A Study On Dehazing With Gaussian Process Regression And Training Example Searching

Outdoor images captured in the haze weather have low visibility and contrast,which causes difficulties to both human perception and computer vision.Therefore,researchers have devoted great efforts to haze removal for image restoration and enhancement.In recent ten years,knowledge based priors have been proposed to capture deterministic or statistical properties of hazy images for dehazing.But prior-based methods are not applicable for most hazy images.Therefore,data-driven approaches have become prevalent for the hazy process and achieve good results.The core of the data-driven model is training data and the performance of the model can be greatly improved by optimal training examples.However,few methods study on training example searching.In the other hand,many data-driven models have poor accuracy and great error in the regression,which leads to some bad haze removal results.Therefore,we study on accuracy increasing and training example searching.In our paper,we first propose a two-layer Gaussian Process Regression to learn the relationship between the hazy image and the transmission map.We proof the advantage of the two-layer model in our paper.The model has high accuracy and small error in the regression.Meanwhile,based on concentration and distribution of haze in the training images,we conduct a two-step selection method to select suitable training set,which further optimizes the regression method.The other study in our paper provides an input-adaptive searching method.For every test example,the searching method automatically collects suitable training examples which have high correlation with test example.Compared to the fixed training set selected by the two-step selection process,the searching method is input-adaptive and has strong robustness.Due to the searching process of every test example,the consumption of time is high,so we also propose two fast searching methods which are k-d tree and Hamming embedding to increase efficiency.Experimental results on the hazy image dataset show the clear and real results of our two methods compared with the state-of-the-art dehazing methods.