Font Size: a A A

Research And Implementation Of Neural Network Verification System

Posted on:2016-06-26Degree:MasterType:Thesis
Country:ChinaCandidate:X LiFull Text:PDF
GTID:2208330473961440Subject:Computer technology
Abstract/Summary:PDF Full Text Request
With the development of the Internet, network security becomes more and more important, and all kinds of security technologies have been put forward. The CAPTCHA is an important way among them to protect the network security. The using of CAPTCHA prevents the automatic program from attacking illegally on the web services, as well as spam wantonly spreading, which can provide a clean and healthy network platform to the enterprises and the users. But more and more CAPTCHAs were maliciously cracked now because of some others purposes. Therefore, the research of CAPTCHA and recognition technology of CAPTCHA gets more and more important. By means of the research on the CAPTCHA recognition technology, one can design easier and safer CAPTCHA. Meanwhile, because many algorithms of machine learning have been used in the CAPTCHA recognition, the research on verification code identification is also a big step toward artificial intelligence.Recently, there are many kinds of verification codes, such as image verification code, text-based CAPTCHA. However, the text-based CAPTCHA is still frequently used in the network, which usually adopts noise interference or distorted adhesion to increase its identifying complexity. At present, the recognition methods of text-based CAPTCHA can be roughly divided into two kinds:one is to base on the segmentation recognition method; the other is to base on the overall recognition method, namely no division.In this paper, the method based on segmentation was adopted to research the recognition of text-based CAPTCHA, and the JINGDONG’s CAPTCHA and CSDN’s CAPTCHA as the identifying object are used to test because they have the features of distortion, partial adhesion and background interference. The method which combines the connected domain segmentation with the projection histogram segmentation for verification code segmentation is presented in this paper. Then, the convolutional neural network in machine learning as identifier for character recognition is used. The main work of this paper is as follows:(1) The research background, variety of CAPTCHA and its current status of research are introduced briefly.(2) The image pre-processing and image segmentation algorithms are briefly introduced. Then the pre-processing algorithms of CAPTCHA are chosen by the experiments, and the segmentation algorithms of two kinds of CAPTCHAs are designed respectively based on the connected domain segmentation and the projection histogram segmentation. Finally, the image size is normalized.(3) The knowledge about the neural network is introduced in detail, and the neural network used in this paper refers to the model of LeNet5 is designed. Finally, the training process of the neural network, optimization methods and training skills are introduced.(4) The requirements analysis is done for the software. The realization of the software is introduced in detail, including the method of capturing the CAPTCHA, making the datasets, the Torch3 library of machine learning and the designing of class diagram. Finally, the VC++/MFC programming language and VS2010 development environment are used to realize the software, and the corresponding analysis for the experiment results is presented.
Keywords/Search Tags:CAPTCHA, Connected Domain Segmentation, Projection Histogram Segmentation, Convolutional Neural Network
PDF Full Text Request
Related items