Research On Captchas Recognition Based On Convolutional Neural Network

Text-based Captchas have been widely deployed across the Internet to defend against undesirable or malicious bot programs,which is an important part of the Internet security.Analyzing the security and usability of the text verification code helps to improve the ability of website to resist malicious attacks,ensure user information security,and maintain a secure network environment.In this paper,we focus on identifying Captchas with various Convolutional Neural Networks,improving the accuracy of recognition and reducing the complexity of the implementation structure.The insights provided in this work can help security experts to revisit the design and usability of text captchas.On the other hand,this work explore a new research field for completing challenging scene text recognition using Convolutional Neural Network.The paper carries out the following work:Firstly,the paper studies the basic theory and implementation model of Convolutional Neural Networks in computer vision problem.The fundamental module of CNN in this paper were studied,including convolution operation and maxpooling,normalization method and fully connected layer,espacailly the applicateable conditions and scpor for activation function used in CNN.The typical Convolutional Neural Network such as CNN,ResNet and Xception were simulated to demonstrate the performance on Captchas recongnition.Secondly,we introduce scaled exponential linear units,which induce self-normalizing properties,as the activations of Convolutional Neural Networks dubbed Self-normalizing Convolutional Neural Networks.Scaled exponential linear units employ strong regularization schemes to make learning highly robust.Furthermore,for activations not close to unit variance,we prove an upper and lower bound on the variance,thus,gradient descent and disappearance are impossible.The overfitting problem in CNN is alleviated and the generalization performance of CNN is improved.Thirdly,a novel structure named Variational Auto-Encoding Convolutional Neural Network is proposed.The Variational Auto-Encoding model is studied.As a combination of a depthlessness model and a probabilistic model,it can make a thorough inquiry for input features and ensure that the compression representation of characterizatio are of probabilistic significance.The combination of Variational Auto-Encoding and Convolutional Neural Networks can learn a meaningful and generalizable representation of the potential space,greatly reducing the time required model training for Captchas character recognition.It provides the interpretability in probability theory for the realizing processing of Convolutional Neural Networks,which abstracts features layer by layer and outputs the final target.Finally,exponential squash function driven Capsule Networks dubbed LECapsNet is proposed to further improve the recognition accuracy.The fact that the output of a capsule is a vector makes it possible to use a powerful dynamic routing mechanism to ensure that the output of the capsule gets sent to an appropriate parent in the layer above.This type of routing-by-agreement should be far more effective than the very primitive form of routing implemented by max-pooling which allows neurons in one layer to ignore all but the most active feature detector in a local pool in the layer below,which is more important for dealing with ambiguity and stacking characters for Captchas recognition.In this paper,by changing the clustering method of feature vector,we construct the LECapsNet,which enlarges the local characteristics of the input data.The recognition accuracy of Tencent verification code is 95% and that of the CCT mechanism verification code is more than 90%.