Research On Captcha Recognition Based On Capsule Network

The network has been widely used in real life,in recent years,social news,such as harassing phone calls,telecom fraud,collation of other people's information for loans has gradually made people realize that their personal information has been exposed unreservedly,the security of personal information used in the website has increasingly aroused people's concern.In view of its easy-to-generate and difficult-to-identify(for computer)features,captcha recognition technical become an important technical means to protect the website information security,however,the pace of selling personal information,access to the designated content of the site to make a large profit will not stop,the machine identification of captcha has become an important technical means to break through the website to obtain information.In order to understand the security factor of the captcha used in the current website under the current technical means,measure the captcha used in the current website in the face of machine attacks in the ability,to better protect the network information security,captcha identification has also become an important technology of the site self-inspection,so as to further enhance the complexity of the current captcha,improve the captcha anti-attack ability lay the foundation.At present,there are many algorithms for captcha recognition task,and the endto-end learning mode of neural network has been widely concerned.In the traditional neural network processing captcha identification task,because of a picture contains more tags,and less training data for the task,and the traditional neural network,on the one hand,requires a large number of training data to learn the expression of characteristics,on the other hand,the neuron node contains too little information stored in scalar data form,pooling layer also makes a large amount of information loss between layers,resulting in the network in the recognition task more difficult.This article use the method for identification is the recently proposed neural network: capsule network(also konw as Capsnet).Capsnet has the capsule neurons,capsule in the special vector form of capsule network enhance the information expression ability of neurons,and implemented in classification layer between class and class discrimination probability values are independent of each other,can better apply to Multi-Label Classification task,dynamic routing algorithm between layers,replace the pooling layer,avoid the information loss,and enhance the interpretability of the flow of information,use less training data,at the same time,the precision of recognition is higher,Therefore,combined with the practical application requirements of captcha identification and the characteristics of Capsnet,carried out the experiment.We selected two multi-label recognition task processing methods: multi-label learning and multi-task learning,and data experiments are carried out.The experimental results of the three methods are compared to evaluate the capability of the Capsnet in the capctha recognition task.The first part of this paper introduces the theory background of captcha and the main recognition methods at home and abroad.The second part gives the development history,propagation algorithm and research status of neural network.In the third part,the structure of the capsule network is analyzed in detail,and the difference between the capsule network and the general neural network.The fourth part introduces the current multi-label classification task theory;In the fifth part,the experimental parameters of the network are listed,and the experimental details are presented.