Design And Implementation On Verification Code Identifying

Nowadays,our life has become more and more comfortable and convenient,because of the development of the Internet.However,the existence of security risks behaves negatively on the experience of users as a dark shadow.CAPTCHA is short for completely automated public turning test to tell computers and humans apart,also known as human interactive proof which is commonly referred to as “verification code”.It is a kind of network security technology and is used to automatically generate and evaluate the end-user telling human and computer program.Currently,this technology has been applied by many domestic and foreign websites,by forcing the human-computer interaction against programming attack.Somehow,it helps to protect our personal information.However,the imperfections of the design framework do really need to be improved.Currently,most of the verification code are composed of numbers and letters blocked or deformed.In this paper,we mainly take three steps to identify those verification codes:pre-processing,segmentation,recognition.We using Otsu threshold algorithm to separate foreground and background in a traditional way.Then according to the actual characteristics of the code,we use appropriate algorithm to reduce noise.Segmentation is the most important part of the whole identifies process,using vertical projection division method and segmentation method based on connected domain.We also present a stroke-based segmentation algorithm,which involves KNN classification algorithm and K-means clustering algorithm;connected-domain is also used as well.This paper also presents a character-based stroke segmentation algorithm,which involves KNN classification algorithm,K-means clustering algorithm,at the same time supporting the use of connected domain method.Convolution neural network has a specially designed structure that is sharing weights,is similar to biological neural networks.It has good robustness on image translation,scaling and rotation.In the recognition stage,we set up a simple seven convolution neural network as a core component.Training and testing data sets are made of standardized character pictures split from actual verification codes.In this paper,verification codes from huawei mall and 12306 are used as testing subjects of segmentation algorithms and neural networks.The accuracy rate of segmentation algorithm to the verification codes from railway website is 72.2%.And the recognition rate ofconvolution neural network for two websites is 78% and 70% respectively,which costs 15 minutes training the network at average.At the same time,the analysis concludes that the main reason for the lower recognition rate is the integrity of characters.