Machine Learning Approaches to CAPTCHA Recognition Requiring Minimal Image Processing

This study focuses on a machine learning approach to the CAPTCHA recognition problem that requires minimal preprocessing of inputs. Both the feed-forward neural nets and the self-organizing maps used in this study took the raw image pixels as input with only simple segmentation and translation performed in advance. The models were trained and tested on four and five letter CAPTCHAs using either letters A-G or AZ, with both models achieving greater than 90% accuracy on four letter CAPTCHAs of letters A-G. Accuracy decreased with the increase in CAPTCHA length, a result of increased crowding of the letters that reduced the performance of the segmentation algorithms. These results show that the difficulty in CAPTCHA recognition is not in the OCR, but rather in the image segmentation. Future work will focus on the image segmentation problem, in addition to ensemble techniques to increase the accuracy of the current models.