Research On Image Captcha Recognition Algorithm Based On Deep Learning

At the beginning of design,captcha is just a Turing test to distinguish human from computer program.With the development of Internet,the current captcha is mainly used as a tool to maintain Internet security.In general,it appears in the user login website,the website registration,the information inquiry,the website post and so on.In these scenarios,we want the user to be a real person,not a computer program.Image character recognition is very simple for human beings,but it is difficult for computer programs.Therefore,the existence of the captcha can maintain a good network environment to a certain extent.There are some scenarios captcha may also affect the user experience,and in order to improve the security of the network,as the academic research to identify the captcha,the existing loopholes can be found.In this paper,deep learning framework Keras is used,choose theano as the back-end,for easy segmentation captcha,After de-noising and segmentation,it has achieved good results through the convolution neural network and the deep neural network.For the captcha which is difficult to segment,With the idea of end to end,the captcha is marked as a whole,and the model is trained by the convolution neural network and the recurrent neural network,and the effect is better than the traditional method.The specific work content and results are as follows:(1):Several common captcha are introduced,and the recognition of each verification code is analyzed.The commonly used methods and principles of verification code acquisition,grayscale,two valued,denoising and segmentation are introduced in detail.(2):This paper introduces the present situation of the development of deep learning,introduces the forward propagation and reverse propagation of the neural network,introduces the principles of convolution and pool operation of the convolution neural network.For the easy segmentation captcha,this paper takes the captcha of the Shandong mobile online business hall as an example,and uses de-noising and segmentation to complete the preprocessing.Convolution neural network modeling and recognition,recognition accuracy of about 90%,while using deep neural network modeling and recognition,the final accuracy rate of 87%.(3):The principle of recurrent neural network and long and short recurrent neural network is introduced.For the captcha which is not easy to be divided,this paper takes the captcha of the Shandong telecom network business hall as an example,and uses the end to end convolution neural network modeling and recognition to identify the correct rate of about 85%,At the same time,convolution neural network and long term recurrent neural network are combined to model and identify,and the final accuracy rate is about 70%.