Recognizing Character-Matching CAPTCHA Using Convolutional Neural Networks with Triple Loss.

Completely Automated Public Turing test to tell Computers and Humans Apart (CAPTCHA) is a widely used type of challenge-response test to determine whether or not the user is human in many web applications. The traditional CAPTCHAs with English and Chinese characters can be automatically recognized with high accuracy. Yet current methods are limited in recognizing new CAPTCHAs such as character-matching CAPTCHA. We present an approach that combines convolution neural network with triple loss to solve character-matching CAPTCHA. We evaluate our approach on five types of CAPTCHAs including character-matching CAPTCHA. The experimental results show that our approach outperforms other four common recognition methods in the aspects of both accuracy and convergence speed.