SG-DDPM: Semantic-Guided Diffusion Model for Low-Light Image Enhancement

Recent advancements in low-light image enhancement (LLIE) through deep learning have been significant. However, these methods often struggle with effectively handling noise in dark regions, hindering the recovery of high-quality texture details obscured by noise. Moreover, existing approaches typically treat the entire image as a uniform entity, neglecting the semantic nuances of distinct regions. This oversight may cause a network to deviate from an area's original color without semantic priors. In response to these challenges, we introduce the Semantic-Guided Denoising Diffusion Probabilistic Model (SG-DDPM). Leveraging a diffusion model for LLIE, our approach employs a sequence of denoising refinement processes to restore realistic details in dark areas. We have optimized the inference process of the diffusion model to enhance its speed. Additionally, we integrate the Segment Anything Model (SAM) to extract semantic information from low-light images, guiding the diffusion model through conditional guidance. Experimental results demonstrate that SG-DDPM exhibits competitive performance across three image enhancement datasets, showcasing improvements in quantitative metrics and visual quality.