Low-light image enhancement via enhancement-guided adaptive refinement and progressive rational mapping

Abstract

Low-light images suffer from coupled degradations of insufficient illumination, low contrast and amplified noise. Curve-based methods needing no paired supervision are efficient, but they correct luminance and rarely address the noise that amplification makes visible. This paper couples an iterative curve estimator with a refinement module conditioned on the enhancement process itself. The enhancement module predicts a per-pixel curve parameter and a gate for each of eight iterations; the refinement module consumes a twelve-channel prior built from those quantities and performs spatially adaptive restoration in the regions where amplification is largest. The refinement module carries the method. Over three independently trained seeds, removing it costs 0.162±0.028 structural similarity index, several times its own spread, against 0.046 across the three main curve families (gamma, rational, polynomial), each measured over three seeds. Removing only its guidance prior costs 0.050±0.049, positive on every seed but not distinguishable from zero on three runs, so we claim the module without apportioning its value internally. The model uses 0.271 M parameters and 35.35 GFLOPs at 256×256. Results are means over three seeds with their spread, since a different seed alone moves LOLv2-Real peak signal-to-noise ratio (PSNR) by up to 1.52 dB. Three findings qualify the design and we report them as measured. Gamma correction attains a higher PSNR than the rational mapping we adopt, so we present it as a characterised design choice, not a demonstrated optimum. The module removes the synthetic noise it is trained on but recovers only 0.33±0.09 dB on real sensor noise, under two per cent of the measured noise level. And performance varies by more than ten decibels with input illumination. Model selection uses paired references from held-out training scenes, so the training objective is reference-free while the pipeline is not.