DDSCNet: Dual-Domain Synergistic Downsampling and Dual-Branch Feature Calibration Network for Low-Light Image Enhancement

Real-world low-light scenarios are complex, and annotated data is scarce. Meanwhile, existing supervised and mainstream unsupervised low-light image enhancement methods typically rely on large-scale paired labeled or unpaired normal-light data for training, which constrains the cross-scene generalization capability of these models. Furthermore, zero-shot low-light enhancement methods based on Retinex theory still exhibit notable performance shortcomings in dark-region noise suppression and illumination component estimation accuracy. To address these challenges, this paper proposes a zero-shot architecture for low-light image enhancement based on dual-domain synergistic downsampling and dual-branch feature calibration, which effectively resolves the core dilemma of the inaccessibility of annotated training data. Specifically, we construct a dual-domain downsampling mechanism with frequency-domain and wavelet complementarity, which provides effective priors for pre-denoising to suppress noise. A dual-branch feature calibration module centered on bidirectional correction and gated weighting is designed to achieve high-fidelity halo-free illumination estimation. To tackle the problems of color distortion and insufficient enhancement in extremely dark scenes, we further propose a multi-dimensional constrained naturalization enhancement module. Extensive experiments on the LOL-v1 and LOL-v2 datasets demonstrate that the proposed method achieves outstanding real low-light enhancement performance and competitive visual results.