Text CAPTCHA Traversal with ConvNets: Impact of Color Channels
Denis O. Ishkov, Valery I. Terekhov · 2022 4th International Youth Conference on Radio Electronics, Electrical and Power Engineering (REEPE) · 2022
Most of the existing studies have investigated fixed-length CAPTCHA recognition. In this paper the authors propose applying convolutional neural networks trained with a special loss function to recognize dynamic character sequences. The paper studies the influence of individual color channels and their linear combination on the models' final quality. Each color channel importance was estimated using trained weighting coefficients in linear combination of color pixel values. The results obtained allowed us to reduce the resource requirements for model training without loss of recognition quality, and at the same time to accelerate training. Model error analysis allowed us to make proposals for improving CAPTCHA designs and ways to counteract automatic recognition are outlined.