The Application of Convolutional Neural Network in Security Code Recognition

Abstract

Security Code or CAPTCHA, which represents Completely Automated Public Turing test to tell Computers and Humans Apart, is used to determine whether the user is human or not. It is widely applied in website management to prevent people from maliciously registering and spreading spam. With the development of machine learning algorithms and artificial intelligence technologies, it is possible to recognize the security codes with machine so that people can have access to the website without registration limits. There are many traditional machine learning algorithms such as Support Vector Machine (SVM) and Random forest are used to recognize security codes, but they have various disadvantages like low efficiency and low learning ability, inability to extract features automatically and difficulty in handling 2-dimension pictures directly. In machine learning area, the convolution neural network (CNN) is famous for its strong learning ability and automatic feature extraction, as well as high learning ability and efficiency, which makes it suitable for 2-dimension image data. Thus, we construct a convolutional neural network for security code recognition. The proposed CNN model is made up with 3 convolutional layers, a flatten layer and a full-connected layer. With the proposed model, we achieve an accuracy of 80.5% in the validation set.