Answers are grounded in this paper and cite their sources.
Low-light image enhancement is challenging due to the strong entanglement between illumination degradation and sensor noise in RGB space, which often causes noise amplification and color distortion. We first present region-aware quantitative evidence that a luminance–chrominance transform provides a more separable degradation structure. On the LOLv2-Synthetic training set with 900 paired samples, YCrCb consistently reduces residual coupling across global, edge, dark, and texture regions, indicating that its decoupling advantage is not merely dominated by smooth background areas. Motivated by this analysis, we propose DuBNet, a luminance–chrominance decoupled framework with a luminance branch (Lum-Net) and a chrominance branch (Chr-Net), where Lum-Net incorporates a lightweight BrightnessEmbed module for exposure-aware adaptation. A fusion/refinement network (Fus-Net) then performs alternating feature-guided fusion using a compact Feature Guidance Block and an efficient attention residual block (ECACoordCAB) that combines ECA and Coordinate Attention for channel- and position-aware modulation. Experiments on LOLv1, LOLv2-Synthetic, LOLv2-Real, and FiveK show that DuBNet achieves competitive fidelity and perceptual quality against strong baselines. We further provide a reduced-complexity distilled variant, DuBNet-KD, to examine whether the proposed decoupled representation and fusion strategy remain effective under a smaller channel width.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: