The Research On CAPTCHA Recognition Technology

With the rapid development of Internet , the network has brought great convenience to people's live. At the same time, network security issues are also prominently increasing. As a widely used means, Network CAPTCHA has played an important role in network security. The research of CAPTCHA recognition technology can help us discover and improve the loopholes of the verification code, so as to prevent websites from malicious attacks by machine program.This paper uses computer vision and pattern recognition to recognize different types of CAPTCHA. We select the representative Mop, Xicihutong and Tianya captcha as the research object. Aiming at the specific code, we proposed targeted breaking method to solve practical problems. The results reveal the possibility of their insecurity. Through comparison of different recognition algorithms, the study has theoretical and practical value. Following is the main work and achievements:1. Focus on the algorithms of BP neural network, convolutional neural network, and give their details of the derivation.2. For the existing divided Mop code, we propose a recognition scheme. It use piecewise linear transformation to remove the image blur. Local OSTU binary threshold got better than the global results. After improving the traditional projection method we propose the minimum segmentation algorithm and it effectively solve the merged characters in Mop code. Then simplified convolutional neural network is used to train and recognize single character. It achieves a high recognition rate at 94.1%.3. For existing Xicihutong verification code, the K-means clustering algorithm combined with vertical projection is proposed to complete division and it solve the problem of stacking and adhesion. Then single character is trained and recognized by simplified convolutional neural network and the recognition rate is 53%.4. For existing tianya captcha, we put forward the overall recognition algorithm based on shape context to verify code. The broken rate is 27.7%. This global recognition thought also gives a new idea to identify other code which is more difficult to split.