Segmentation free Bangla OCR using HMM: Training and recognition

Abstract

The wide area of the application of HMM is in Speech Recognition where each spoken word is considered as a single unit to be recognized from the trained word network. Using this concept some research has been done for character recognition. In this paper, we present the training and recognition mechanism of a Hidden Markov Model (HMM) based multi font supported Optical Character Recognition (OCR) system for Bangla character. In our approach the central idea is separate HMM model for each segmented character or word. We emphasize on word level segmentation and like to consider the single character as a word when the character appears alone after segmentation process is done. The system uses HTK toolkit for data preparation, model training from multiple samples and recognition. Features of each trained character are calculated by applying Discrete Cosine Transform (DCT) to each pixel value of the character image where the image is divided into several frames according to its size. The extracted features of each frame are used as discrete probability distributions that will be given as input parameter to each HMM model. In case of recognition a model for each separated character or word is build up using the same approach. This model is given to the HTK toolkit to perform the recognition using Viterbi Decoding. The experimental result shows significant performance.