Image Dehazing Method Based on Diffusion Model

In this paper, we design a distribution transformation-based method for accelerating image dehaze for diffusion models. The distribution transformation architecture solves the problem of blind transformation of the distribution of the diffusion model during image dehaze processing by providing additional constraints for the model. The efficient use of the dual network structure in the model inference process is achieved through a redesigned fusion sampling method. Meanwhile, in order to ensure the consistency of the output data distribution, the data are distribution normalised after each fusion sampling. The accelerated sampling strategy proposed in this paper optimises the network structure employed by the model and effectively reduces the computational burden of the model, while maintaining the image quality and improving the inference efficiency. Experiments on multiple synthetic and real foggy image datasets show that the network structure proposed in this paper significantly improves the inference speed of the model and achieves a good dehaze effect, with clear details and thorough fog removal in the dehazed image, which effectively restores the background information and the real colours of the image.