Advancing Ambient Lighting Normalization via Diffusion Shadow Generation

Ambient Lighting Normalization (ALN) finds extensive applications in images captured in complex real-world scenarios and can be extended to more general image restoration tasks. To address the ALN issue in real-world highresolution images, this paper proposes Advancing Ambient Lighting Normalization via Diffusion Shadow Generation. Specifically, we design a mask-free shadow removal network based on a $U$-Net architecture without global residual connections, which can effectively handle a wider range of image restoration tasks and exhibit high throughput during training on large-scale high-resolution images. During the training of the mask-free shadow removal model, we dynamically overlay diverse shadow conditions generated by a diffusion model onto the restored images, facilitating iterative learning within the shadow removal network. This results in an efficient general image restoration network capable of handling high-resolution images under complex ambient lighting conditions. Experimental results on official datasets confirm the superiority of our method.