Multi-scale spatial diffusion under frequency information-guidance For low-light image enhancement

Denoising Diffusion Probabilistic Models (DDPMs) have recently been employed for low-light image enhancement (LLIE), formulated as a low-light image conditioned generation process that maps Gaussian noise to a normal light image. However, their restorations usually depend on only naive condition information, which lacks enhanced information guidance. In this case, DDPMs tend to calculate the region-level enhancement inflexibly and hardly generate accurate background structures and noise-covered details. To respond to this issue, we propose a Multi-scale Spatial Diffusion model under Frequency Information Guidance (MSFIG-Diff) for low-light enhancement. Specifically, the MSFIG-Diff exploits the low-light image as condition information, and incorporates features enhanced by a regression network from the low-light input as auxiliary guidance. Towards more effective information extraction for guidance in low-light images, a Fourier convolution-based regression network is employed to decouple the brightness and noise information. Furthermore, to compensate for the missed details during feature down-sampling, a spatial attention-based multi-scale mechanism is proposed to gradually integrate multi-level information into the denoising network for accurate restoration. Extensive comparisons of the widely used benchmarks demonstrated the effectiveness of introducing enhanced information guidance and verified the superior performance of our method.