਠ ਡ ਢ ਣ ਤ ਥ ਦ ਧ ਨ ਪ ਫ ਬ ਭ ਮ ਯ ਰ ਵ ੜ Consonants ੳ ਅ ੲ Vowel bearers ਸ਼ ਜ਼ ਖ਼ ਫ਼ ਗ਼ Additional consonants Vowel Modifiers Auxiliary signs ਹ ਰ ਵ Half Characters Figure 1. Components in Gurmukhi Script Due to the similarity of the various characters in this script, recognition becomes very difficult. 78 Copyright c 2016 SERSC

3 Top strip ਲ ਧ ਆਣ Header line Bottom strip Core strip Figure 2. Various Strips in Gurmukhi Character Generally each Indian script is having some particular set of set of laws for the amalgamation of consonants, vowels and modifiers. Basically modifiers may be attached to the consonants or can be attached to a vowel. The words in the Gurmukhi script can be broken down into three strips namely top strip, bottom strip and core strip. Top and core strips can be alienated by a header line and core and lower or bottom strip can be separated by a virtual base line. Feature extraction is one of the major fields in character recognition. Generally features are of two types :(1) Statistical features (2) Structural features. Structural features have been derived from various statistical distributions of points like moment, histogram etc.structural features includes loops, direction of strokes, end points, intersection of strokes. Character recognition is carried out using two approaches namely template and feature based approach. In the first approach the unidentified test pattern will be compared straightforwardly by the idyllic pattern. Conversely in feature based approach, we will extract features from test pattern and the extracted features will be used in classification. These features can be used in various classification models like ANN, HMM (Hidden Markov model), SVM etc. 3. How Recognition Takes Place For doing the recognition, there are some steps which must be followed and that steps are as following: Digitization Pre-processing Segmentation Feature Extraction Classification Copyright c 2016 SERSC 79

4 Figure 3. Steps for Character Recognition Image Acquisition: In this section the image digitization takes place. Initially some characters are written on the sheet of paper. Now the paper is scanned using a scanner. After scanning the image a bitmap image will be obtained. The process of converting the document typically in the handwritten form into an electronic format is termed as Digitization. After digitization an output image i.e bitmap image is given to pre-processing stage as a input. Pre-processing: It is basically the operations that are applied on digital image. A bitmap image is normalized to a window of 100x100 size. The output of pre-processing stage will be a normalized bitmap image. Preprocessing involves a number of steps like binarization, detection of edges, dilation of images, and filling of holes present in image. In binarization, with the help of thresholding the gray image is transformed into the binary image. Feature Extraction: It plays an important role in the recognition of characters. The efficiency of recognition relies a lot on the features extracted. There are various techniques through which features can be extracted such as, power & parabola arc fitting, diagonal feature extraction, transition feature extraction which can be used for the recognition of offline Gurmukhi character recognition. (a) Diagonal feature extraction: To attain high recognition efficiency the diagonal features are very important. By moving along the diagonals of each zone the diagonal features can be mined from the pixels of each zone. 80 Copyright c 2016 SERSC

5 Figure 4. Diagonal Feature Extraction Steps for extracting the diagonal features are as following: I: Firstly the image has to be converted into 100 zones & size of each zone should be 10*10 pixels. II: Now if one move beside the diagonals of each zone then one can extract the diagonals features. III: Every zone is having nineteen diagonals; forefront pixels which are available along the every diagonal are added to have a solitary sub feature. IV: Now the average of nineteen sub-features values is taken to make a particular value & is positioned into the respective zone as its feature. V: Now analogous to the zones whose diagonals don t have a foreground pixel, the feature value is considered as zero. By utilizing the above steps, one can have n features related to every zone. (b) Transition feature extraction To find out the transition information, the image is scanned from top to bottom & left to right. Steps for mining the transition features are as following: I: Firstly the image has to be converted into 100 zones & size each zone should be pixels. II: Find the no. of transitions for each zone. By utilizing the above steps, one will have n features corresponding to each zone. (c) Parabola curve fitting feature extraction: Firstly the image has to be converted into 100 zones as shown in Figure 4. A parabola is fixed to the series of ON pixels in each zone by applying least square method. A parabola y=a+bx+cx 2 is defined by three Copyright c 2016 SERSC 81

6 4. Literature Review Table 1. Literature Review in Tabular Form Paper ID Year Technique Used/Feature extracted Conclusion [1] 2000 Statistical features In this paper, the author has outlined a scheme to integrate different information sources in order to recognize the Devanagari script. These information sources are habitually statistical in nature for OCR [17]. [2] 2002 Fuzzy Multifactorial Analysis In this article the author has found that continuation of stirring characters is a foremost difficulty to plan the efficient character segmentation process. A novel method is offered for recognition & segmentation of touching characters. [3] 2008 Elastic matching In this the author has projected two stages for the recognition. In the first phase unidentified strokes are recognized and in the second stage the author has evaluated the characters with the help of strokes that are found in the first stage [5]. Using the elastic matching a maximum recognition rate of 90.08% is achieved. [4] 2009 Small line segment In this the author has proposed a method called small line segment for the recognition of online handwritten Gurmukhi characters[7] [5] 2010 Top-Down Technique & windowing technique [6] 2010 Moment based Feature extraction The authors concluded that after doing the windowing, character were interpreted correctly. In this article the author has discussed various moments like Zernike moments, Pseudo Zernike moments etc [6].Out of all these moments the pseudo Zernike moments give best results. [7] 2011 K-NN classifier Using K_NN classifier & diagonal features a accuracy of 94.12% is achieved [8] 2011 Zoning,Directional,diagonal feature extraction techniques-knn & HMM classifiers [9] 2011 SVM with RBF kernel, Zoning The authors have found that by using these techniques we can compare the handwriting of different authors. In this article the author has found that by using various techniques a accuracy of 94.53% for known writers and unknown writers can be achieved. [9]. [10] 2011 Nil The authors have presented various stages for hand written character recognition system [11]. [11] 2012 SVM,PNN,ANN,KNN & NNN [12] 2012 Binarization,morphological operations (erosion and dilation) The author has reviewed a number of techniques and classifiers for Gurmukhi character recognition like SVM, PNN, ANN, KNN & NNN.The author has emphasized on SVM for offline handwritten Gurmukhi character recognition [8]. According to author that by using various features likes zone density, projection histograms and geometric features such as area, perimeter a 82 Copyright c 2016 SERSC

7 maximum accuracy of 98% can be achieved [12] SVM & KNN classification In this article the author has found that a accuracy of 95.11% and % can be achieved using SVM & KNN classifiers respectively[13] [14] 2012 HMM In this article the author has proposed two recognition techniques based on HMM i.e. lexicon driven technique and lexicon free technique. Both these techniques are quite useful in recognition systems and gives good recognition accuracy [18]. [15] 2013 Support vector machine [16] 2014 K-NN classifier SVM classifier According to author that the training size has a great impact on the accuracy of offline Gurmukhi HCR In this article the author has found that by using classification techniques like k-nn and SVM a accuracy of and % can be achieved respectively[10] Table 2. Comparative Table for Features Extracted and Classifier Techniques Paper id Worked on Feature extraction Classifier Technique Efficiency [1] Recognizing online handwritten Gurmukhi Characters [2] Offline handwritten gurmukhi character and numeral recognition [3] HCR of Gurumukhi script [4] [5] Feature extraction techniques Offline HCR of Gurumukhi script [6] Recognition system for handwritten gurmukhi Position of stroke, area, length, curliness and slope Diagonal features, transition features, zoning density features, projection histogram features. Zoning density features, background directional distribution (BDD) features Zoning, diagonal, Directional, changeover, intersection and open end points, gradient and chain code features. Diagonal feature, crossroads and open end points feature extraction Area, perimeter, major and minor length axis, orientation, eccentricity, zone density, project Characters histogram. parameters: a, b and c. This will provide 3n features for a given character. Nil SVM ANN NNN K-NN SVM K-NN SVM SVM Feed Forward multilayer perceptron 94.59% 95.04% with SVM 94.12% with K-NN 92.78% with NNN with ANN 95.04% 98.10% with K-NN % with SVM 94.29% 98% Copyright c 2016 SERSC 83

8 Steps to Extract features: I: Firstly the image has to be converted into 100 zones of equal size. II: For each zone, a parabola is fitted by applying least square method & find the values of a, b and c. III: Now the zones which do not have a foreground pixel, we will take the values of a, b and c as zero. (d) Power curve fitting feature extraction: Firstly the image has to be converted into 100 zones of equal size. A power curve is fitted to the series of ON pixels in every zone by applying least square method. A power curve of the form y=ax b is exclusively defined by two parameters: a & b. This will thus provide 2n features for a specified character. Steps utilized to mine these features are as following: I: Firstly the image has to be converted into 100 zones of equal size. II: In each zone, fit a power curve using least square method & determine the values of a & b. III: Analogous to the zones that do not have a foreground pixel, take the value of a & b as zero. 4. Methodology Generally in character recognition we follow the following procedure: Firstly we take a image which may be a scanned image or camera captured consists of various words and characters. Secondly we have to apply some segmentation technique in order to segment the image i.e break the words into individual characters so as to recognize each word. In the third step we have to extract the features from the image and these features will be used in classification stage. We can use number of classifiers like K-NN, SVM, and ANN etc.various classification techniques are summarized in the next section. Segmentation & Classification: An image having series of characters is broken down into sub-images of individual character with the aid of segmentation. Classification is basically the phase where decision making is done. In this, we make use of the features that we extract in the feature extraction stage. Decision making is done in the classification using various classifiers. k-nn classifier has been used by a number of researchers for the purpose of recognition. The Euclidean distance in K-NN classifier is found by using the formula as following : d= Where, N Gross number of features, X k Value of library stored feature and Y k Value of the candidate feature. K-NN: K-Nearest neighbor technique is used for classifying the objects which are based on closest training examples present in feature space.k-nn technique is also known as lazy learning. In this algorithm an object can be classified by the majority votes of the K-Neighbors. If the value of K is equal to 1 then we can say that object is allocated to the class of nearest neighbors. The Euclidian distance is calculated to discover the K- neighbors. After finding the distance we will put the distances in ascending order for further classification. SVM:Generally SVM classifier accepts input data & classify the input data into two distinct classes.svm classifier is trained by with some training data & a model is made for the classification of test data which is based upon this model. Sometimes there is 84 Copyright c 2016 SERSC

9 a problem arises named as multiclass classification. For this there will be a need to design multiple binary classifiers. There are number of kernel functions in this algorithm: (1) Polynomial kernel (2) Lineal kernel (3) Gaussian radial basis function (4) Sigmoid or tangent kernel ANN: Artificial neural network (ANN) is basically a information processing model which works in the similar fashion as our brain works.ann is used in pattern recognition & data classification via some learning process. Information processing elements are neurons and ANN consists of neurons which help in solving specific problems. 6. Proposed Technique and Methodology Initially a image containing Gurmukhi text will be occupied and this image will be fed to the pre-processing stage to remove the noise. After removing the noise, filtered picture will be received. Now using some appropriate segmentation technique each line in the image will be segmented. After line segmentation, words in the corresponding lines will be segmented. After the segmentation of the words, we will have individual words having different characters. At last we will segment each character and finally we will recognise each character. 7. Conclusion and Future Scope Figure 5. Methodology for Word Recognition In this article an endeavor has been made to summarize the offline handwritten Gurmukhi character recognition system. Various features extraction techniques like diagonal features extraction, transition features extraction, parabola curve fitting, power curve fitting based extraction techniques have been discussed. Furthermore the Copyright c 2016 SERSC 85

10 classification techniques like K-nearest neighbor Support vector machine has been discussed in this article. The results using various classification techniques have been reported in the literature review. This article will provide an preliminary help for the researchers who wish to research in this particular area. Furthermore the current efforts by the various authors are limited to character recognition only. This can be extended to word recognition level. A brief methodology for doing the same has also been discussed in the previous stage. References [1] Munish Kumar,M. K. Jindal, Size of Training Set Recognition Accuracy of HCR System, JETWI, Vol. 5, No. 4, November [2] R.K. Sharma, Munish Kumar K-NN Based Offline Handwritten Gurmukhi Character Recognition ICIIP, [3] Rajiv Kumar, Amardeep singh, Detection and Segmentation of Lines and Words in Gurmukhi Handwritten Text 2nd International Advance Computing Conference, IEEE 2010 [4] M.K. Jindal, R.K. Sharma Classification of Characters and Grading Writers in Offline Handwritten Gurmukhi Script, ICIIP, [5] R.K Sharma, Anuj Sharma, Online Handwritten Gurmukhi Character Recognition Using Elastic Matching, Congress on Image and Signal Processing, 2008 [6] Renu dhir, Moment based Invariant Feature Extraction Techniques for Bilingual Character Recognition, ICETC,2010. [7] Anuj Sharma et.al Recognizing Online Handwritten Gurumukhi Characters using Comparison of Small line segments, IJCTE, Volume.1, Number.2, June2009, [8] Anoop Rekha, Offline Handwritten Gurmukhi Character and Numeral Recognition using Different Feature Sets and Classifiers - A Survey IJERA, Volume. 2, Issue , PP [9] Rajneesh Rani, Renu Dhir Handwritten Gurumukhi Character Recognition Using Zoning Density and BDD Features, IJCSIT, Volume. 2, 2011, PP [10] R. K. Sharma, M.K. Jindal, Efficient Feature Extraction Techniques for Offline Handwritten Gurmukhi Character Recognition, National. Academy. Sci. Letters, 2014) PP [11] R. K. Sharma, Munish Kumar, Review on OCR for Handwritten Indian Scripts Character Recognition, pp , Springer-Verlag Berlin Heidelberg 2011 [12] Mandeep Kaur, Sanjeev Kumar, A Recognition System For Handwritten Gurumukhi Characters, IJERT Volume. 1 Issue 6, [13] Gita Sinha, Rajneesh Rani, Handwritten Gurmukhi Character Recognition Using K-NN and SVM Classifier, IJARCSSE, Volume 2, Issue 6, June [14] Sujoy Roy, Vipin Narang, Devanagari Character Recognition in Scene Images, ICDAR, [15] Adwait Dixit, Yogesh Dandawate, Handwritten Devanagari Character Recognition using Wavelet Based Feature Extraction and Classification Scheme, INDICON, [16] Neha Sahu,Nitin kaliraman, An Efficient Handwritten Devnagari Character Recognition System Using Neural Network, Automation, Computing, Communication, Control & Compressed Sensing,2013. [17] Veena Bansal et.al (2000) Integrating Knowledge Sources in Devanagari Text Recognition System IEEE transactions on systems, man, and cybernetics,systems & Humans, Volume 30, number.4, PP [18] A.Bharath et.al (2012), HMM-Based Lexicon-Driven and Lexicon-Free Word Recognition for Online Handwritten Indic Scripts, IEEE transactions on pattern analysis and machine intelligence, Volume 34, number 4, Copyright c 2016 SERSC