基于深度学习的验证码图像识别  

Captcha Recognition Based on Deep Learning

作　　者：徐星 宋小鹏[1] 杜春晖 XU Xing;SONG Xiaopeng;DU Chunhui(School of Information and Communication Engineering, North University of China, Taiyuan 030051, China;Taiyuan Institute of China Coal Technology and Engineering Group, Taiyuan 030006, China)

机构地区：[1]中北大学信息与通信工程学院,山西太原030051 [2]中国煤炭科工集团太原研究院有限公司,山西太原030006

出　　处：《测试技术学报》2019年第2期138-142,共5页Journal of Test and Measurement Technology

摘　　要：验证码是一种区分用户是计算机还是人的公共全自动程序.为了尽可能大批量地获取某网站的信息,就需要让机器可以全自动地识别该网站的验证码.为了破解验证码,对深度学习的验证码图像识别方法进行了研究.提出使用图像标注的方法来生成验证码图像中的字母序列.实验采用深度学习框架Caffe,将卷积神经网络与循环神经网络相结合进行训练.将卷积神经网络的输出用于训练循环神经网络,来不断地预测出序列中下一个最有可能出现的字母.训练的目标是将输出的词尽量和预期的词一致.测试结果表明,该模型能够对该网站的验证码图像做到97%的识别准确率.该方法比只采用卷积神经网络进行识别效果好.Captcha is a public fully automated program that distinguishes whether a user is a computer or a person. In order to get the information of other people's websites as much as possible, it is necessary to let the machine full-automated recognize the verification code of the website. In order to crack the captcha, the method of image recognition for deep learning verification code is studied. The feature extracted from the convolutional neural network is proposed for the generation of recurrent neural network statements, which is used to verify the method of captcha image annotation. Using Caffe, a deep learning framework to train. The output of the convolutional neural network before the last fully connected layer is used for feature extraction, and the recurrent neural network is used to continuously predict the next most likely word. The goal of training is to match the output words to the expected as far as possible. The test results show that the model can identify 97% recognition accuracy of the captcha image of the website. This method is better than only using the convolutional neural network.

关 键 词：验证码 深度学习 卷积神经网络 循环神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]