Feature Refine Network for Text-Based CAPTCHA Recognition

Text-based CAPTCHA is a widely used security mechanism. Text-based CAPTCHA recognition aims to automatically detect characters in a text-based CAPTCHA. It reveals the weakness of current CAPTCHA and improves the security ability. In this paper, we propose a novel Feature Refine network (FRN) for text-based CAPTCHA with small-size characters. FRN consists of convolutional layers and deconvolution layers. The convolutional layers enhance the feature extraction capabilities of the network and expand the receptive field. The deconvolution layers increase the resolution of the feature map and restore the details of texts. In addition, our model uses skip ROI pooling to extract multi-scale features with multi levels of abstraction. We test our model on five popular text-based CAPTCHAs, namely eBay, Baidu, Hotmail, Sina and NetEase. The experimental compared with the state-of-the-art methods demonstrate the ability of FRN. The recognition rates are improved above 90%, and these results achieve the new state-of-the-art for real website CAPTCHAs.