基于DINOv2到EfficientViT的多损失蒸馏Patch特征少样本异常检测OA

To address the problem in few-shot industrial anomaly detection where large-scale models are difficult to deploy and lightweight models show limited performance,a lightweight anomaly detection method based on knowledge distillation is proposed.DINOv2 as the teacher model and EfficientViT as the student model,transferring knowledge are used to achieve high detection performance under resource constraints.A feature alignment adapter is designed to reduce the representation gap between the two models,and a multi-objective distillation loss combining Huber loss,cosine similarity loss,and Gram matrix loss guides the student from multiple dimensions of magnitude,direction,and spatial correlation.Experiments on the MVTec AD dataset show that the student model,trained only with normal samples,achieves a 0.170 improvement in image-level AUROC over the baseline model without distillation,while maintaining comparable performance to the teacher model with about 23%fewer parameters and 4 times faster inference.The effectiveness of the proposed framework is confirmed by ablation and visualization studies.