DPCDNet: dual-path color-decoupled network for low-light image enhancement

Abstract

The fundamental bottleneck in low-light image enhancement is the persistent coupling between luminance and chrominance throughout the pipeline, from input representation and feature interactions to the final output. This coupling prevents existing methods from simultaneously achieving brightness enhancement and color fidelity. To address this problem, we propose a dual-path color-decoupled network, termed DPCDNet, which progressively decouples luminance and chrominance at three stages: source, interaction, and output. Specifically, the Gated Color Fusion (GCF) module fuses complementary luminance cues from HVI and YUV spaces to achieve initial decoupling; the Complementary Mamba Interaction Module (CMIM) combines Mamba-based global modeling with frequency-aware feature enhancement to facilitate complementary interaction between luminance and chrominance features while mitigating feature re-coupling. The Illumination-Aware Attention Module (IAAM) uses illumination-branch features to guide spatially adaptive refinement, thereby completing decoupling at the output stage. On datasets such as LOLv1, LOLv2, MIT-Adobe FiveK, and DICM, DPCDNet achieves significantly better PSNR and SSIM than existing methods. Our method helps to mitigate the full-pipeline luminance-chrominance coupling problem, offering a new approach for enhancement techniques that balance brightness and color.