188宝金博页面版

Computer Engineering and Applications计算机工程与应用 2018，54（3）1 引言随着互联网技术的快速发展，网络安全成了人们关注的一个重点内容。验证码主要是用于区分机器自动程序与人类用户的差异性，抵御恶意程序，防止滥用网络资源。验证码自动识别技术可以提高现有验证码安全性，并帮助设计出更安全的验证码，进而有效地确保网络安全。国内外学者在验证码的识别领域已有了较多的研卷积神经网络识别汉字验证码范 望，韩俊刚，苟 凡，李 帅FAN Wang, HAN Jungang, GOU Fan, LI Shuai西安邮电大学 研究生学院，西安 710121School of Postgraduate, Xi’an University of Posts and Telecommunications, Xi’an 710121, ChinaFAN Wang, HAN Jungang, GOU Fan, et al. Chinese character CAPTCHA recognition based on convolution neuralnetwork. Computer Engineering and Applications, 2018, 54（3）：160-165.Abstract：CAPTCHAs（Completed Automated Public Turing test to tell Computers and Humans Apart）have alreadybeen widely applied in various fields of social life. Automatic recognition of CAPTCHAs consisting of English letters andArabic numerals has already reached an advanced level. While with general methods identifing the CAPTCHAs consistingof Chinese characters seems too difficult and the accuracy needs to be promoted. This paper mainly proposes a method ofautomatic identification CAPTCHAs which is based on convolutional neural network to improve the accuracy of charactersrecognition. In order to improve the generalization performance of the model by which adopting the framework of Kerasconvolution neural network and designing of multilayer convolution to extract deep-layer image information of whichidentifing Chinese characters CAPTCHAs and alphanumeric CAPTCHAs respectively. The experimental results indicatethat the accuracy of identification has been promoted remarkably. The identification rate of Chinese characters is up to99.4%. Meanwhile, the maximum of the identification rate of alphanumeric four-character CAPTCHAs is as high as99.3%. These findings show that the Deep Neural Network possesses an excellent perceptivity against complex structures.It can be seen from the comparative experiments that the framework of Keras convolution neural network has better per-formance than other frameworks in CAPTCHAs recognition.Key words：CAPTCHAs（Completed Automated Public Turing test to tell Computers and Humans Apart）; Chinese characterCAPTCHAs; CNN; Keras framework摘 要：验证码今已广泛应用在各个领域，常见的英文字母与数字组合的验证码自动识别准确率已达到较高的水准，而汉字因其字符复杂，用传统方法进行自动识别难度很大。提出一种基于卷积神经网络的验证码自动识别方法来提高字符的识别准确率。采用Keras卷积神经网络框架，设计多层卷积来提取深层次图像信息，分别对汉字验证码和字母数字验证码进行识别，以提高模型的泛化性。实验结果表明用该方法汉字验证码的单字识别率已达到99.4%；传统四字符字母数字验证码的识别率最高达到99.3%。这一结果表明深度神经网络对验证码复杂结构的感知能力很强大，通过对比实验发现Keras框架在验证码识别领域有较好效果。关键词：验证码；汉字验证码；CNN；Keras框架文献标志码：A 中图分类号：TP391 doi：10.3778/j.issn.1002-8331.1706-0304作者简介：范望（1994—），男，硕士研究生，研究领域为基于深度学习的图像识别，医学图像识别，E-mail：978683267@qq.com；韩俊刚（1943—），男，二级教授，研究领域为软件和硬件的形式化验证，图形处理器和新型计算机体系结构；苟凡（1993—），男，硕士研究生，研究领域为基于深度学习的图像识别；李帅（1992—），男，硕士研究生，研究领域为计算机图像识别。收稿日期：2017-06-23 修回日期：2017-08-08 文章编号：1002-8331（2018）03-0160-06160