A Research Of CAPTCHA Recognition Base On Shape Context

With the development of the Internet, the problem of proliferation of spam andnetwork security comes to appear. The emergence of CAPTCHAS to some extent reducesthat. Today, many CAPTCHA systems remain unsafe and many of them are easy to beskipped. The study of the recognition of CAPTCHA can boost not only the development ofCAPTCHS’s design, but also the development of related field, such as image processing,character recognition and so on. The detected curves which recognized by the existingCAPTCHA recognition algorithm in denoising stage are not smooth enough. The existingCAPTCHA recognition algorithms cannot detect intermittent curves, its accuracy is not highand its speed is slow.The dissertation aiming at the existing problems of line interference detectionalgorithm. First, using5x5neighborhoods to replace the original algorithm’s3x3neighborhood to solve big-tilt-angle-curve-detection problem. Secondly, selecting the pointswhich torsion angle less than90degrees from the neighborhood as the points which on theline of interference, in order to ensure the smoothness of the curves. In order to solve thedetection of the interrupted curves by setting the threshold value. Recognition algorithm forthe CAPTCHA in the recognition rate is not high, and the character changes in somevalidation code image (there may be a certain amount of rotation and scaling), in order toadapt to the requirements, this dissertation resolve these problems by the statistics of eachside of the tangent vectors to improve shape context. This dissertation paper calculates theshape context by selecting the representative points to match to improve the problem ofslow recognition speed.This paper use improved shape context recognition algorithm to identify the two sites’CAPTCHA(OSCHINA, CSDN). The recognition rate is98%and54%, respectively.Compared with original algorithm, the improved algorithm’s accuracy is higher, especially;the accuracy of recognition of the CAPTCHA which with the strong noise increases in largeextent the algorithm execution time is reduced to about1%of the original.