Research On Character Verification Code Recognition Based On Deep Learning

CAPTCHA is a turing test method designed to identify which operations are performed by humans and which are performed by computer programs.This test is usually used to maintain the security of human-computer interaction in the Internet.It is commonly used in stages such as user registration,login,and information verification.Because the production of character CAPTCHA is simple and easy to maintain,it has become the most commonly used CAPTCHA among many types of CAPTCHA.Research on character CAPTCHA recognition can discover the defects in character CAPTCHA design,thereby improving its design mechanism,enhancing the website’s ability to resist illegal intrusion,protecting users’ online privacy,and providing a safe online environment for users。Currently,researchers have proposed many methods for character captcha recognition,including those based on deep learning and those based on segmentation algorithms.However,the neural network models used in deep learning-based methods are relatively complex,while traditional segmentation-based methods suffer from low segmentation accuracy and unsatisfactory recognition results.Therefore,this paper proposes two character captcha recognition algorithms to address the above issues,and the specific work is as follows:1.The character CAPTCHA recognition algorithm based on character segmentation is proposed to address the low segmentation accuracy for sticky and distorted characters in traditional segmentation algorithms and the interference of the CAPTCHA image background.In the preprocessing stage,this algorithm uses knowledge of image processing to preprocess the CAPTCHA image.In the character segmentation stage,the algorithm combines the characteristics of the vertical projection segmentation algorithm and the color filling segmentation algorithm,and improves them by adding a step for secondary segmentation of characters.Finally,this article introduces global average pooling to replace the fully connected layer and constructs a 9-layer convolutional neural network to recognize characters.The experimental results show that our method has improved the segmentation efficiency and character recognition accuracy for both non-sticky and sticky CAPTCHAs.2.A lightweight network-based character captcha recognition algorithm is proposed to address the drawbacks of complex network models and high computational costs associated with deep learning methods in the field of character captcha recognition.This algorithm improves the Le Net-5network by first setting the input layer to a unified size of 160*60 pixels to adapt to the size of the input character CAPTCHA image.Secondly,we use dilated convolution instead of standard convolution in the convolutional layer to obtain a larger receptive field.We then introduce the BN layer to receive the input of each convolutional layer and normalize the received parameters using BN to improve the network performance.In addition,we improve and optimize the activation and loss functions by introducing the Leaky Re LU activation function and the Softmax cross-entropy loss function.Experimental results show that the proposed algorithm performs well and achieves a certain improvement in accuracy while maintaining a fast calculation speed.