Handwritten Bangla digit recognition employing hybrid neural network approach

Handwritten Bangla digit recognition is one of the most attractive area for researchers who have interest in image processing and pattern recognition field. In our everyday activities like bank check identification, passport and document analysis, number plate identification and especially in our postal automation service, recognition of handwritten digits plays a significant role. That's why a rich body of literature already has been published in this area. But most traditional techniques are generally based on complex feature extraction approach that introduces a great overhead in recognition tasks. Recently a novel approach Back-propagation algorithm is used for recognition which simplifies the recognition process but the main drawback is, the network takes lots of iterations to converge. This paper addresses a faster and efficient Hybrid Neural Network Solution called (BAM+BPNN) which is a combination of Bidirectional Associative Memory and Back-Propagation Neural Network. BAM is used for dimensional reduction and BPNN is used to train the neural network with the set of input patterns for acquiring separate knowledge of each digit. This research can take a decision that Hybrid Neural Network algorithm (BPNN with BAM) takes less iteration to train and less time to recognize digits than Back-propagation algorithm (BPNN). Experimental study shows the effectiveness of our proposed technique.