A Dehazing Algorithm Based on Local Adaptive Template for Transmission Estimation and Refinement

Outdoor images are vulnerable to environment and may suffer various distortions. Therefore, preprocessing for images captured in bad weather is particularly important for computer vision system. One of the most common conditions is haze. Image dehazing, especially single image dehazing is a challenging topic since it's an ill-posed problem and needs to rely on extra information or prior. In this paper, we discussed the shortcomings of existing algorithms and proposed a novel step called Local Adaptive Template. The template is used in transmission estimation and transmission refinement. Starting from the target pixel, the template is extracted under the guidance of the similarity function and only contains pixels related to the center point, thus avoids the influence of adjacent objects, even those with blur edges. We then used the template to improve the Dark Channel Prior(DCP) and the Guided Filter(GF) respectively, and effectively avoided the block effect in DCP and the blur in GF. The obtained transmission map is much more accurate, and free from the halo effect. The dehazing result is much clearer and still looks natural without haze residual. Experiments on natural images and synthetic images show that our method achieves better dehazing results than several state-of-art algorithms and can adapt to different situations.