| 基于DINOv3的无监督微小目标异常检测方法 |
| 点此下载全文 |
| 引用本文：王铭巍,杜奕,龚晓立,郑世良,曹晓夏.基于DINOv3的无监督微小目标异常检测方法[J].上海第二工业大学(中文版),2026,43(2):201-209 |
| 摘要点击次数: 258 |
| 全文下载次数: 5 |
|
| 基金项目:国家自然科学基金(41672114, 41702148) 资助 |
|
| 中文摘要:使用预训练视觉模型进行异常检测是近年来工业视觉检测领域的重要研究方向, 如何获取通用特征表示是提升异常检测模型性能的关键因素。为此, 针对无监督微小目标异常检测问题, 本文提出一种基于DINOv3 的无监督微小目标异常检测算法, 命名为DINOv3-Dinomaly Frozen Transfer Anomaly Detection (D3D-FTAD)。该模型将DINOv3 视觉大模型作为特征编码器, 与Dinomaly 框架有机结合, 实现输入图像的高质量特征提取。实验结果表明,D3D-FTAD 模型在图像级AUROC、像素级AUROC 等各项检测指标上均优于CDO、MSFLOW 等经典异常检测模型, 表明了高质量的通用视觉表示有助于提升异常检测性能。 |
| 中文关键词:视觉异常检测 无监督学习 DINOv3 特征表示 |
| |
| Unsupervised Tiny Object Anomaly Detection Method Based on DINOv3 |
|
|
| Abstract:Using pre-trained visual models for anomaly detection has emerged as a significant research direction in the field of industrial visual inspection in recent years. Developing effective universal feature representations is a key factor in enhancing the performance of anomaly detection models. To address the challenge of implementing unsupervised tiny object anomaly detection, this paper introduces an unsupervised small target anomaly detection algorithm based on DINOv3, named DINOv3-Dinomaly Frozen Transfer Anomaly Detection (D3D-FTAD). This model seamlessly integrates the DINOv3 visual large model as a feature encoder with the Dinomaly framework, enabling high-quality feature extraction from input images. Experimental results demonstrate that the D3D-FTAD model outperforms classic anomaly detection models such as CDO and MSFLOWin various detection metrics, including image-level AUROC and pixel-level AUROC, underscoring the significance of high-quality universal visual representations in enhancing anomaly detection performance. |
| keywords:visual anomaly detection unsupervised learning DINOv3 feature representation |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|