Progressive Spatial-Frequency Domain Fusion for Low-Light Image Enhancement

Images captured in low-light conditions often suffer from insufficient luminance, detail loss, and increased noise, which negatively affects both visual quality and the performance of downstream visual tasks. Existing low-light image enhancement methods typically perform pixel-level adjustments in the spatial domain or or follow a multi-stage approach, where global adjustments are made in the frequency domain before enhancement in the spatial domain. However, these methods do not fully leverage frequency domain information, leading to issues such as noise amplification or detail loss in the enhanced images. To address these problems, we propose a novel Progressive Spatial-Frequency Domain Fusion (PSFDF) network for low-light image enhancement. Specifically, based on the U-Net architecture, we progressively extract spatial and frequency domain features at different scales in the Encoder. In the Decoder, we introduce skip connections and design a Spatial-Frequency Domain Fusion Module (SFDFM) to fuse the futures from both domains. This progressive strategy enables the extraction and fusion of spatial and frequency domain features across multiple scales, leveraging the advantages of frequency domain information to enhance global luminance and improve details, thus achieving complementary information and producing detail-rich, naturally realistic enhanced images. Quantitative and qualitative results on multiple low-light datasets demonstrate that the proposed PSFFN outperforms state-of-the-art methods, significantly enhancing both the visual quality and recognizability of low-light image enhancement results.