Restoration Of Single Image Degraded By Haze

The ultimate goal of computer vision is to understand and explain the rich three-dimensional world. With the development of basic science (cognitive neuroscience and visual computing technologies), and with the growing of computer cost performance, the applications of computer vision become more and more extensive, such as security detection, industrial measurement and monitoring, automatic navigation, monitoring without anyone and so on. But the current visual system in the design cannot do with the disturbance of the adverse weather. So it can only be used in fine weather conditions. However, for any visual systems applied in outdoor, they inevitably face all kinds of weather conditions, such as haze, fog, rain, snow and so on. A robust outdoor vision system must have the ability to work well under any weather conditions. This is an issue of great practical significance.A simplified degradation model and the premise of the approximate dark existence is the physical fundamental of the method. Using the image segmentation, object recognition, belief propagation inference and Laplace repair, the restoration of a single haze image is realized. The main contents of this dissertation are as follows:(1) Repair transmission map using matting Laplacian matrix. Firstly, the dark channel is calculated from a single input image to obtain rough transmission map. Then, the input image is segmented and the transmission map is uniformly segmented. Finally, based on the segmentation the transmission map is repaired using laplace matting matrix.(2) Estimate the tranission map using Belief propagation. When the assumption of the approximate dark existence does not hold,, the transmission map is rectified by belief propagation algorithm. The rectified map accurately reflects the influence of haze..(3) Estimate atmospheric light automatically. Firstly, the sky region separated from the input image. Then, the value of atmospheric light is decided by the brightest pixel in the sky region. For the image without the sky region, the belief propagation algorithm is used to exclude the wrong points in transmission map. Then, the brightest pixel in the region is chose as the atmospheric light.The experimental results demonstrate the method is able to remove the haze layer as well as provide a reliable relative depth map.