融合扩散模型与知识蒸馏的无监督工业缺陷检测

doi:10.3778/j.issn.1002-8331.2506-0028

摘要/Abstract

摘要： 近年来，基于无监督学习的工业缺陷检测模型取得了显著的性能提升。然而，现有的缺陷合成策略依赖外部数据源，导致合成缺陷与真实缺陷存在较大差异，严重制约了模型的泛化性能。此外，现有的方法存在特征细节信息丢失问题，导致模型出现误检现象。为此，引入一种多源缺陷合成策略，协同利用扩散模型生成的图像和DTD数据集的图像合成更符合真实缺陷分布的缺陷样本。利用合成的缺陷样本微调教师网络对缺陷的表征能力，并引入异常屏蔽模块解决教师学生网络同构导致的过度泛化问题。构建细节修补模块，通过跨层级特征融合增强学生网络对教师特征的细节重建能力。在MVTec AD标准数据集上进行了定量和定性实验，与基准模型相比，取得了更优的图像级和像素级指标得分。

关键词: 无监督学习, 缺陷检测, 扩散模型, 反向蒸馏

Abstract: In recent years, industrial defect detection models based on unsupervised learning have achieved significant performance improvements. However, the existing defect synthesis strategies rely on external data sources, resulting in significant differences between the synthesized defects and some real defects, which seriously restricts the generalization performance of the model. Furthermore, the existing reverse distillation methods have the problem of losing feature detail information, resulting in false detection phenomena in the model. To this end, a multi-source defect synthesis strategy is first introduced. The images generated by the diffusion model and the image synthesized from the DTD dataset are more in line with the defect samples of the real defect distribution. Then, the synthetic defect samples are used to fine-tune the representational ability of the teacher network for defects. Subsequently, an anomaly masking module is introduced to address the issue of excessive generalization caused by teacher-student network isomorphism. Finally, a detail repair module is constructed to enhance the student network’s ability to reconstruct the details of the teacher’s features through cross-level feature fusion. Quantitative and qualitative experiments are conducted on the MVTec AD standard dataset. Compared with the benchmark model, the proposed method achieves better performance in terms of both image-level and pixel -level AUROC scores.

Key words: unsupervised learning, defect detection, diffusion model, reverse distillation

刘明明, 史伟峰, 范学慧, 张海燕. 融合扩散模型与知识蒸馏的无监督工业缺陷检测[J]. 计算机工程与应用, 2025, 61(23): 205-211.

LIU Mingming, SHI Weifeng, FAN Xuehui, ZHANG Haiyan. Unsupervised Industrial Defect Detection Method Based on Diffusion Model and Knowledge Distillation[J]. Computer Engineering and Applications, 2025, 61(23): 205-211.

参考文献

[1] 殷旭鹏, 赵兴强. YOLOv11-MAS: 一种高效的PCB缺陷检测算法[J]. 计算机工程与应用, 2025, 61(17): 102-111.
YIN X P, ZHAO X Q. YOLOv11-MAS: Efficient PCB defect detection algorithm[J]. Computer Engineering and Applications, 2025, 61(17): 102-111.
[2] 查健, 陈先中, 王文财, 等. 基于改进的YOLOv5s刨花板表面小目标缺陷检测算法[J]. 计算机工程与应用, 2024, 60(17): 158-166.
ZHA J, CHEN X Z, WANG W C, et al. Small defect detection algorithm of particle board surface based on improved YOLOv5s[J]. Computer Engineering and Applications, 2024, 60(17): 158-166.
[3] 李斌, 屈璐瑶, 朱新山, 等. 基于多尺度特征融合的绝缘子缺陷检测[J]. 电工技术学报, 2023, 38(1): 60-70.
LI B, QIU L Y, ZHU X S, et al. Insulator defect detection based on multi-scale feature fusion[J]. Transactions of China Electrotechnical Society, 2023, 38(1): 60-70.
[4] 李维刚, 叶欣, 赵云涛, 等. 基于改进YOLOv3算法的带钢表面缺陷检测[J]. 电子学报, 2020, 48(7): 1284-1292.
LI W G, YE X, ZHAO Y T, et al. Strip steel surface defect detection based on improved YOLOv3 algorithm[J]. Acta Electronica Sinica, 2020, 48(7): 1284-1292.
[5] 罗东亮, 蔡雨萱, 杨子豪, 等. 工业缺陷检测深度学习方法综述[J]. 中国科学: 信息科学, 2022, 52(6): 1002-1039.
LUO D L, CAI Y X, YANG Z H, et al. Survey on industrial defect detection with deep learning[J]. SCIENTIA SINICA Informationis, 2022, 52(6): 1002-1039.
[6] KINGMA D P, WELLING M. Auto?encoding variational Bayes[C]//Proceedings of the International Conference on Learning Representations, 2014: 1-9.
[7] GOODFELLOW I, POUGET A J, MIRZA M, et al. Generative adversarial nets[C]//Proceedings of the 28th International Conference on Neural Information Processing Systems, 2014: 2672-2680.
[8] HO J, JAIN A, ABBEEL P. Denoising diffusion probabilistic models[C]//Advances in Neural Information Processing Systems, 2020: 6840-6851.
[9] ZHANG H, WANG Z, ZENG D, et al. DiffusionAD: norm-guided one-step denoising diffusion for anomaly detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(8): 7140-7152.
[10] MOUSAKHAN A, BROX T, TAYYUB J. Anomaly detection with conditioned denoising diffusion models[J]. arXiv: 2305.15956, 2023.
[11] 戚银城, 武学良, 赵振兵, 等. 嵌入双注意力机制的 Faster R-CNN 航拍输电线路螺栓缺陷检测[J]. 中国图象图形学报, 2021, 26(11) : 2594- 2604.
QI Y C, WU X L, ZHAO Z B, et al. Bolt defect detection for aerial transmission lines using faster R-CNN with an embedded dual attention mechanism[J]. Journal of Image and Graphics, 2021, 26(11): 2594-2604.
[12] ROTH K, PEMULA L, ZEPEDA J, et al. Towards total recall in industrial anomaly detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2022: 14298-14308.
[13] CAO Y K, WAN Q, SHEN W M, et al. Informative knowledge distillation for image anomaly segmentation[J]. Knowledge-Based Systems, 2022, 248: 108846.
[14] DENG H Q, LI X Y. Anomaly detection via reverse distillation from one-class embedding[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2022: 9727-9736.
[15] ZHANG X, LI S Y, LI X, et al. DeSTSeg: segmentation guided denoising student-teacher for anomaly detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2023: 3914-3923.
[16] ZAVRTANIK V, KRISTAN M, SKOCAJ D. DRAEM—A discriminatively trained reconstruction embedding for surface anomaly detection[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2021: 8310-8319.
[17] CHIU L L, LAI S H. Self-supervised normalizing flows for image anomaly detection and localization[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway: IEEE, 2023: 2927-2936.
[18] JIANG J L, ZHU J L, BILAL M, et al. Masked swin transformer unet for industrial anomaly detection[J]. IEEE Transactions on Industrial Informatics, 2023, 19(2): 2200-2209.
[19] LIU Z K, ZHOU Y M, XU Y S, et al. SimpleNet: a simple network for image anomaly detection and localization[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2023: 20402-20411.
[20] TIEN T D, NGUYEN A T, TRAN N H, et al. Revisiting reverse distillation for anomaly detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2023: 24511-24520.