Character Recognition In CAPTCHA Image Based On Deep Neural Network

CAPTCHA is a public automatic program to distinguish human and computer script,widely applied for internet security.Deploying a CAPTCHA prevents the website from malicious attack such as automatic registration,comment,and voting.Text Captcha,which provides users with an image consisting of characters to be recognized,is the most widely employed for its flexibility and ease of use.Different from sequences in natural scenes,text-based CAPTCHAs lack semantic information that can be used to infer context and add various security features such as deformation,distortion,and overlap to resist segmentation and recognition.The recognition algorithms suitable for natural scene images often achieve limited performance when applied to the CAPTCHA directly.At the same time,due to the flexibility and frequent updating of the CAPTCHA,the method of training network models based on a large amount of labeled data is prohibitive.Based on the existing research,this dissertation explores the application of deep learning in CAPTCHA recognition task,which mainly includes three aspects:improving recognition performance,lightweight model size and enhancing generalization performance on small samples.Based on the properties of characters in CAPTCHA images,this dissertation proposes a variety of more accurate,efficient and robust deep neural networks.The research and corresponding contribution in this dissertation are concluded as follows:1.This dissertation proposes a model based on character feature sharing and task correlation,which significantly improves the recognition accuracy of various character types.Prior neural networks often use fully shared feed-forward network to extract features from multiple characters.Due to the differences in font,texture and ranking position among characters,such network structure has significant gaps in the recognition accuracy of characters at different positions,which restricts the final overall recognition rate.Inspired by multi-task learning network model,this dissertation puts forward a network structure based on the property of sharing and task correlation of the characters,the network uses a reasonable design of hierarchical structure of sharing and task specific layer division to make the network sharing features and independent of the characters of different position to learn at the same time,significantly improving the recognition performance of characters in the central characters of the sequence and the overall CAPTCHA recognition performance.2.This dissertation proposes a convolutional network based on explicit position information encoding.Compared with the prior convolutional neural networks-recurrent neural network hierarchical model,the network significantly reduces the computational cost and the number of parameters of the model under the premise of the same or better recognition performance.The character structure in the CAPTCHA is highly correlated with the spatial location.Although the existing convolutional neural network(convolutional neural network)-recurrent neural network(RNN)hierarchical model achieves the optimal performance in a variety of CAPTCHA recognition tasks,it is difficult to efficiently utilize the spatial location information in the CAPTCHA image in the process of feature extraction.Convolution structure design in this dissertation implements the explicit of feature location information encoding.Compared with the implicit position encoding in traditional convolution,explicit position encoding make the features extracted by convolutional layers more representative in CAPTCHA recognition task.This structure realizes the function of recurrent neural network modeling character sequence correlation in hierarchical model.On this basis,the simple structure and lightweight model size make the network fast,efficient and easy to deploy.3.This dissertation proposes an effective character-level self-attention mechanism for learning aliasing structure properties.The hierarchical network composed of Transformer designed based on this mechanism has excellent generalization performance on small training sets,and only 600-1000 annotated images are needed to effectively crack the CAPTCHA of the target website.Both convolutional neural network and Transformer network have certain limitations in the way of sensing regions in the task of CAPTCHA recognition.For CAPTCHA sequences of characters in the characteristics of the pixel level,this dissertation analyzes the local and global awareness of the convolution neural network the attention of the advantages and disadvantages of the Transformer,this dissertation proposes a character level since the attention mechanism make the network focused on the relationship between the connected characters,can effective learning CAPTCHA characters of structural characteristics of aliasing,The model can successfully crack the character CAPTCHA under the training of a small number of samples.This low-cost,high-precision identification means that websites need to re-evaluate their current security.