SCK - Sparse Coding Based Key-point Detectors for Knowledge Discovery

点击翻译按钮获取中文摘要

Detection of key-points or distinctive points in images is a very important step in numerous vision-based applications i.e. image matching, image stitching and object detection/recognition. The research in the field has seen successful introduction of many key-point detectors such as Harris, SIFT, SURF, MSER, SFOP. However, the very nature of being developed based on a human pre-defined structure (a corner, blob or junction, etc.) makes these classical (hand-crafted) detectors inflexible when being used in unintended situations. In recent years, deep learning has been applied to key-point detection. Yet, the main stream of learning based methods remains using hand-crafted detectors to provide input to their training algorithms, essentially making them hand-crafted as well. These facts clearly indicate that the development of new classical detectors is still an indispensable task. 1.We will analytically verify the property of scale and rotational invariance possessed by SCK. It is expected that, by incorporating this property, the application range of SCK will be significantly broadened to handle images of very different scales and rotational angles. 2.We will investigate how SCK can be used to enhance the available detectors, classical or learning based alike. By fusing SCK and the hand-crafted detectors, it is expected that the robustness of SCK against non-uniform lighting, scale and rotational changes will improve the quality of classical detectors. A learning based detector with SCK providing reliable input will also be studied. 3.We will propose and develop a new feature descriptor, to replace the popular SIFT descriptor which possesses limited information discrimination power. The new descriptor is expected to work coherently with the SCK detector. 4.We will apply the developed algorithms in the previous three tasks to multiple practical vision tasks, especially focusing on two areas in which we have high level of expertise and broad experience: non-invasive biometrics and multimedia information processing. The expected outcome of the proposed research is the generation of a novel and complete framework with associated tools for key-point detection and feature descriptor extraction. The gained knowledge will generate significant societal benefits to Canada and provide enabling technology for applied research such as image matching, image stitching, human-computer interaction, multimedia information retrieval, and solution to a broad range of real world problems, including autonomous vehicles, high quality photography, new media production/delivery, preservation of cultural heritage, e-health, smart homes, gaming and e-entertainment, and creating life-like experience in immersive multimedia environment.