In-vehicle low-light face image enhancement with physical-semantic constrained diffusion and gated selective...

Driver fatigue monitoring in intelligent cockpits faces unique challenges in nighttime scenarios: the coexistence of extreme non-uniform illumination and high-intensity sensor noise. Existing algorithms hit bottlenecks: discriminative models, limited by regression-to-the-mean effects, over-smooth high-frequency features like eyelids and pupils; meanwhile, probabilistic generative models, lacking structural constraints, generate spurious ocular structures (semantic hallucinations), misleading downstream tasks. To address this, we propose STDN-MD, a physical-semantic synergistic restoration framework based on Retinex theory. Eschewing black-box mapping, it constructs a cascaded paradigm progressing from physical illumination estimation to semantic texture restoration. To alleviate the noise propagation tendency of linear state-space scanning in standard Vision Mamba under noisy low-light conditions, we design a Gated Recurrent 2D Selective Scan (GR-SS2D) module. Using nonlinear gating to replace linear recurrence, this module dynamically blocks noise propagation, constructing a precise implicit illumination field. For reflectance restoration, we introduce face parsing masks as strong empirical priors. Injecting facial topological information into the conditional diffusion process constrains uncertainty, forcing the recovery of authentic structures and suppressing identity drift. Experiments on our YaWDD-Dark benchmark show STDN-MD achieves an LPIPS of 0.068, reducing perceptual error by 52.8% versus state-of-the-art methods. The method significantly improves visual perception and exhibits superior performance in preserving key fatigue features, achieving dual breakthroughs in visual fidelity and machine perception accuracy. The code is available at https://github.com/PanchengZhang/STDN-MD.