Conquering image imperfectness by priors with applications to image recovery

Techniques on improving the quality of digital visual data are getting more and more attention with the fast growing pervasive high quality display equipment. However, those image enhancement tasks are extremely difficult due to the under-determined nature, since only a limited portion of the ideal signal can be gathered by current digital equipments with limited observations. Powerful image prior is of critical importance to address this problem.In this dissertation, I investigate several image priors. I propose the edge smoothness prior to remove the jaggy edge artifacts. The SoftCuts metric is first derived in the literature as an analytical form of this prior with a specific geometric explanation, which enables a principled approach to obtain smooth image edges. In addition, I study the sharp edge prior, which captures an important property of clear well-focused images of still scenes. Moreover, these priors can be smoothly integrated with an α-channel image description scheme, which characterizes the image edge transition precisely, and in a concise way.I apply these priors successfully on several challenging image enhancement tasks. The general motion from blur problem with one single input image is addressed for the first time in the literature, including estimating both parametric and non-parametric forms of the motion field, as well as multiple motion estimation and segmentation. In addition, I propose a novel approach for super resolving one single input image. With the help of the SoftCuts metric, the jaggy edge artifacts are greatly reduced. I also investigate the partial blur removal problem for single input images, based on the analysis of the generation process of partial blur, and on the usage of powerful image priors and a user assistant method.Extensive experimental results demonstrate the effectiveness, efficiency, scalability and robustness of the proposed approaches.