FEGAN: A High-Performance Font Enhancement Network for Text CAPTCHA Preprocessing

This study aims to address performance deficiencies in CAPTCHA preprocessing methods that impede the accurate recognition of text CAPTCHAs, which are crucial for identifying security vulnerabilities. To improve CAPTCHA preprocessing methods, a similar font is initially searched and acquired by manually removing obstructing pixels from a target CAPTCHA and retaining the font part. Using the found font, a pseudo-dataset is generated containing a large number of clean and dirty pairs to train to the proposed supervised Font Enhancement Generative Adversarial Network (FEGAN), which is designed to effectively eliminate non-font-related interferences and preserve the font outlines. Test results show that FEGAN can improve the recognizer’s accuracy by approximately 16% to 50% on the M-CAPTCHA dataset (a publicly available dataset on Kaggle) and 5% to 35% on the P-CAPTCHA dataset (generated using the Python ImageCaptcha package), substantially outperforming the Multiview-filtering-based preprocessing approach.