A Texture Based Approach to Word-Level Script Identification from Multi-script Handwritten Documents

Script identification from handwritten document images is an open document analysis problem especially for multilingual environment like India. To design the Optical Character Recognition (OCR) system for multi-script document pages, it is essential to recognize different scripts prior to employing an OCR engine of a particular script. The present work describes a texture based approach to word-level script identification from five handwritten scripts namely, Malayalam, Oriya, Tamil, Telugu and Roman. A 92-element feature vector has been designed in which 80 features consists of selected coefficients of Discrete Cosine Transform (DCT) and the remaining 12 features have been taken from the Moment invariants. Experimentations are conducted on a database consisting of 1000 word images of each script which are evaluated using multiple classifiers. The Multi Layer Perceptron (MLP) classifier is found to be a best choice for the said purpose which is then applied comprehensively using different cross validation folds and different epoch sizes. The average success rate for the present technique of word-level handwritten script identification is found to be 93.56% for 5-fold cross validation with epoch size 1000, which is quite encouraging.