Convolutional Neural Network Approach for Extraction and Recognition of Digits from Bank Cheque Images

Abstract

Recognition of handwritten character and digits is one of the major challenges in the computer vision system. One of the major application areas of character and number recognition is the financial sector, which deals with enormous amount of document data. In this paper, we have proposed to provide an automation system for bank cheque processing. The proposed system takes the cheque images as input and extracts the digits from account number, date and amount field, respectively. Initially, cheque images are preprocessed, and then digits are segmented using simple connected component analysis. Extracted digit images are normalized and given as input to convolutional neural network (CNN) classifier. Classifier is trained using large data set: MNIST and built-in MATLAB digit data set along with our data set. Total samples of 91,000 images are used for training and testing. Post-processing is done to construct the whole numbers for account number, date and amount. Recognition rate achieved for 50 cheque images is 95.59%.