VEX: Chroma-Stable Virtual Exposure Routing for Low-Light Image Enhancement

Abstract

Low-light image enhancement (LLIE) seeks to improve visibility while suppressing amplified noise, preserving color, and preventing highlight over-enhancement. Most existing methods infer a normally exposed image from a single observed representation, requiring one feature stream to reconcile shadow brightening, highlight protection, denoising, and chromatic correction. We propose VEX, a chroma-stable virtual exposure routing network for single-image LLIE. A Chroma-Stable Virtual Exposure Generator (CVEG) decomposes the input into luminance and log-chroma components and produces five learnable virtual exposure states by perturbing luminance in a bounded space under a shared log-chroma constraint. A shared multi-scale encoder extracts comparable features from all states. At each of four scales, a Noise-Saturation-aware Exposure Router (NSER) combines learned feature evidence with explicit luminance, mid-tone, saturation, and detail-variation priors to predict pixel-wise softmax weights over the exposure states. An Exposure State Mixer (ESM) then performs gated multi-view recalibration of the routed bottleneck, after which a routed-skip decoder predicts a residual correction. VEX therefore casts LLIE as spatially adaptive selection among internal exposure hypotheses rather than as direct single-state regression. Extensive experiments on CDD-11 and LOL demonstrate that VEX consistently outperforms representative traditional and learning-based methods across fidelity, structural, and perceptual criteria.