Knowledge distillation-based unsupervised anomaly detection has achieved strong performance in industrial inspection. However, existing Reverse Distillation (RD) methods suffer from two fundamental limitations: the bottleneck compresses semantically entangled CNN features into a one-class representation that inadequately characterizes normal structural regularity, and the pairwise distillation objective leaves the co-learned teacher-student geometry semantically underconstrained. To address these issues, we propose Semantically Grounded Reverse Distillation (SGRD), which integrates a frozen vision foundation model (DINOv2 in this work) as a fixed semantic reference into the RD framework. Specifically, a foundation-aware semantic bottleneck conditions the bottleneck encoding pathway with foundation-model patch semantics to produce a structurally coherent one-class code. Furthermore, a foundation-semantic representation anchoring module jointly constrains both teacher and student representations toward the foundation-model semantic subspace, preventing semantically arbitrary representation geometry. Experiments on MVTec AD achieve 98.5% image-level AUROC and 98.3% pixel-level AUROC, while experiments on VisA achieve 99.8% image-level AUROC and 94.2% pixel-level AUPRO, with the proposed method attaining the strongest overall average performance among the compared methods on both benchmarks.

Few-shot object detection suffers from limited annotations, redundant background interference, insufficient feature interaction, and severe sample imbalance. Existing meta-learning-based methods extract class prototypes from support images but often fail to effectively suppress background noise or align channel-wise features between support and query branches. To address these issues, we propose a few-shot object detection method based on background reconstruction and multi-channel interactive feature fusion. First, a background reconstruction module is designed to suppress redundant background interference by applying random region masking to support set images, thereby generating robust class prototype features that are resistant to background noise. Second, a multi-channel interactive feature fusion module is designed, which leverages depthwise separable convolution to enable effective channel-wise feature interaction and information alignment between support class prototypes and query features, thereby enhancing cross-branch feature interaction and fusion. Finally, to address the uneven sample distribution and the foreground–background imbalance in few-shot scenarios, we proposed a category-aware weighted loss. By appropriately weighting the contributions of different object categories and background samples, the proposed loss encourages balanced optimization, resulting in faster convergence and improved detection performance. Experimental results demonstrate that the proposed method improves detection accuracy and generalization performance under few-shot settings. On the Pascal VOC dataset (Split1), the proposed method achieves 45.1%, 62.9%, and 67.1% under 1-shot, 3-shot, and 10-shot settings, respectively, outperforming the baseline; consistent improvements are also observed on the MS COCO dataset and the DIOR dataset.

Knowledge distillation-based unsupervised anomaly detection has achieved strong performance in industrial inspection. However, existing Reverse Distillation (RD) methods suffer from two fundamental limitations: the bottleneck compresses semantically entangled CNN features into a one-class representation that inadequately characterizes normal structural regularity, and the pairwise distillation objective leaves the co-learned teacher-student geometry semantically underconstrained. To address these issues, we propose Semantically Grounded Reverse Distillation (SGRD), which integrates a frozen vision foundation model (DINOv2 in this work) as a fixed semantic reference into the RD framework. Specifically, a foundation-aware semantic bottleneck conditions the bottleneck encoding pathway with foundation-model patch semantics to produce a structurally coherent one-class code. Furthermore, a foundation-semantic representation anchoring module jointly constrains both teacher and student representations toward the foundation-model semantic subspace, discouraging semantically arbitrary representation geometry. Experiments on MVTec AD achieve 98.5% image-level AUROC and 98.3% pixel-level AUROC, while experiments on VisA achieve 99.8% image-level AUROC and 94.2% pixel-level AUPRO, with the proposed method attaining the strongest overall average performance among the compared methods on both benchmarks.