Handwritten Amharic Bank Check Recognition Using Hidden Markov Random Field

Amharic, a working language in Ethiopia, has its own writing system which is totally different from that of the Latin alphabet based languages. Amharic handwriting recognition is challenging due to the huge number of symbols, significant interclass similarity and also intra-class variability. In this paper the application of Hidden Markov Random Field (HMRF) for handwriting recognition of the legal amount field of Amharic bank check is presented. The three main contributions of this paper are the following. First, a new feature extraction technique is used which tries to extract natural features as perceived by human beings. The features extracted by this technique show a significant performance improvement. Second, a classification technique by estimating likelihood using a method known as pseudo-marginal probability is developed. The third contribution is the application of contextual information based on the syntactical structure of Amharic checks. Such context information is important in recognition process because even humans fail to recognize symbols correctly without any context. A noticeable difference is observed between results obtained with and without the application of contextual information. On the whole, despite the huge interclass similarity and also intra-class variability of handwritten Amharic characters, attractive results are found.