Image Enhancement by Super-resolution, Focus Editing and Exposure Composition.

摘要

Although significant progress has been made in imaging devices during the past few decades, the photographs acquired by digital cameras are still far from perfection due to the physical limitations of hardware such as aperture, lens and sensor. This fact brings out the demand for study on image enhancement: a computational technique that aims to improve the interpretability or perception of information in photographs for human viewers. The work in this thesis mainly focuses on three tasks in image enhancement.;Firstly, since the camera sensor has limited resolution, the acquired images cannot capture the scene very detailedly. Hence, people often resort to a postprocessing technique called super-resolution (SR) to enhance the resolution of the captured images. In the first part of this thesis, two approaches are presented to address the challenging single image SR problem, which is to recover a high-resolution (HR) image from one low-resolution (LR) input. Specifically, a novel learning-based framework is designed specifically for face image SR task from the perspective of DCT domain. In addition, an efficient two-step scheme is developed to super-resolve generic image by exploiting the salient edges of the input LR image.;Secondly, due to the limitation of lens and aperture, some cameras cannot produce pleasant photographs with desired focus setting. For example, portrait photography that requires shallow depth of field (DOF) is not allowed when using the compact point-and-shoot cameras. In the second part of this thesis, a new and complete postprocessing-based focus editing system that is able to handle the tasks of focus map estimation, image refocusing and defocusing, is developed to overcome the optical limitations and create different kinds of novel photos with desired focus setting from an imperfect photo.;Finally, since the radiance of the real world spans several orders of magnitude and its dynamic range dramatically exceeds the capability of the current digital cameras, there often exist some undesirable over- or under-exposed regions in a photograph. The third part of this thesis aims at producing one great looking well-exposed image that is virtually impossible with a single exposure by compositing a stack of photos at different exposures taken with a conventional camera. Particularly, a simple but effective method is presented to describe how to take advantage of the gradient information to accomplish exposure composition in both static and dynamic scenes. Compared to conventional high dynamic range (HDR) imaging work, the proposed approach is quite appealing in practice since it is computationally efficient and easy to use, and frees users from the tedious radiometric calibration and tone mapping steps.;Throughout this work, extensive experiments on various real and synthetic image data are conducted to evaluate the performance of the proposed algorithms.