Transformer Encoder for Efficient CAPTCHA Recognize

Abstract

CAPTCHAs have become a key factor affecting the efficiency of Robotic Process Automation. In this study, we constructed a large-scale CAPTCHA dataset encompassing up to 60 diverse stylistic variations, offering robust data support for model resilience and adaptability. We designed a CAPTCHA image prediction reconstruction task, allowing for unsupervised feature encoding that implicitly models the image spatial domain features. This enables the acquisition of feature representations for various CAPTCHA styles in a common subspace. Additionally, character length prediction and character set prediction tasks were established to explicitly model text label features, fine-tuning network model parameters in alignment with recognition tasks. Through lightweight sequence modeling and text prediction modules, our design achieves high predictive accuracy while maintaining a controlled model size. Our proposed model attained a full-word accuracy rate of 92.47% on the multi-style CAPTCHA dataset, surpassing existing mainstream approaches.