An Efficient ConvNet for Text-based CAPTCHA Recognition

Text-based CAPTCHA is a widely used security mechanism to protect websites from malicious operations. The CAPTCHA recognition based on deep learning is a representative method to verify the security of CAPTCHAs deployed on the website. The ConvNet is a typical model for a wide variety of vision tasks and proves to be effective when it recognizes characters in different scenarios. However, the ConvNet applied to solve CAPTCHAs is still limited to high computation and complicated processing. In this work, we propose an efficient end-to-end network entirely consisting of standard ConvNet modules. According to the property of CAPTCHA, we reduce the redundant convolution calculation in the previous ConvNet and introduce a novel group convolution operation along the width of the image with improved performance and efficiency. The experiment shows our ConvNet successfully solves the CAPTCHAs from the highly-visited website, Sina.com, with a high accuracy above 90%, and the number of trainable parameters of it is less than 1/5 of the model in prior work based on prominent ConvNets such as ResNet and Inception network.