LGHA-Net: Local-global hybrid attention for pixel-wise multi-illuminant estimation under low-bit-depth degradation

Pixel-wise illuminant estimation in multi-illuminant scenes remains challenging when spatially varying illumination is coupled with low-bit-depth degradation, because quantization weakens local chromatic boundaries while distant regions may still share correlated illumination context. To address this problem, we propose LGHA-Net, an RGB-only Transformer-enhanced U-Net for pixel-wise multi-illuminant estimation. The core component is a Local-Global Hybrid Attention (LGHA) block inserted at the bottleneck stage, where 7 × 7 local window attention preserves illumination-boundary details and global self-attention models long-range illumination-context dependencies. A learnable residual weighting strategy adaptively balances the local and global updates. The network predicts a two-channel UV chromaticity map, which is converted into RGB illumination vectors so that angular supervision and evaluation are performed in the same illumination-vector space. Experiments on the LSMI dataset show that LGHA-Net reduces the overall mean and median angular errors from 1.80 ∘ and 1.47 ∘ to 1.59 ∘ and 1.23 ∘ , respectively, compared with the U+PP baseline. Controlled low-bit-depth experiments further demonstrate improved robustness under 12-bit and 10-bit quantization. Ablation studies show that the LGHA backbone provides the main gain, while the optional physics-constrained post-processing further refines tail cases.