Optical Character Recognition of Heavily Distorted Text Segments

The fact that artificial text segments can be generated, which are recognizable by humans but not by artificial intelligence, shows that research in the field of artificial intelligence in the range of optical character recognition is not as advanced as it could be. Such artificial text segments serve as human interaction proofs in the field of computer security mechanisms, also known as CAPTCHAs. Taking one of the most used CAPTCHAs—the reCaptcha system—as a guideline for heavily distorted text segments, this thesis pursues the goal of developing a robust recognition algorithm, which is easily adaptable to other systems and use cases. In the scope of this thesis a strong character classifier based on Fourier descriptors is developed. The test results indicate that the proposed algorithm outperforms current state-of-the-art character recognition systems in single character recognition with a mean recognition rate of % vs. %. Furthermore, with a recognition rate of % it is able to break the reCaptcha system and beat human recognition performance by nearly %. On top of that, it is shown to be adaptable to other CAPTCHA systems and, despite being developed for character recognition, also to numbers. The complete algorithm solely depends on pre-rendered instances of characters without any knowledge or instances of applied distortions and is able to recognize and separate merged characters.