CTC Loss-Based OCR Model for Efficient Text CAPTCHA Recognition

CAPTCHAs are widely employed to safeguard systems against automated bots by differentiating human interactions from machine activities. They exist in various formats, including text, audio, video, and image-based challenges, with text-based tests being the most prevalent. These rely heavily on Optical Character Recognition (OCR) technology but often pose significant challenges due to the distorted or complex characters designed to mislead bots. Recent advancements have leveraged machine learning and neural networks, particularly the BiConvLSTM algorithm, which effectively combines convolutional layers for feature extraction and bidirectional Long Short-Term Memory (LSTM) layers for sequential data processing. This study introduces an OCR-based approach for text CAPTCHA recognition, employing Connectionist Temporal Classification (CTC) loss to improve model performance. Trained and tested on a publicly available dataset containing 1070 CAPTCHA images, the BiConvLSTM model achieves a character-level accuracy of 99.85% and a word-level accuracy of 97%. The model demonstrates enhanced accuracy over existing solutions, providing a robust framework for reinforcing security systems against automated threats and improving CAPTCHA recognition capabilities.