A very high accuracy handwritten character recognition system for Farsi/Arabic digits using Convolutional Neural Networks

In this paper, a new method is presented for recognizing the handwritten Farsi/Arabic digits by fusing the recognition results of a number of Convolutional Neural Networks with gradient descent training algorithm. Convolutional Neural Networks are a type of neural networks that are biologically inspired from human visual system which combines feature extraction and classification stages. This paper is concentrated on two main contributions. The first one is automatic extraction of input pattern's features by using a CNN for Farsi digits and the second one is fusing the results of boosted classifiers to compensate the recognizers' errors. The difference between competing systems is in the training set, which the frequency of samples that are “hard to recognize” were become higher in boosted classifiers. In addition, two rejection strategies were proposed and evaluated to find out “hard to recognize” samples. The experiments were conducted on extended IFH-CDB test database. The results reveal a very high accuracy classifier outperforming most of the previous systems. The achieved result shows 99.17% in recognition rate. In addition, the result was grown up to 99.98% after rejection of ten percents of “hard to recognize” samples.