A novel edge-aware À-Trous filter for single image dehazing

Single image dehazing has been a particularly important and active area of research in computer vision. One of the key problems of physics-based methods is the fast estimation of the transmission. Previous single image based methods, however, require some time-consuming optimization algorithms, making them impracticable for real-time applications. In this paper, we propose a novel edge-aware filter based on the orthogonal B-spline wavelets. The filter is constructed by equipping low-pass filters with edge-stopping functions and leads to an efficient method for edge-preserving smoothing, which can be used in the transmission estimation according to our analysis of the dichromatic atmospheric scattering model and the dark channel prior. Instead of using conventional multiresolution processing which leads to halo effects in previous methods, we adjust Mallat algorithm to process signals from coarse scales to fine scales and obtain halo-free image filters without increasing any computational complexity. Compared with the bilateral filter, our method is much faster and has better behavior near the edges as illustrated by our noise removal testing. Extensive experiments on single image dehazing with comparable or better performance to the state-of-the-art techniques show the effectiveness of the proposed method.