The Application Of Guided Image Filter In Single Image Restoration

Image restoration has been an active area of computer vision. Haze anddefocus are two main reasons that cause the image degradation in practice.Although various methods have been proposed to deal with the problems of singleimage dehazing and image deblurring, some flaws are still remaining, such as thenoise magnification, high computational complexity and the poor estimationaccuracy of blur map.To solve these problems, this thesis will applies the guided image filter to dealwith the restoration of the single haze image and single defocus image based on thedegradation model and some prior constraints. The guided image filter can beconsidered as the first iteration of solving the image reverse problem, and it hasvery good edge-preserving smoothing property. This thesis introduces three mainprominent problems in great detail on the single image restoration.Firstly, since single image dehazing is a typical ill-posed problem, the thesisproposed an optimization framework and designed two specific guided image filters,which are joint bilateral filter and bilateral filter, to deal with this problem. Thejoint bilateral filter is applied to refine the rough transmission map, which isestimated based on dark prior, under the guidance of original image. Anotherbilateral filter is then applied to obtain the dehazing image, which can realizeimage denoising in the same time. Besides, the thesis introduces a color correctionfunction to redress the color distortion.Secondly, we derive an optical degradation model which enables us to adopt apixel scheme to realize image multi-focusing. It can greatly improve thecomputational efficiency.At last, a rough blur map of the input image, which reflects the amount ofdefocus blur at edge locations, is obtained by local robust blur estimation, and thenthe guided image filtering method is applied to propagate the blur value from edgelocations into the unknown regions. Based on a simple geometry prior ofphotograph, a depth map can obtained from transforming the refined blur map.Then we can directly produce different styles of images with the blur map anddepth map.