基于深度学习的物体识别验证码破解方法

摘要

为了提高验证码的安全性,图像验证码应运而生,然而类似于12306官网的图像验证码,它的安全性就一定很好吗？针对这一问题,提出了一种基于深度学习的图像识别破解12306验证码的方法.由于深度学习(Deep learning)在图像识别上具有明显的优势,因此利用CNN(Convolutional Neural Networks)在图像分类上的经典模型-Alexnet去训练12306图像数据集.实验结果表明,在像素值过低的情况下,图像与文字的识别准确率分别达到了93.2％和99.5％.最后再根据12306验证码的结构编写破解脚本并调用生成的两个模型文件自动识别验证码上的图像和文字,输出需要点击的图像或坐标.即做成图像识别的一个应用实例--12306验证码破解工具.%In order to improve the security of the CAPTCHA,the image CAPTCHA came into being.But similar to the 12306 official website of the image CAPTCHA,it's security must be good.To solve this problem,a method based on deep learning to break it was presented in this paper.Deep learning has obvious advantages in image recognition,so the classical model of image classification of Convolutional Neural Networks(CNN) called Alexnet was used to train the 12306 image dataset.Experimental results show that the recognition accuracy of image and text is 93.2％ and 99.5％ respectively when the pixel value is too low.Finally,according to the structure of the 12306 CAPTCHA,a crack script was writen and the generated two model files were called to identify the image and text automatically on the CAPTCHA,and then the images or coordinates to be click were output.We implemented all these steps into an image recognition application,the 12306 CAPTCHA breaking tool.