Atmospheric light estimation in hazy images based on color-plane model

In this paper, a novel air light recovery method based on color-plane model is proposed. The method aims to improve the robustness of air light estimation for single image dehazing applications. Traditional methods estimating air light rely on user input or dense haze region. It is unstable when dense haze region is not included. The methods based on the color-line model and haze-line model achieve significant improvements on air light estimation without identifying dense haze regions. However, both of the models are lack of generality. Therefore, we propose the color-plane model which combines the color-line model and haze-line model. It is motivated by the fact that colors with a common direction widely exist and tend to stay together in natural images. These colors spread on color-plane after being blended with haze, which indicates the orientation of the air light. An algorithm based on region growing is designed to extract potential color-plane in haze images. RANSAC is employed to estimate the air light orientation. The magnitude is estimated based on the assumption that the intensities of several bright pixels in different depth ranges should be similar. A novel algorithm based on energy minimization is proposed to estimate depth ranges. Experimental results show that the proposed method performs better than state-of-the-art methods.