Low-light image enhancement aims to recover usable brightness, color fidelity, and fine details from underexposed imagery. In real-world conditions, however, enhancement models often suffer from noise amplification, color shifts, and local over-/under-exposure, with severe dark regions being particularly prone to texture loss and visual artifacts. To address these challenges, we propose DBENet, a low-light enhancement network that integrates attention modeling with cross-branch interaction. The proposed architecture adopts a dual-branch design to jointly model luminance enhancement and chrominance/detail restoration, introduces lightweight attention at the feature level to strengthen informative representations, and incorporates a cross-branch interaction module to enable complementary guidance between branches. This design facilitates unified optimization of brightness boosting, noise suppression, and structural detail recovery within a single framework. Moreover, considering the spatially non-uniform illumination and diverse degradation patterns in real low-light images, we establish a systematic evaluation protocol covering both real and synthetic low-light data, and conduct comprehensive quantitative and qualitative comparisons on multiple representative benchmarks, including paired and unpaired settings. Experimental results demonstrate that the proposed method achieves more consistent overall improvements across objective metrics and perceived visual quality, effectively alleviating color distortion and dark-region artifacts, and enhancing robustness and practical usability in real low-light scenarios.
Lin et al. (Mon,) studied this question.