基于多尺度标准化流模型的布匹缺陷检测

doi:10.3901/JME.2025.10.241

摘要/Abstract

摘要： 基于深度学习的布匹缺陷检测算法依赖大量带标签的缺陷样本，但在实际生产过程中难以大量获取且标注烦琐。针对这一问题提出一种基于多尺度标准化流模型的布匹缺陷检测算法。该方法首先通过加载预训练权重的特征提取网络获取正常样本的特征，然后通过训练标准化流模型，建立正常样本特征分布与标准高斯分布之间的映射关系，以此评估待测图像特征经转化后在标准高斯分布中的概率密度来识别和定位缺陷。为减轻深层特征低分辨率引起的纹理细节损失，提出一种融合策略融合不同尺度下变换后的特征分布。检测时计算待测布匹图像特征的概率密度来评估每个像素的异常程度，并根据设定的阈值划分缺陷区域。试验表明：所提出的方法在不需要缺陷样本的情况下即可完成布匹缺陷检测，且检测精度可达像素级。在不同纹理背景与不同缺陷类型的数据上表现优异，图像级AUC高达100%，像素级AUC达98.5%，显示了良好的泛化能力。

关键词: 布匹缺陷检测, 正常样本, 标准化流, 多尺度, 深度学习

Abstract: Deep learning-based fabric defect detection methods rely heavily on a substantial collection of labeled defect samples, which can be challenging to obtain and annotate during actual manufacturing processes. To address this issue, a fabric defect detection method based on multiscale normalizing flow model is proposed. The method initially extracts features from defect-free samples via a feature extraction network with pre-trained weights. Subsequently, it establishes a mapping relationship between the feature distribution of defect-free samples and the standard Gaussian distribution through training the normalizing flow model. This mapping enables the assessment of the probability density of a test image’s features within the standard Gaussian distribution to identify and localize defects. In order to mitigate the loss of texture details caused by the low resolution of deep features, a fusion strategy is proposed to fuse the feature distributions at different scales. During the detection process, the probability density of the features for the fabric image is calculated to evaluate the anomaly degree of each pixel, and defect regions are delineated based on a predefined threshold. Experimental results demonstrate that the proposed multiscale flow model-based detection approach can accomplish fabric defect detection without requiring defect samples, achieving pixel-wise accuracy. It exhibits outstanding performance on fabric defect dataset with varying texture backgrounds and defects, achieving a 100% image-wise AUC metric and a 98.5% pixel-wise AUC metric, showcasing robust generalization capabilities.

Key words: fabric defect detection, defect-free samples, normalizing flow, multiscale, deep learning

中图分类号:

TS101

余智祺, 徐洋, 王元飞, 盛晓伟. 基于多尺度标准化流模型的布匹缺陷检测[J]. 机械工程学报, 2025, 61(10): 241-249.

YU Zhiqi, XU Yang, WANG Yuanfei, SHENG Xiaowei. Fabric Defect Detection Based on Multiscale Normalizing Flow Model[J]. Journal of Mechanical Engineering, 2025, 61(10): 241-249.

参考文献

[1] HANBAY K，TALU M F，OZGUVEN O F. Fabric defect detection systems and methods：A systematic literature review[J]. Optik，2016，127(24)：11960-11973.
[2] CZIMMERMANN T，CIUTI G，MILAZZO M，et al. Visual-based defect detection and classification approaches for industrial applications：A survey[J]. Sensors，2020，20(5)：1459.
[3] ZHENG X Q，ZHENG S，KONG Y G，et al. Recent advances in surface defect inspection of industrial products using deep learning techniques[J]. International Journal of Advanced Manufacturing Technology，2021，113(1-2)：35-58.
[4] CELIK H I，DULGER L U，TOPALBEKIROGLU M. Development of a machine vision system：Real-time fabric defect detection and classification with neural networks[J]. Journal of the Textile Institute，2014，105(6)：575-585.
[5] JING J F，ZHANG H H，WANG J，et al. Fabric defect detection using Gabor filters and defect classification based on LBP and Tamura method[J]. Journal of the Textile Institute，2013，104(1)：18-27.
[6] ZHU D D，PAN R R，GAO W D，et al. Yarn-dyed fabric defect detection based on autocorrelation function and glcm[J]. Autex Research Journal，2015，15(3)：226-232.
[7] HE K，ZHANG X，REN S，et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition， New York：IEEE，2016：770-778.
[8] SZEGEDY C，VANHOUCKE V，IOFFE S，et al. Rethinking the inception architecture for computer vision [C/CD]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition，F. New York：IEEE，2016.
[9] JUN X，WANG J A，ZHOU J，et al. Fabric defect detection based on a deep convolutional neural network using a two-stage strategy[J]. Textile Research Journal，2021，91(1-2)：130-142.
[10] 罗维平，徐洋，陈永恒，等. 基于迁移学习和改进ResNet50网络的织物疵点检测算法[J]. 毛纺科技，2021，49(2)：71-78.LUO Weiping，XU Yang，CHEN Yongheng，et al. Fabric defect detection algorithm based on migration learning and improved ResNet50 network[J]. Wool Textile Journal，2021，49(2)：71-78.
[11] 郭同建，徐洋，陈慧敏，等. 基于深度学习的机织印花布疵点实时检测方法研究[J]. 东华大学学报，2023，49(1)：33-38.GUO Tongjian，XU Yang，CHEN Huimin，et al. Research on real-time detection method of woven printed fabric defects based on deep learning[J]. Journal of Donghua University，2023，49(1)：33-38.
[12] 赵树煊，张洁，汪俊亮，等. 基于两阶段深度迁移学习的面料疵点检测算法[J]. 机械工程学报，2021，57(17)：86-97.ZHAO Shuxuan，ZHANG Jie，WANG Junliang，et al. Fabric defect detection algorithm based on two-stage deep transfer learning[J]. Journal of Mechanical Engineering，2021，57(17)：86-97.
[13] REN S，HE K，GIRSHICK R，et al. Faster R-CNN：Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence，2017，39(6)：1137-1149.
[14] JING J F，WANG Z，RATSCH M，et al. Mobile-Unet：An efficient convolutional neural network for fabric defect detection[J]. Textile Research Journal，2022，92(1-2)：30-42.
[15] TAO X，ZHANG D P，MA W Z，et al. Automatic metallic surface defect detection and recognition with convolutional neural networks[J]. Applied Sciences-Basel，2018，8(9)：1575.
[16] RONNEBERGER O，FISCHER P，BROX T. U-net：Convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention (MICCAI)，October 5-9，2015，Munich，Germany. 2015：234-241.
[17] TAO X，GONG X，ZHANG X，et al. Deep learning for unsupervised anomaly localization in industrial images：A survey[J]. IEEE Transactions on Instrumentation and Measurement，2022，71：1-21.
[18] 王孟涛，李岳阳，杜帅，等. 基于灰度梯度共生矩阵和SVDD的织物疵点检测[J]. 丝绸，2018，55(12)：50-56.WANG Mengtao，LI Yueyang，DU Shuai，et al. Fabric defect detection based on gray-level gradient co-occurrence matrix and SVDD[J]. Journal of Silk，2018，55(12)：50-56.
[19] AIGER D，TALBOT H. The phase only transform for unsupervised surface defect detection[C]//2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition，New York：IEEE，2010：295-302.
[20] 余文勇，张阳，姚海明，等. 基于轻量化重构网络的表面缺陷视觉检测[J]. 自动化学报，2022，48(9)：2175-2186.YU Wenyong，ZHANG Yang，YAO Haiming，et al. Visual inspection of surface defects based on lightweight reconstruction network[J]. Acta Automatica Sinica，2022，48(9)：2175-2186.
[21] 张宏伟，刘舒婷，陆帅，等. 基于多尺度卷积自编码器的色织面料缺陷检测[J]. 纺织高校基础科学学报，2021，34(2)：45-51.ZHANG Hongwei，LIU Shuting，LU Shuai，et al. Defect detection of yarn-dyed fabric based on multi-scale convolutional encoder[J]. Basic Sciences Journal of Textile Universities，2021，34(2)：45-51.
[22] DINH L，KRUEGER D，BENGIO Y. Nice：Non-linear independent components estimation[J]. arXiv Preprint arXiv：1410.8516，2014.
[23] YU J，ZHENG Y，WANG X，et al. Fastflow：Unsupervised anomaly detection and localization via 2D normalizing flows[J]. arXiv Preprint arXiv：211107677，2021.
[24] MA X，ZHOU C，LI X，et al. FlowSeq：Non- autoregressive conditional sequence generation with generative flow[C]//Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019：4282-4292.
[25] TSANG C S C，NGAN H Y T，PANG G K H. Fabric inspection based on the Elo rating method[J]. Pattern Recognition，2016，51：378-394.
[26] SILVESTRE-BLANES J，ALBERO-ALBERO T，MIRALLES I，et al. A public fabric database for defect detection methods and results[J]. Autex Research Journal，2019，19(4)：363-374.
[27] BERGMANN P，FAUSER M，SATTLEGGER D，et al. MVTec AD—A comprehensive real-world dataset for unsupervised anomaly detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2019：9592-9600.
[28] BERGMANN P，LÖWE S，FAUSER M，et al. Improving unsupervised defect segmentation by applying structural similarity to autoencoders[J]. arXiv Preprint arXiv：1807.02011，2018.
[29] SANDLER M，HOWARD A，ZHU M，et al. Mobilenetv2：Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition，New York：IEEE, 2018：4510-4520.
[30] SIMONYAN K，ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[J]. arXiv Preprint arXiv：1409.1556，2014.
[31] AKCAY S，ATAPOUR-ABARGHOUEI A，BRECKON T P. Ganomaly：Semi-supervised anomaly detection via adversarial training[C]//14th Asian Conference on Computer Vision，December 2-6，2018，Perth，Australia. 2019：622-637.
[32] 张宏伟，汤文博，李鹏飞，等. 基于去噪卷积自编码器的色织衬衫裁片缺陷检测[J]. 纺织高校基础科学学报，2019，32(2)：119-125，132.ZHANG Hongwei，TANG Wenbo，LI Pengfei，et al. Defect detection and location of yarn-dyed shirt piece based on denoising convolutional autoencoder[J]. Basic Sciences Journal of Textile Universities，2019，32(2)：119-125，132.
[33] LIU Z，HUO Z，LI C，et al. DLSE-Net：A robust weakly supervised network for fabric defect detection[J]. Displays，2021，68：102008.
[34] WANG J，XU G，LI C，et al. SDDet：An enhanced encoder-decoder network with hierarchical supervision for surface defect detection[J]. IEEE Sensors Journal，2023，23(3)：2651-2662.