Visualizing Residuals for Optimized Text CAPTCHA Recognition

Text CAPTCHAs are critical components of network security and are widely deployed in various information systems. Analyzing text CAPTCHA models is crucial for identifying system vulnerabilities. We used a single residual connection as a baseline for visual performance comparison against multiple residual connections (non-centroids), utilizing a clustering graph based on training and validation accuracy. A Gradient-weighted Class Activation Mapping (Grad-CAM) plot then identified key residual connections, guiding further optimization. Finally, a per-character accuracy change curve refined the analysis of different residual performance on individual characters. These visualization methods optimize residual connections without increasing model parameters, resulting in a more robust CAPTCHA recognizer.