HVR-SSLE: Hierarchical Visual Reasoning for Self-Supervised Low-Light Image Enhancement
Dongwon Choo, Qikang Deng, TaeWon Park, DoHoon Lee · IEEE Access · 2026
Low-light image enhancement (LLIE) is a fundamental problem in computational photography, aiming to recover images degraded by coupled noise, color distortion, and detail loss under insufficient illumination. While recent Transformer and diffusion approaches can improve perceptual quality, their high computational cost and reliance on small paired datasets limit practical deployment and reliable evaluation. In this work, we reinterpret LLIE as a hierarchical visual reasoning problem and propose HVR-SSLE (Hierarchical Visual Reasoning for Self-Supervised Low-Light image Enhancement), a compact recurrent framework that alternates low-level local refinement and high-level global restoration in a coarse-to-fine schedule for progressive enhancement. The recurrence is trained efficiently via a one-step gradient approximation, enabling multi-step refinement with low memory overhead. We further quantify train–test scene overlap in LOL-v1/v2, revealing substantial duplication and cross-split overlap that can inflate benchmark scores. To reduce reliance on LLIE-specific paired data, we train HVR-SSLE in a self-supervised manner on the general-purpose COCO dataset by synthesizing diverse low-light inputs using a parametric degradation curve with controllable cutoff, compression, and nonlinearity. Trained solely on COCO, HVR-SSLE contains only 0.34M parameters yet generalizes zero-shot to standard paired benchmarks (LOL-v1/v2 and LSRW) and real-world unpaired datasets (DICM, LIME, MEF, and NPE), achieving competitive PSNR/SSIM and the best PIQE/BRISQUE on LIME and MEF. Code is available at https://github.com/dwchoo/HVR-SSLE.