Image Dehazing based on Iterative-Refining Diffusion Model

One of the hardest problems in computer vision is image dehazing. In the single picture dehazing job, the approaches based on pixel domain mapping and prior knowledge of physical models have achieved amazing results. Sadly, most deep dehazing algorithms now in use have limited generalization capabilities, which makes it challenging to apply them to data samples with foggy conditions that exhibit a wide range of variation. We suggest an Iterative-Refining Diffusion Model built on the U-Net architecture to solve the issue. We show that the suggested approach may be used to the dehazing problem. It is based on Denoising Diffusion Probabilistic Models (DDPM) [14] and the denoising score matching. An empirical data distribution is created from the conventional normal distribution by a sequence of repeated refining stages that are comparable to the Langevin dynamics process. The U-Net architecture [27], the model's network architecture, is trained with dehazing targets to progressively eliminate different haze levels from the output. Extensive analyses demonstrate that the proposed model outperforms the state-of-the-art methods on multiple benchmarks.