Neural-Based Hit-Count Feature Extraction Method for Telugu Script Optical Character Recognition

Abstract

The recognition accuracy and efficiency of any OCR system greatly depend on the feature extraction methods. There are several feature extraction methods each has its own characteristics. These methods differ in terms of the number features that they extract and the complexity. With less number of features, the recognition accuracy may be low, and with more number of features, the recognize time may be more. The features are to be selected in such a way that they could distinguish one character from other with minimum comparisons and gives less false positives and false negatives. The accuracy of an OCR can be improved by changing the feature extraction methods. Telugu is called Italian of the east. But it is surprising that there are not many OCRs that could detect Telugu characters with fairly good accuracy. The accuracy of OCRs available in the market are either highly objectionable or the price is very high. To address this issue, we took up this project. Other problems include the segmentation of overlapped characters and right feature extraction. We tried to solve these issues, by taking a segmented character from a word and check to find a correct match for it or tell that the character does not exist so that the particular character can be re segmented. In this work, a hit-count-based feature extraction method with neural networks is used for the fast recognition even though the training time is more. The experimental results show that the proposed hit-count-based feature method greatly reduces the time by maintaining the recognition accuracy.