Recognition of Multi-Fontstyle Characters Based on Convolutional Neural Network

Abstract

Convolutional Neural Networks are popularly used in OCR and document recognition. This paper applies stochastic diagonal Levenberg-Marquardt method into a convolutional network, which is presented by Simard. The relations between the sample class number, global learning rate and the network's convergence speed are discussed, Experiments on different train sets showed that class number is an essantial factor to the neural network's convergence. We have successfully expeanded Simard network into recognition of multi-font style little character set like Baidu CAPTCHA and got a recognition rate as 98.4% in single Baidu CAPTCHA character, and 93.5% as the overall rate. Experiments in this paper has confirmed that Convolutional Neural Network can be successfully used in recognition of multi-fontstyle little character set.