The Recognizition Of 12306 CAPTCHA Based On CNN

CAPTCHA is an automatic program that distinguishes between a user is a person or a machine.It is generated by machine according to the specific forms of human cognition,and then it can judge the user's answer to avoid brute force with a specific program batch operation.The Chinese part of 12306 website verification code is the non-standard font,will be randomly distorted,deformed and noise.The most of the image part are the very common objects in the majority of Chinese people's life.But there are a small number of objects due to people's age,education and other factors led to the identification of difficulties,coupled with the picture overall resolution is low,blurred,which give people eye recognition has brought a lot of unfavorable factors.In recent years,with the development of machine learning technology and computer hardware,the research of convolutional neural network(CNN)is becoming more and more in-depth,and the results obtained are more and more abundant.Nowadays,in many fields,convolutional neural networks have been widely used.Especially in the traditional image classification,convolutional neural network has made a breakthrough.From the early recognition of handwritten digits problem,and then to today's ImageNet competition,people witnessed the Lenet-5 to Vgg?ResNet and Inception network model.This is the result of a tireless effort by the researchers to try to adjust and improve the structure and parameters of the convolution neural network in a variety of ways to maximize performance.In this paper,based on the characteristics of the 12306 specific text and image verification code,on the basis of in-depth understanding and research of convolutional neural network theory and research results at home and abroad,the main work is as follows:(1)A migration learning method based on fine-tune small-scale data sets on pretrained network model is implemented.By using the comparison experiment of network model training and two kinds of migration learning methods,the training method suitable for this data set is selected.(2)Complete the data set.To crawl with a certain number of 12306 web site verification code to remove duplication,marking,classification and data augmentation.(3)Using vgg-16,vgg-19,inception_v3 and other classic network to compare and analyze the results of the data set after fine-tune,the network model suitable for this data set is selected.(4)This paper presents an integrated solution to identify 12306 text + image verification code.