粘连字符的图片验证码识别.doc

?图形、图像、模式识别? 粘连字符的图片验证码识别 王 璐，张 荣，尹 东，詹金春，吴陈洋 WANG Lu，ZHANG Rong，YIN Dong，ZHAN Jinchun，WU Chenyang 中国科学技术大学 电子工程和信息科学系，合肥 230027 Department of Electronic Engineering Information Science，University of Science Technology of China，Hefei 230027，China WANG Lu，ZHANG Rong，YIN Dong，et al.Breaking visual CAPTCHA of merged characters.Computer Engineering and Applications，2011，47（28）：150-153. Abstract：CAPTCHA has made great contributions on maintaining the Internet security and preventing malicious machine at- tack.However，through the existing pattern recognition techniques，some CAPTCHAs can also be recognized.This paper focus- es on breaking mop and xicihutong web CAPTCHA with merged characters.The difficulty is the segmentation of merged characters in image.Aiming at mop CAPTCHA which has blur merged characters，the paper proposes a segmentation algo- rithm based on local teeny value and minimum projective value.As to xicihutong web CAPTCHA with interlaced merged characters，the color clustering is proposed by combining with the vertical projection to achieve the purpose of segmentation. It uses the convolution neural network to train and recognize them，reaching a high recognition rate. Key words：Completely Automated Public Turing test to tell Computer and Humans Apart（CAPTCHA）；merged character； segmentation；convolutional neural network 摘 要：验证码在维护互联网安全、防止机器恶意攻击做出了很大贡献。但通过现有的模式识别技术仍然可以破解部分验证 码。着重于有粘连字符的猫扑和西祠胡同网站验证码进行识别，难点在于分割图片中的粘连字符。对字符是模糊粘连的猫扑验 证码，提出了基于局部极小值和最小投影值的方法来分割；对有交错粘连的西祠胡同验证码，通过颜色聚类与竖直投影结合来达 到分割字符的目的。最终均采用卷积神经网络进行训练和识别，达到了较高的识别率。 关键词：验证码；粘连字符；分割；卷积神经网络 DOI：10.3778/j.issn.1002-8331.2011.28.041 文章编号：1002-8331（2011）28-0150-04 文献标识码：A 中图分类号：TP391.43 验证码（CAPTCHA，Completely Automated Public Tur- ing test to tell Computer and Humans Apart），即全自动区分 计算机和人类图灵测试的简称，是一种区分用户是计算机还 是人的公共全自动程序，这个程序必须生成并评价人类容易 通过但计算机却不能通过的测试。验证码形式一般为基于图 片和基于声音，本文识别的重点在图片验证码，下文提到的验 证码统指图片验证码。 研究验证