DehazeDiff: When Conditional Guidance Meets Diffusion Models for Image Dehazing

The existing deep dehazing methods usually adopt end-to-end pixel-level optimization, generative adversarial networks or normalized flow methods, which would suffer several problems, e.g. smooth recovery results, unstable training and specific network structure design. Recently, the denoising diffusion probability model (DDPM) has shown great potential in image generation. Consequently, we propose a new DehazeDiff framework for image dehazing based on DDPM, aiming to explore a new solution. Firstly, a feature extraction network is introduced to extract the potential features from haze images to clear images, which are then fed into the diffusion model for further feature refinement. It can improve the learning ability of the model, and realize the haze removal. Secondly, by learning the noise distribution on the basis of the given hazy image and the clear image, the feature-guided conditional guidance is used in the image generation process to fully learn the input image features and improve the image quality. Experimental results on both indoor and outdoor datasets show that the proposed method outperforms the mainstream methods in terms of objective index and visual effect.