Please use this identifier to cite or link to this item:
|Title:||Offline Handwritten Gurmukhi Script Recognition|
|Supervisor:||Sharma, R. K.|
|Abstract:||Over the last few years, a good number of laboratories all over the world have been involved in research on handwriting recognition. Handwriting recognition is a complex problem owing to the issues of variations in writing styles and size of the characters etc. The main objective of this work is to develop an offline handwritten Gurmukhi script recognition system. Gurmukhi is the script used for writing Punjabi language which is widely spoken in certain regions of north India. This thesis is divided into eight chapters. A brief outline of each chapter is given in the following paragraphs. The first chapter introduces the process of OCR and various phases of OCR like digitization, pre-processing, segmentation, feature extraction, classification and post-processing. Applications of offline handwritten character recognition system are also discussed in this chapter. In an overview of Gurmukhi script, the nature of handwriting in Gurmukhi script and character set of Gurmukhi script has also been presented. Major contributions and assumptions in this research work have also been discussed in this chapter. Chapter 2 contains a review of literature on various methods used for non-Indian and Indian scripts recognition. In this chapter, a detailed literature survey on established procedures for numeral and character recognition techniques has been presented. We have reviewed literature for different scripts, namely, Arabic, Bangla, Devanagari, French, Gujarati, Gurmukhi, Kannada, Japanese, Malayalam, Oriya, Roman, Tamil, Telugu and Thai in this thesis. Chapter 3 describes essential phases of an offline handwritten Gurmukhi script recognition system. These have been discussed in four sections entitled data collection phase, digitization phase, pre-processing phase and segmentation phase. In data collection phase, we have collected 300 samples of offline handwritten Gurmukhi script documents. These documents have been divided into three categories. Category 1 consists of one hundred samples of offline handwritten Gurmukhi script documents where each Gurmukhi script document is written by a single writer. Category 2 contains one hundred samples where each Gurmukhi script document is written ten times by ten different writers. In category 3, one Gurmukhi script document is written by one hundred different writers. These samples of offline handwritten Gurmukhi script documents of different writers have been collected from schools, colleges, government offices and other public places. In digitization phase, the procedure to produce the digital image of a paper based handwritten document has been presented. In pre-processing phase, size normalization and thinning of text has been done. In segmentation phase, a new technique has been proposed for line segmentation of offline handwritten Gurmukhi script document. Line segmentation accuracy of about 98.4% has been achieved with the use of this technique. Water reservoir based method has also been implemented for touching character segmentation with an accuracy of 93.5%. Chapter 4 presents a framework for grading of writers based on offline Gurmukhi characters. Samples of offline handwritten Gurmukhi characters from one hundred writers have been taken in this work. In order to establish the correctness of our proposed approach, we have also considered Gurmukhi characters taken from five Gurmukhi fonts. These fonts are: amrit, GurmukhiLys, Granthi, LMP_TARAN and Maharaja (F1, F2, …, F5, respectively). For training data set of handwriting grading system, we have used printed Gurmukhi font Anandpur sahib. Some of statistical features, namely, zoning features, diagonal features, directional features, intersection and open end points features have been used to assign a unique classification score to a writer. The gradation results are based on the values obtained by two classifiers, namely, Hidden Markov Model (HMM) and Bayesian classifier. Chapter 5 presents curve fitting based novel feature extraction techniques, namely, parabola curve fitting based features and power curve fitting based features for offline handwritten Gurmukhi character recognition. In order to assess the quality of these features in offline handwritten Gurmukhi character recognition, the performance of the recently used feature extraction techniques, namely, zoning features, diagonal features, directional features, transition features and intersection and open end points features have been compared with these proposed feature extraction techniques. Each technique has been tested on 5600 samples of isolated offline handwritten Gurmukhi characters. The classifiers that have been employed in this work are k-Nearest Neighbours (k-NN) and Support Vector Machine (SVM) with three flavors, i.e., Linear-SVM, Polynomial-SVM and RBF-SVM. The proposed system achieves maximum recognition accuracy of 97.9%, 94.6%, 94.0% and 92.3% using k-NN, Linear-SVM, Polynomial-SVM and RBF-SVM classifier, respectively, when power curve fitting based features are used in classification process. As such, the results obtained using power curve fitting based features are promising. It has also been seen that the results achieved using parabola curve fitting based features are also better than the other recently used feature extraction techniques. A maximum recognition accuracy of 95.4% has been achieved when the parabola curve fitting based features were used with k-NN classifier. In Chapter 6, we have presented an offline handwritten Gurmukhi character recognition system using zoning based novel feature extraction methods and k-fold cross validation technique. In this work, we have used various feature extraction techniques, namely, zoning features, diagonal features, directional features, intersection and open end points features, transition features, shadow features, centroid features, peak extent based features and modified division point based features for offline handwritten Gurmukhi character recognition. For classification, we have considered k-NN, Linear-SVM, Polynomial-SVM and MLPs classifier. In this study, we have considered 5600 samples of isolated offline handwritten Gurmukhi characters. We have concluded that peak extent based features are preeminent features than other feature extraction techniques. Using 5-fold cross validation technique, we have achieved recognition accuracy, with peak extent based features, of 95.6%, 92.4%, 95.5% and 94.7% with Linear-SVM, Polynomial-SVM, k-NN and MLPs classifier, respectively. Chapter 7 presents a Principal Component Analysis (PCA) based offline handwritten Gurmukhi character recognition system. PCA is used for extracting more representative features for data analysis and to reduce the dimensions of data. In this work, we have collected 16,800 samples of isolated offline handwritten Gurmukhi characters. These samples are of three categories. In category 1, each Gurmukhi character has been written 100 times by a single writer (5600 Samples). In category 2, each Gurmukhi character has been written 10 times by 10 different writers (5600 Samples). For category 3, we have again collected each Gurmukhi character written by 100 writers (5600 Samples). Here, we have also used different combinations of classifiers as LPR (Linear-SVM + Polynomial-SVM + RBF kernel), LRK (Linear-SVM + Polynomial-SVM + k-NN), PRK (Polynomial-SVM + RBF-SVM + k-NN) and LRK (Linear-SVM + RBF-SVM + k-NN) for recognition purpose. We have used different combinations of output of each classifier in parallel and recognition is done on the basis of voting scheme. The partition strategy for selecting the training and testing patterns has also been experimented in this work. We have used all 16,800 images of offline handwritten Gurmukhi characters for the purpose of training and testing. The proposed system achieves a recognition accuracy of 99.9% for category 1 samples, of 99.7% for category 2 samples and of 92.3% for category 3 samples. In this chapter, we have also presented a hierarchical technique for offline handwritten Gurmukhi character recognition. In this technique, we have proposed a strong feature set of 105 feature elements using four types of topological features, namely, horizontally peak extent features, vertically peak extent features, diagonal features, and centroid features. We have also applied various feature set reduction techniques, namely, Principal Component Analysis (PCA), Correlation Feature Set (CFS) and Consistency Based Feature Set (CON). We have seen that PCA performs better than other feature selection techniques for character recognition. A maximum recognition accuracy of 91.8% has been achieved with hierarchical technique when we considered PCA based feature set and Linear-SVM classifier with 5-fold cross validation technique. Finally, Chapter 8 presents the conclusion drawn from the results of various experiments conducted in this thesis. Also, some pointers to the future research on the topics considered in this thesis are discussed briefly.|
|Appears in Collections:||Doctoral Theses@SOM|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.