A Multi-scale Patch Approach with Diffusion Model for Image Dehazing

Enhancing visual clarity in single images through dehazing is crucial for improving computational processes like object recognition and semantic segmentation. The advancements in deep learning, especially with Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), have significantly advanced image dehazing techniques. Notably, denoising diffusion probabilistic models (DDPMs) have shown great potential in text-to-image synthesis and have been adapted for image restoration tasks, including dehazing. However, current DDPM applications mainly use a patch-based approach and have not fully explored the benefits of training on multi-scale patches. Recognizing that objects in images often display similar textures and hues across various scales, there’s a need for multi-scale training approaches to better identify these features. We introduce a novel framework, the Multi-Scale Patch-Based Denoising Diffusion Probabilistic Model (MP-DDPM), which leverages multi-scale patch sampling during training to enhance the dehazing capabilities of the conditional diffusion model. To our knowledge, this constitutes the first instance of diffusion models being expressly tailored to the task of single image dehazing. Our extensive experimental evaluations, conducted on well-established dehazing benchmark datasets, corroborate the preeminence of the MP-DDPM over existing methodologies.