Robust vehicle re-identification under strong illumination and haze conditions via multi-scale feature enhancement

Vehicle re-identification (Re-ID) is essential for intelligent transportation, yet its performance is often hindered by strong illumination and haze, which obscure vehicle appearance and reduce matching accuracy. To address this, we propose a multi-scale feature enhancement framework that improves robustness under adverse visual conditions. A dual-branch preprocessing module first suppresses glare and restores hazy details. A convolutional–transformer backbone then extracts global semantics and local textures across multiple scales. To ensure condition-invariant representations, a contrastive domain adaptation loss aligns features between normal and degraded samples. Experiments on public datasets and a newly collected strong-light/haze benchmark demonstrate clear improvements over state-of-the-art methods in both mean Average Precision (mAP) and Rank-1 accuracy. These results confirm the effectiveness and practicality of our method for large-scale vehicle Re-ID in challenging real-world environments.