Deep Learning Based Research On The Universality Of Text-based CAPTCHA Recognition Algorithm

As a kind of Turing test,text-based CAPTCHA plays an important role in network security,such as preventing malicious script programs from illegal attacks.As a means of network security protection,text-based CAPTCHA can be used to distinguish the operator from the human or the machine simply and effectively.Therefore,text-based CAPTCHA is widely used in various websites to prevent malicious batch operations,prevent dictionary attacks and ensure the authenticity of online voting.By studying the text-based CAPTCHA recognition,we can promote the development of text-based CAPTCHA recognition algorithm,and find out whether the type of text-based CAPTCHA design is safe,so as to enhance the network security.In order to recognize text-based CAPTCHA,various recognition methods have been proposed.According to whether segmentation involved in the recognition algorithm or not.These methods can be divided into segmentation based algorithm and non-segmentation based algorithm.This paper mainly studies the universality of text-based CAPTCHA recognition in the framework of deep learning.The main research contents are as follows:1.We propose a fully convolutional neural network framework based on sliding window to recognize text-based CAPTCHA.The traditional method of text-based CAPTCHA recognition based on segmentation has high requirements for segmentation algorithm,because segmentation is easy to cause the lack of character structure,which affects the recognition result.Based on the pixel level segmentation results of fully convolution neural network,this paper proposes to use sliding window to get the recognition results,which effectively solves the disadvantages of traditional segmentation algorithm.2.We propose a framework of corrective recognition network.The text-based CAPTCHA recognition algorithm of direct recognition method is greatly influenced by the type of character style.Character conglutination and distortion can be reduced by more datasets and stronger network structure,but character skew is difficult to be solved by these two solutions.In this paper,the framework of affine transformation is proposed to correct the character,and then recognition,which can effectively improve the recognition rate of direct recognition algorithm.3.We propose a text-based CAPTCHA recognition framework based on object detection.Although the type of text-based CAPTCHA is changeable,the transformation of single character is limited.Based on the limitation of single character transformation,this paper proposes a method of target detection to crack the verification code,and uses the target detection network identification verification code to achieve the universality of the verification code recognition algorithm.Based on the limitation of single character transformation,this paper proposes a method of object detection to recognize text-based CAPTCHA,and achieve the universality of the text-based CAPTCHA recognition algorithm.