Optimization has been widely used in many separate image processing tasks. However, there have been few works focused on general frameworks that can be applied in different image processing problems simultaneously. The difficulty lies in the essential difference in the intrinsic characteristics of different image processing problems. If we narrow down our focus in image enhancement only, however, it becomes possible to formulate a general optimization framework for multiple enhancement problems. In this thesis we study the optimization problems in image enhancement. A general framework is proposed, from which different reduced embodiments are derived and effectively solved.

We then apply the optimization framework to different practical applications, including downsampling, supe...[ Read more ]

Optimization has been widely used in many separate image processing tasks. However, there have been few works focused on general frameworks that can be applied in different image processing problems simultaneously. The difficulty lies in the essential difference in the intrinsic characteristics of different image processing problems. If we narrow down our focus in image enhancement only, however, it becomes possible to formulate a general optimization framework for multiple enhancement problems. In this thesis we study the optimization problems in image enhancement. A general framework is proposed, from which different reduced embodiments are derived and effectively solved.

We then apply the optimization framework to different practical applications, including downsampling, super resolution, denoising, bit depth increasing, and dehazing. The intrinsic characteristics of specific applications are carefully handled, either in the cost function or in the constraints of the optimization framework. Weighted optimization is also studied, where the importance of different pixels are modeled in an additional weighting function, with the standing point that the statistical characteristics are content adaptive. Experiment results show the effectiveness of the proposed methods for each enhancement problem.

We further study the cases where the general optimization framework is difficult to use, for example image dehazing. The difficulty lies in finding a universally valid image prior and designing a good cost function based on it. We have found that machine learning is a better choice, for it is designed to learn the implicit mapping between input and output signals. With machine learning techniques, the learning data can be synthesized where groundtruth-hazy image pairs are readily available, and the intrinsic relation between them can be automatically learned. This scheme sheds some light on solving a large category of image enhancement problems.