SG-CIDNet: structure-guided low-light image enhancement network in HVI space

作者：Chenyang Zhou, Zhongxiang Liu, Shichu Zhang · 年份：2026 · DOI：10.1117/12.3109885 · 研究领域：Image Enhancement Techniques、Generative Adversarial Networks and Image Synthesis、Computer Graphics and Visualization Techniques

Images captured in suboptimal lighting conditions are often characterized by reduced visibility, obscured structural textures, and significant noise corruption. These factors collectively undermine the reliability of subsequent computer vision applications. To mitigate these limitations, we present a novel framework named SG-CIDNet, a multi-branch low-light image enhancement network that synergizes illumination priors with a structure guidance mechanism. To achieve efficient separation between intensity (I) and chromatic (HV) data, a two-stream framework rooted in the HVI domain is developed. Acting as a supplementary pathway, the Structure Guidance Network (SGN) leverages dilated convolutions at multiple scales. This module is tasked with estimating illumination maps and and generate Retinex-based structural priors to guide feature reconstruction. The backbone incorporates cross-attention mechanisms to facilitate interaction between luminance and color branches, while CBAM modules are integrated to refine dark regions and edge details. Moreover, a Multi-Scale Adaptive Gating Fusion module is integrated to adaptively merge deep features with the generated structural priors. Empirical tests on the LOL-v1 and LOL-v2 datasets indicate that our method outperforms state-of-the-art competitors. Specifically, it excels in both metric scores and perceptual quality, offering enhanced detail preservation and natural color recovery.