engleski

PATTERN RECOGNITION WITH PIXEL INTENSITY COMPARISONS ORGANIZED IN DECISION TREES

This thesis investigates computer-vision algorithms that are based on pixel-intensity comparisons organized in decision trees. The motivation for this topic were previously published papers that report competitive results at a very high processing speed. High processing speed is particularly relevant for methods running on small devices with limited resources, such as mobile phones and embedded hardware. The thesis starts by providing a discussion of the origins of the method. Next, a computational framework for reconstructing a small image patch from pixel-intensity comparisons is presented and experimentally verified. The result indicates that pixel-intensity comparisons are capable of encoding the appearance of the patch to a sufficient degree. This serves as the justification for pursuing further research in this area. The following chapters of the thesis introduce novel approaches for solving problems of broad practical interest: face detection, eye-pupil localization and face alignment. All methods are thoroughly experimentally tested on several publicly available datasets and compared to the state-of-the-art approaches. The results show that the proposed methods achieve sufficient accuracy for a wide range of applications. Furthermore, their very high processing speed makes them especially suitable for real-time applications running on small devices with limited processing power and low battery life. This indicates their superiority over state-of-the-art approaches based on convolutional neural networks for such applications. In the end, the thesis gives a summary of the most relevant conclusions derived from previous chapters.

decision trees ; pixel-intensity comparisons ; face alignment ; face detection ; eye- pupil localization

nije evidentirano