Residual Feature-Driven Knowledge Distillation for Reliable Open-Set Scene Understanding Under Distribution Shift

Reliable scene understanding under open-world conditions requires intelligent perception systems to accurately recognize known semantic categories while remaining robust to out-of-distribution (OOD) inputs, distribution shifts, and uncertain environmental conditions. This challenge becomes increasingly important for resource-constrained edge intelligence, where lightweight models are expected to provide reliable predictions without sacrificing computational efficiency. Although knowledge distillation has achieved remarkable success in compressing deep neural networks, existing methods primarily transfer classification semantics and often neglect the uncertainty representations that are critical for reliable open-set perception. To address this issue, we propose Residual Feature-driven Knowledge Distillation (RFKD), a lightweight uncertainty-aware distillation framework for reliable open-set scene understanding under distribution shift. Instead of directly distilling output confidence or energy scores, RFKD reconstructs uncertainty within the student’s latent feature space through a compact residual uncertainty branch. The proposed framework combines confidence-aware supervision, relational uncertainty distillation, and energy-guided relative ordering to preserve teacher-induced uncertainty geometry while enabling the student to learn discriminative feature-level uncertainty representations. The present study is evaluated on unimodal image data and does not claim empirical validation for multimodal perception. Extensive experiments on CIFAR-100 using multiple OOD benchmarks demonstrate that RFKD consistently improves uncertainty estimation while maintaining high computational efficiency. Compared with the ResNet-50 Teacher (Energy), RFKD increases the average AUROC from 0.7793 to 0.8446 while reducing the model size from 23.71 M to 11.29 M parameters and computational complexity from 1.31 G to 0.56 G FLOPs; the ImageNet-style student baseline obtains an AUROC of 0.7528. A score-specific sensitivity analysis shows that the energy detector is strongest when the auxiliary ordering score uses the residual branch alone, whereas moderate coupling with the student’s log-sum-exp potential improves the standalone OOD head. These results demonstrate that explicitly modeling representation-level uncertainty offers an effective and efficient solution for reliable scene understanding, providing a practical reliability enhancement for future intelligent perception systems operating in open and dynamic environments.