Enhanced RefineDNet for Single Image Dehazing
Jingyu Ren, Lei Yang · 2024
Haze-free images are the preconditions of many vision systems, and thus single image dehazing is of great significance in computer vision. RefineDNet [1] is a two-stage weakly supervised dehazing framework which combines the merits of prior-based and learning-based approaches together. However, under high image saturation or adverse weather conditions, its performance is limited. In this work, a modified RefineDNet is proposed by leveraging the saturation line prior (SLP) [2] method in the first stage and a plug-in based on a patch-based diffusion model [3] for image preprocessing to enhance the dehaze ability of the framework. Extensive experiments demonstrate that our modified RefineDNet framework achieves superior haze removal and produces visually pleasing results.