Recognition Of Character Verification Code Based On Convolution Neural Network

As an effective means of network security protection,identifying code has been widely used on the Internet.The research of identifying code recognition can not only design more secure and easy-to-use identifying code from the perspective of anti-identification of identifying code,but also discover the security vulnerabilities of identifying code in time,so as to improve the work efficiency of network servers and the security of users access to the Internet.In recent years,deep learning technology has been widely used in various research fields.However,the application and research of deep learning,especially convolutional neural network,in recognition of different identifying codes is a little scarce.Therefore,convolutional neural network is applied to recognition of identifying codes pictures consists of numerals,letters and Chinese characters commonly used in the Internet.For verification code pictures consisting of numbers and letters,the feasibility of target detection algorithm SSD based on convolutional neural network applied to the recognition of verification code pictures is analyzed.It is pointed out that SSD network will reduce the ability of extracting global information in the process of stacking convolutional layers continuously.Based on this,non-local design idea is introduced into SSD network.The NL_SSD network structure designed in this thesis is put forward by adding non-local module to the five continuous convolution modules.The influence of the enhancement of horizontal flip data and the setting of detection threshold in detection on the recognition result is analyzed and discussed.For the verification code picture consisting of common Chinese characters,the recognition process of Chinese character verification code is divided into two parts:character segmentation and character recognition.The reason why NL_SSD network in Chapter three can not be directly applied to Chinese character verification code recognition is analyzed.The application of NL_SSD algorithm in character segmentation stage of Chinese character verification code is described in detail.The problems of existing classification network and corresponding modules in Chinese character recognition stage are analyzed.The core module of Chinese character verification code recognition design is given,two parallel convolution branches merge feature map channels and make module residuals.The core module is stacked with convolution layer,pooling layer and full connection layer to form CHNet network.The experimental results show that the classified network composedof modules in this paper maintains a high recognition accuracy and greatly speeds up the convergence speed and reduces the network complexity.