PRIN: A Perceptual-Structural Retinex-Inspired Network for Low-Light Image Enhancement
Low-light image enhancement (LLIE) focuses not only on visibility recovery but also on preserving structurally meaningful content and perceptual naturalness. Existing learning-based methods may achieve favorable pixel-wise fidelity while still producing over-smoothed textures, color deviations, or spatially incoherent illumination. Retinex-inspired methods provide interpretable decomposition into reflectance and illumination components, however, the asymmetric restoration roles of these components are rarely encoded through branch-specific frequency-domain refinement. To address this issue, we propose a Perceptual-Structural Retinex-Inspired Network (PRIN), which integrates Retinex priors with restoration-role-aware enhancement. PRIN employs a shared encoder and asymmetric decoders to estimate reflectance and illumination components. The reflectance branch performs attention-guided and edge-aware wavelet refinement to preserve textures and boundaries, whereas the illumination branch promotes spatially coherent brightness adjustment. After the recomposition of reflectance and illumination branches, a curve-based post-composition enhancement head calibrates residual tone and exposure deviations. Experiments on LOL-v1, LOL-v2-Syn, and LOL-v2-Real show that the proposed network ties for the best structural similarity index (SSIM) and obtains the second-lowest learned perceptual image patch similarity (LPIPS) on LOL-v1, achieves the best SSIM and LPIPS on LOL-v2-Syn, and obtains the highest peak signal-to-noise ratio (PSNR) and the second-lowest LPIPS on LOL-v2-Real.