Abstract:

In human-computer intelligent interaction, it is a basic technology for machine to recognize captcha automatically. Text-based captcha recognition usually preprocesses the captcha images, then cuts them, and finally classifies and recognizes the characters. The accuracy of character cutting directly affects the final recognition result. This paper proposes a method based on adversarial network to recognize text captchas. First, a Pix2pix network is trained to preprocess the captcha image, and then a pair of segmentation and recognition networks are obtained by using the method of confrontation training. The segmentation network can not only separate the pasting characters, but also filter out the hard-to-divide verification codes. The identification network adopts a context-dependent multi-channel convolution network, which can effectively solve the problem that the information cannot be identified due to information loss during the segmentation process. Experimental results show that this method can improve the accuracy of text-based captchas recognition.

Key words: adversarial learning, captcha cracking, segmentation, neural network

摘要：

在人机智能交互中，让机器自动识别验证码是机器模拟人的一项基础技术。基于文本的验证码识别一般先对验证码图片进行预处理，然后切割，最后对字符分类识别。字符切割的准确程度直接影响最终识别结果。提出一种对抗学习方法识别文本型验证码。先训练一个Pix2pix网络对验证码图片进行预处理，然后对抗训练出一对分割和识别网络。分割网络不仅能分割粘贴字符，而且可以筛选出难以分割的验证码结果。识别网络采用上下文相关的多通道卷积网络，能有效解决分割过程中因信息丢失而无法识别的问题。实验结果表明，该方法能提高文本验证码识别的准确率。

关键词: 对抗学习, 验证码识别, 分割, 神经网络