Abstract:

CAPTCHA (Completely Automated Public Turing Test to Tell Computers and Humans Apart) is a crucial human-machine distinction tool that websites employ to thwart automated malicious program attacks. Investigating CAPTCHA recognition can reveal weaknesses in CAPTCHA systems. By leveraging deep learning and computer vision techniques, the very purpose of CAPTCHAs can be circumvented. A Deep Convolutional Neural Network model is employed to identify CAPTCHAs, eliminating the need for traditional image processing techniques such as location and segmentation. Our research proposes a CAPTCHA recognition system focusing on the central area of feature maps using the DCNN model, which we customized and combined with the attention mechanism. This approach helps distill the character information that needs to be learned during training in the complex context of CAPTCHAs with lots of noise. The experimental findings illustrate that our model has exceptional identification capabilities on CAPTCHAs that contain background noise and character adhesion distortion. It achieves excellent accuracy and a low character mistake rate across several datasets.