A Transformer Network for CAPTCHA Recognition

Abstract

Websites can improve their security and prevent harmful Internet attacks by incorporating CAPTCHA verification to dictate whether the end-user is a human being or a robot. It is critical to improve the CAPTCHA design method and promote the CAPTCHA design level that its recognition technology can drive. In this paper, the neural network algorithm is used to study CAPTCHA recognition. First, to address the issues in the traditional BP neural network, such as uncertain structural parameters, low convergence rate, and quickly accept a local minimum. This paper proposes to use a convolutional neural network (CNN) to learn the words’ feature in an image. Second, existing methods are inadequate for CAPTCHAs with colored impedimental lines, character adhesion, rotation, distortion, and scaling interference. This paper presents a new method-based transformer scheme for CAPTCHA identification. The image is divided into individual letters firstly with image pre-processing as an option to improve accuracy and then the detected letters are spelled into words. The proposed method is more efficient and verified by a collection of data.