Deep learning models trained on curated datasets often fail when deployed in complex visual scenes involving illumination drift, occlusion, adverse weather, and cross-domain sensor variation. We propose an end-to-end framework that jointly addresses representational robustness and domain-invariant transfer through two integrated components. First, a scene-aware contrastive learning strategy generates positive pairs via physically grounded scene transformations—simulating fog, rain, illumination gradients, and structured occlusion—and employs an adaptive difficulty weighting mechanism that steers the contrastive gradient toward degradation conditions the encoder finds hardest to resolve. Second, a multi-scale domain-invariant feature alignment method applies distributional matching at multiple depths of the encoder hierarchy, coupled with a cross-scale consistency regularizer that prevents semantic drift between layers. The framework requires no target-domain data and operates with minimal or no annotation. Experiments on PACS and a cross-domain scene recognition benchmark (BDD-City) demonstrate that our method outperforms representative supervised, self-supervised, domain generalization, and domain adaptation baselines, with average cross-domain accuracy gains of 3.5 and over 4% points, respectively. Gains are most pronounced under severe degradation conditions, where baseline methods suffer the steepest performance drops. Ablation analysis confirms that each component contributes independently and that their interaction is synergistic.
Wenjun Li (Thu,) studied this question.