Prior-Anchored Local Statistical Representation Rectification for Low-Light Image Enhancement

Existing low-light image enhancement methods are increasingly built on deep representation learning. However, most of them still encode features as deterministic points in the embedding space, making it difficult to capture the statistical uncertainty. Under complex low-light scenes, latent neighborhood relations tend to reflect degradation patterns rather than semantic content or local structures, which leads to statistical shifts and reduced discriminability of representations. Therefore, we propose Prior-Anchored Local Statistical Representation Rectification (PaLSR), which models low-light enhancement as a representation learning process driven by local statistical rectification. PaLSR first learns a normal-light statistical prior with K Gaussian anchors, which provides a shared coordinate system for local representation. Instead of enhancing degraded features in the Euclidean embedding space, PaLSR represents each feature as mean-variance offsets relative to these anchors, describing its content displacement and uncertainty. With these offsets rectified, local representations are reorganized under complex low-light degradations. Extensive experiments on multiple low-light benchmarks and different network architectures show that PaLSR achieves consistent improvements in restoration quality. These results validate the effectiveness of prior-anchored local statistical representation rectification under complex degradation conditions.