Research On Enhancement Of Images Degraded By Fog

Contaminated by fog, device-perceived images usually suffer from problems like contrast degradation, color deviation and so on, which results in a drastic deterioration in visibility. Taking traffic surveillance system for example, it is no doubt that both accuracy and robustness of the system will be influenced badly. Current dehazing methods are not taking noise into consideration so that due to the inherent defects of economical cameras large amounts of noise are often brought into the hazy image and exasperated after dehazing, which has a rather bad effect on dehazing quality.In this thesis, the enhancement techniques of foggy images, mainly on the model-based methods, are systematically studied. And edge-preserving denoising algorithms are specially studied for the purpose of image denoising after haze removal. The main work and achievements of this thesis include:(1) Considering that an ideal transmission map should be globally smooth, perseverating prominent edges and free of textures and details, a multi-scale fusion algorithm based on dark channel prior dehazing is proposed. By applying L0 gradient minimization to smoothing the pixel-wise transmission map while remaining salient edges, and removing the false high-frequency components in patch-wise transmission map using large scale Gaussian Filter, the final refined transmission map is gained through fusion. At last, a haze-free image is obtained with the use of optical model describing the function of haze. It is demonstrated that the proposed algorithm works better than other state-of-art approaches in halo effect and haze removal in small regions of depth discontinuity occluded by foreground. At the same time, compared with the accelerated dark channel prior algorithm based on guided filter, the proposed algorithm’s speed is also raised by 7.52 times.(2) Assuming that a pixel’s similar patterns should appear in the areas sharing the same or adjacent depth and transmission is the exponential function of depth, this thesis proposes a modified non-local means algorithm using transmission to conduct pre-classification and similarity weight refinement, aimed to remove the noise after applying dark channel prior image dehazing. By utilizing the transmission map provided by the process of haze removal, similar pixels pre-classification is carried out to speed up the algorithm. Then a transmission-based revising factor is formed to adjust the similarity weight of the pixels to be denoised. At the same time, an adaptively parameter optimizing model based on absolute deviation of gradients is built in order to select appropriate parameters according to the intensity of details in different local areas, which ensures an ideal detail preservation together with effective noise removal. Experiments demonstrate a better denoising power over original non-local means algorithm and state-of-the-art denoising methods for dehazing.(3) A GPU-based parallel algorithm is designed and realized under the architecture of CUDA. In order to speed up the proposed algorithms and meet the requirement of real-time application, time-consuming procedures like patch-like dark channel map acquisition, transmission refining and weight of similarity between pixels computation are assigned to GPU to conduct mission-level and pixel-level parallelization through texture attachment. Compared to original ones, parallelized algorithms can reach an acceleration rate of 134 or more.