基于判别式增强的蒸馏学习自监督缺陷检测

doi:10.16182/j.issn1004731x.joss.24-0206

摘要/Abstract

摘要：

针对异常缺陷数据稀缺、类型未知，以及传统知识蒸馏缺陷检测方法缺乏对异常表示多样性的问题，提出一种基于判别式增强的自监督蒸馏学习方法。提出一种基于注意力的多尺度融合模块，通过放大学生网络和教师网络多尺度的特征差异来提高对异常表示的能力；设计一种由特征重加权模块和解码器组成的判别网络，通过进一步强调教师网络中的异常特征来生成更加精准的异常分数图，从而提高缺陷检测分割精度。结果表明：该方法超过了现有的知识蒸馏缺陷检测方法，验证了方法的有效性与优越性。

关键词: 缺陷检测, 知识蒸馏, 基于注意力的多尺度融合模块, 特征重加权模块, 判别网络

Abstract:

To address the issues of scarce and unknown types of abnormal defect data and the lack of diversity in anomaly representation in conventional knowledge distillation defect detection methods, a self-supervised distillation learning method based on discriminative enhancement is proposed. An attention-based multi-scale feature fusion module is proposed, which enhances the capability of anomaly representation by amplifying the multi-scale feature differences between the student network and the teacher network. A discriminative network composed of a feature reweighting module and a decoder is designed to generate more accurate anomaly score maps by further emphasizing the anomaly features in the teacher network, thereby improving the accuracy of the defect detection segmentation. Experiments show that the performance of the proposed method surpasses the existing knowledge distillation defect detection methods, verifying the effectiveness and superiority of the proposed method.

Key words: defect detection, knowledge distillation, attention-based multi-scale fusion module, feature reweighting module, discriminative network

中图分类号:

TP391.4

冯志远,陈莹 . 基于判别式增强的蒸馏学习自监督缺陷检测[J]. 系统仿真学报, 2025, 37(6): 1499-1511.

Feng Zhiyuan,Chen Ying . Self-supervised Defect Detection via Discriminative Enhancement-based Distillation Learning[J]. Journal of System Simulation, 2025, 37(6): 1499-1511.

图/表 13

图1

图2

图3

图4

图5

表1

表2

图6

表3

表4

图7

表5

图8

参考文献 25

1	Roth Karsten, Pemula L, Zepeda J, et al. Towards Total Recall in Industrial Anomaly Detection[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 14298-14308.
2	于豪, 蒋锦霞, 赖晓翰, 等. 基于自适应感受野的电力设备表面缺陷检测方法[J]. 系统仿真学报, 2023, 35(7): 1572-1580.
	Yu Hao, Jiang Jinxia, Lai Xiaohan, et al. Surface Defect Detection of Power Equipment Using Adaptive Receptive Field Network[J]. Journal of System Simulation, 2023, 35(7): 1572-1580.
3	Zavrtanik Vitjan, Kristan Matej, Skočaj Danijel. DRÆM-a Discriminatively Trained Reconstruction Embedding for Surface Anomaly Detection[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 8310-8319.
4	Schlegl T, Seeböck Philipp, Waldstein S M, et al. f-AnoGAN: Fast Unsupervised Anomaly Detection with Generative Adversarial Networks[J]. Medical Image Analysis, 2019, 54: 30-44.
5	Watanabe Haruna, Ren Togo, Ogawa Takahiro, et al. Bone Metastatic Tumor Detection Based on AnoGAN Using CT Images[C]//2019 IEEE 1st Global Conference on Life Sciences and Technologies (LifeTech). Piscataway: IEEE, 2019: 235-236.
6	王可, 张辉, 曹意宏, 等. 面向医药生产的智能机器人及其关键技术研究综述[J]. 计算机集成制造系统, 2022, 28(7): 1981-1995.
	Wang Ke, Zhang Hui, Cao Yihong, et al. Intelligent Robots and Key Technologies for Pharmaceutical Production[J]. Computer Integrated Manufacturing Systems, 2022, 28(7): 1981-1995.
7	Liu Y, Li C L, Póczos B. Classifier Two Sample Test for Video Anomaly Detections[C]//BMVC. Newcastle: BMVA, 2018: 71.
8	Pang Guansong, Yan Cheng, Shen Chunhua, et al. Self-trained Deep Ordinal Regression for End-to-end Video Anomaly Detection[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 12170-12179.
9	Bogdoll Daniel, Nitsche Maximilian, Marius Zöllner J. Anomaly Detection in Autonomous Driving: A Survey[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Piscataway: IEEE, 2022: 4487-4498.
10	Vojir Tomas, Šipka Tomáš, Aljundi Rahaf, et al. Road Anomaly Detection by Partial Image Reconstruction with Segmentation Coupling[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 15631-15640.
11	Defard Thomas, Setkov Aleksandr, Loesch Angelique, et al. PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization[C]//Pattern Recognition. ICPR International Workshops and Challenges. Cham: Springer International Publishing, 2021: 475-489.
12	Cohen Niv, Hoshen Yedid. Sub-image Anomaly Detection with Deep Pyramid Correspondences[EB/OL]. (2021-02-03) [2023-04-20]. .
13	Kingma Diederik P, Welling Max. Auto-encoding Variational Bayes[EB/OL]. (2022-12-10) [2023-04-20]. .
14	Goodfellow I, Pouget-Abadie Jean, Mirza Mehdi, et al. Generative Adversarial Networks[J]. Communications of the ACM, 2020, 63(11): 139-144.
15	戴俊, 王俊, 朱忠奎, 等. 基于生成对抗网络和自动编码器的机械系统异常检测[J]. 仪器仪表学报, 2019, 40(9): 16-26.
	Dai Jun, Wang Jun, Zhu Zhongkui, et al. Anomaly Detection of Mechanical Systems Based on Generative Adversarial Network and Auto-encoder[J]. Chinese Journal of Scientific Instrument, 2019, 40(9): 16-26.
16	Schlegl T, Seeböck Philipp, Waldstein S M, et al. Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery[C]//Information Processing in Medical Imaging. Cham: Springer International Publishing, 2017: 146-157.
17	Li Chunliang, Sohn K, Yoon J, et al. CutPaste: Self-supervised Learning for Anomaly Detection and Localization[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 9659-9669.
18	Tsai M C, Wang Shengde. Self-supervised Image Anomaly Detection and Localization with Synthetic Anomalies[C]//2023 10th International Conference on Internet of Things: Systems, Management and Security (IOTSMS). Piscataway: IEEE, 2023: 90-95.
19	Bergmann Paul, Fauser Michael, Sattlegger David, et al. Uninformed Students: Student-teacher Anomaly Detection with Discriminative Latent Embeddings[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 4182-4191.
20	Salehi Mohammadreza, Sadjadi Niousha, Baselizadeh Soroosh, et al. Multiresolution Knowledge Distillation for Anomaly Detection[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 14897-14907.
21	Wang Guodong, Han Shumin, Ding Errui, et al. Student-teacher Feature Pyramid Matching for Anomaly Detection[EB/OL]. (2021-10-28) [2023-04-20]. .
22	Bergmann Paul, Fauser Michael, Sattlegger David, et al. MVTec AD-a Comprehensive Real-world Dataset for Unsupervised Anomaly Detection[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 9584-9592.
23	Deng Hanqiu, Li Xingyu. Anomaly Detection via Reverse Distillation from One-class Embedding[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 9727-9736.
24	Luo Jiaxiang, Zhang Jianzhao. A Method for Image Anomaly Detection Based on Distillation and Reconstruction[J]. Sensors, 2023, 23(22): 9281.
25	Liu Tongkun, Li Bing, Zhao Zhuo, et al. Reconstruction from Edge Image Combined with Color and Gradient Difference for Industrial Surface Anomaly Detection[EB/OL]. (2022-10-26) [2023-04-20]. .

	类别	US	MKD	PaDiM	EdgRec	DRAEM	RD	DRAD	STPM	本文
纹理类	地毯	91.6	79.3	99.8	97.4	97.0	98.9	99.3	—	99.8
	网格	81.0	78.0	96.7	99.7	99.9	100	99.9	—	100
	皮革	88.2	95.1	100	100	100	100	99.9	—	100
	瓷砖	99.1	91.6	98.1	100	99.6	99.3	99.2	—	100
	木材	97.7	94.3	99.2	94.0	99.1	99.2	99.6	—	99.7
对象类	瓶底	99.0	99.4	99.9	100	99.2	100	100	—	100
	电缆	86.2	89.2	92.7	97.9	91.8	95.0	98.6	—	98.2
	胶囊	86.1	80.5	91.3	95.5	98.5	96.3	98.1	—	97.6
	榛子	93.1	98.4	92.0	98.4	100	99.9	100	—	100
	齿轮	82.0	73.6	98.7	97.3	98.7	100	99.6	—	100
	药片	87.9	82.7	93.3	99.0	98.9	96.6	96.8	—	99.1
	螺钉	54.9	83.3	85.8	89.9	93.9	97.0	97.6	—	98.4
	牙刷	95.3	92.2	96.1	100	100	99.5	97.2	—	99.4
	晶体管	81.8	85.6	97.4	99.8	93.1	96.7	98.9	—	98.2
	拉链	91.9	93.2	90.3	98.3	100	98.5	98.8	—	98.1
	平均值	87.7	87.8	95.5	97.8	98.0	98.5	98.9	95.5	99.2

类别		US	MKD	EdgRec	DRAEM	PaDiM	DRAD	RD	STPM	本文
纹理类	地毯	—/87.9	95.6/—	99.4/96.9	95.5/—	99.1/96.2	98.7/—	98.9/97.0	98.8/95.8	99.5/98.0
	网格	—/95.2	91.8/—	99.2/97.0	99.7/—	97.3/94.6	98.9/—	99.3/97.6	99.0/96.6	99.2/96.9
	皮革	—/94.5	98.1/—	99.7/98.8	98.6/—	99.2/97.8	99.1/—	99.4/99.1	99.3/98.0	99.8/99.4
	瓷砖	—/94.6	82.8/—	98.6/96.3	99.2/—	94.1/86.0	95.2/—	95.6/90.6	97.4/92.1	98.4/96.1
	木材	—/91.1	84.8/—	91.4/77.5	96.4/—	94.9/91.1	94.9/—	95.3/90.9	97.2/93.6	97.2/94.7
对象类	瓶底	—/93.1	96.3/—	98.3/94.3	99.1/—	98.3/94.8	98.6/—	98.7/96.6	98.8/95.1	99.3/97.4
	电缆	—/81.8	82.4/—	97.7/88.7	94.7/—	96.7/88.8	96.9/—	97.4/91.0	95.5/87.7	96.6/90.5
	胶囊	—/96.8	95.9/—	95.2/82.2	94.3/—	98.5/93.5	98.9/—	98.7/95.8	98.3/92.2	98.7/95.8
	榛子	—/96.5	94.6/—	99.4/95.4	99.7/—	98.2/92.6	98.9/—	98.9/95.5	98.5/94.3	99.4/96.5
	齿轮	—/94.2	86.4/—	98.0/91.2	99.5/—	97.2/85.6	98.5/—	97.3/92.3	97.6/94.5	97.6/92.8
	药片	—/96.1	89.6/—	98.7/96.1	97.6/—	95.7/82.7	99.1/—	98.2/96.4	97.8/96.5	99.3/97.9
	螺钉	—/94.2	96.0/—	97.7/89.3	97.6/—	98.5/94.4	99.2/—	99.6/98.2	98.3/93.0	99.6/98.2
	牙刷	—/93.3	96.1/—	99.2/94.9	98.1/—	98.8/93.1	98.8/—	99.1/94.5	98.9/92.2	99.0/93.7
	晶体管	—/66.6	76.5/—	94.3/87.7	90.9/—	97.5/84.5	91.4/—	92.5/78.0	82.5/69.5	92.5/73.5
	拉链	—/95.1	93.9/—	98.7/96.0	98.8/—	98.5/95.9	98.5/—	98.2/95.4	98.5/95.2	99.0/95.9
	平均值	—/91.4	90.7/—	97.7/92.1	97.3/—	97.5/92.1	97.7/—	97.8/93.9	97.0/92.1	98.3/94.5

新指标	ResNet18	ResNet50	WideResNet50
DET-AUROC	98.1	98.6	99.2
SET-AUROC	97.6	98.0	98.3
PRO-AUROC	93.1	93.9	94.5

师生网络	判别网络	AMFF	DET-AUROC	SET-AUROC	PRO-AUROC
√			95.0	97.1	92.3
√	√		98.7	98.1	94.1
√		√	98.2	97.7	93.9
√	√	√	99.2	98.3	94.5

方法	DET-AUROC	SEG-AUROC	PRO-AUROC
师生网络	95.0	97.1	92.3
MFF+师生网络	97.4	97.3	92.8
AMFF+师生网络	98.2	97.7	93.9