Research On Recognition Algorithms Of Text-based Captcha Based On Convolutional Neural Network

As one of the important means to ensure users’ safe use of the Internet,the captcha technology is widely used in various portals that need to verify identity,account registration,system login,operation confirmation and other important links,avoiding many network security risks.Research on the identification method of captcha can discover the current vulnerabilities of this technology from another perspective,promote researchers to design a more secure captcha system,provide a more secure network environment for people,and promote the development of technology.This thesis studies the recognition algorithm of character verification code.The verification code recognition process based on convolutional neural network generally includes two parts: character detection and character recognition.The overall method recognition only includes the process of character recognition.Captcha character detection algorithms are roughly divided into regression-based detection algorithms and segmentation-based detection algorithms.Therefore,by improving the algorithm based on convolutional neural network,this thesis proposes three kinds of verification code recognition algorithms: image detection + classification recognition algorithm,image segmentation + classification recognition algorithm and overall method recognition algorithm,and analyzes the three algorithms in verification code recognition.application features.The main work and innovations are as follows:(1)For most regression-based verification Captcha detection algorithms,the image features in different layers of the convolutional network cannot be effectively used,the local information is lost in the layer-by-layer extraction process,and the gradient caused by the deep network is unstable,etc.Problem,propose the RDB-Yolov5 algorithm.Drawing on the method of extracting multi-layer features in Residual Dense Net(RDN),Residual Dense Block(RDB)is used to improve the original feature extraction network.Make full use of the layered features of each convolutional layer,improve the feature fusion method,and reduce the information loss and the missed detection of small-scale verification code characters.(2)Aiming at the problems of character sticking,distortion,deformation and complex background in CAPTCHA,a segmentation-based CAPTCHA character detection algorithm is proposed: ASPP-DBNet algorithm.The algorithm can better distinguish characters and background interference information,and improve the segmentation effect while ensuring the inference speed.In the segmentation module,the encoder-decoder structure combining the feature pyramid and the improved version of ASPP(Atrous Spatial Pyramid Pooling)is used to obtain a larger receptive field and improve the ability of feature extraction.The simulation results show that the two improved CAPTCHA characters detection algorithms based on regression and segmentation can effectively detect CAPTCHA characters in simple data sets and CAPTCHA characters in difficult data sets with complex background,adhesion and deformation,and achieve higher accuracy.Rate.(3)Aiming at the situation that the characters in the verification Captcha are different in size and the image background is complex,a verification Captcha character recognition algorithm is proposed: SPP-Efficient Net algorithm.In order to expand the receptive field,integrate local features and global features,and learn from the SPPNet network architecture,the SPP(Spatial Pyramid Pooling)module is added to Efficient Net,which not only increases the depth of the network,but also improves the expressiveness of the input feature map,enhance the acceptance range of backbone features,and improve the accuracy of verification code character recognition.The experimental results show that the algorithm in this thesis performs well on simple data sets and difficult data sets,and has a certain improvement in the recognition rate of captcha images with complex backgrounds that are difficult to identify.(4)In order to study the application of the ensemble method in verification code recognition,this thesis improves the traditional end-to-end recognition algorithm,and proposes a segmentation-free ensemble method based on the CRNN(Convolutional Recurrent Neural Network)model,CAPTCHA recognition algorithm.When the traditional CRNN algorithm extracts character information,complex text information is easily ignored as noise information.The CBAM(Convolutional Block Attention Module)attention mechanism is introduced into the CNNs framework to improve the feature expression of the feature map.Ability to highlight key features and suppress other features.The experimental results show that the method proposed in this thesis performs well,the recognition rate is improved to a certain extent,and it has certain advantages over the other two methods in terms of speed.