An Approach for Chinese Character Captcha Recognition Using CNN

Abstract

An approach using convolution neural network (CNN) to recognize Chinese Character Captcha captured from Tencent Security Center is proposed. 3500 commonly used Chinese characters are applied in the verification code system, so the problem is to detect, recognize and match the characters occurred in the two Captcha images. Firstly the characters in a Captcha image are analyzed and segmented, and each single character is a minimum unit in this paper. Secondly, a recognizing CNN is built and trained with a character library with 620,000 different character modes made by ourselves. Thirdly, the segmented characters from the Captcha image are input the network to finish recognition. The experiment result shows that the identification rate of the test data is as high as 99.4 percent after 25 iterations, and it is 1.75 percent higher when compared with another CNN model. It shows that the approach using CNN to do Chinese character Captcha recognition is feasible and it can provide a high level of accuracy.