Ocr framework for offline tamil handwritten scripts

Abstract

In Pattern Recognition, Optical Character Recognition (OCR) is one of the popular and challenging research areas. OCR is a technology by which machine understandable code can be obtained from the character image. The applications of OCR include handwritten and printed ancient documents reading, such as handwritten characters in bank cheque, recognizing the characters from palm scripts, identifying the characters from inscriptions, and so on. This thesis focuses on recognizing the Offline handwritten characters of the cursive scripts collected from people written in the Tamil language. Even though Tamil is a South Indian language, it is also a highly speaking language in Canada, Singapore, Malaysia, SriLanka, and so on. The character sets of Tamil language consists of 12 vowels, 18 consonants, and 216 vowelized consonants and special characters. It has unique character sets to represent Tamil numbers, years, and months. The general structure of the Tamil language is versatile and cursive in nature. Numerous complexities are observed from the handwritten shapes of the Tamil Characterand#8223;s structure. They are, similar shapes introduced by writers, minor variations for the same character, characterand#8223;s structure connectivity, unnecessary curves, touching letters, shape variations, unwanted loops, angle variations leading to shape variations, unknown shaped characters similar in shape and character discontinuity etc. To provide the solution for these issues during the recognition of the versatile Tamil 247 character set, four interrelated component selection approaches are proposed in this thesis before feature extraction process. These component selections are facilitated to choose the proper characterand#8223;s structure before applying thefeature extraction algorithms. They are Zone and Junction Point-based Component Selection (ZJPCS), Junction Point-based Component Selection (JPCS), Zone-based Component Selection (ZCS), and a Junction Point Elimination based Component Selection (JPECS). To extract the required features from the character images, few test experiments are carried out with a proper analysis of various recognition works of Tamil and other language characters. The structures of the characters are connected with uneven components. By the way of human intervention, the characters are identified by the connectivity and location of the connected component. As per the feature selection in the Image Processing techniques, choosing the proper component is an essential and crucial way to recognize by a machine. Hence, it plays a major role in the proposed methodologies. For each component selection approaches, suitable feature extraction approaches are adapted to reveal the essential features from it. These features are extracted statistically (based on the pixel availability) and structurally (based on the shape properties). From these collections of feature extraction techniques attempted, four different frameworks are proposed and implemented in a hybrid way to implement a successful Tamil handwritten character recognition system. newline