Enhancing unsupervised shadow removal via multi-intensity shadow generation and diffusion modeling

Shadow removal is crucial for enhancing image quality and facilitating downstream computer vision tasks. However, acquiring paired shadow datasets is costly and challenging. This paper presents an unsupervised shadow removal algorithm leveraging a diffusion model. It initially generates a shadow-free image through a brightness enhancement network, which pairs with the original shadow image to train a shadow generation network. A shadow intensity control module ensures diverse shadow intensities, addressing data scarcity. During shadow removal, a resampling module constrains effects within shadow areas, while a boundary artifact removal network eliminates artifacts. Experimental results demonstrate the method’s superiority over existing unsupervised methods, achieving state-of-the-art performance on benchmark datasets with improved PSNR (+ 1.46 dB) and reduced RMSE ( \(-\) $-$  1.4) in shadow regions. The source code and pre-trained models are available at https://github.com/Donghui-Wang/SMGDM-SRA