HVD-Net: low-light image enhancement in HSV space via continuous hue encoding and dual-branch restoration

Abstract

Low-light image enhancement (LLIE) aims to improve visibility while suppressing noise and correcting colour distortion. however, the strong inter-channel coupling in RGB space often causes colour instability during nonlinear enhancement, and commonly used HSV-like colour spaces exhibit discontinuities at hue boundaries, which may further introduce artefacts in transitional regions. to address these issues, we propose HVD-Net, a low-light image enhancement network in HSV space via continuous hue encoding and dual-branch restoration. specifically, the input image is first transformed into HSV space, where hue is represented by a continuous sine-cosine encoding to alleviate boundary discontinuities during convolutional learning. on this basis, an asymmetric dual-branch architecture is constructed to model luminance restoration and chrominance denoising separately, with cross-branch gated interaction at the bottleneck for complementary feature exchange. in the luminance branch, a Hybrid Channel-Spatial Attention (HCSA) module is introduced to adaptively enhance illumination distribution and contrast. in the chrominance branch, a Multi-Scale Feature Modulation (MSFM) module exploits low-frequency colour priors to guide detail recovery while suppressing chrominance noise. experiments on multiple paired benchmark datasets and real-world unpaired datasets show that HVD-Net achieves a favorable trade-off between enhancement quality and computational cost.