在织物生产中表面缺陷检测是质量控制的关键手段。无监督深度学习重建方法可有效解决织物缺陷样本稀缺、标注成本高昂以及先验知识匮乏等问题。然而目前无监督重建模型在将缺陷图像高质量地重建为无缺陷图像的任务上仍然面临挑战，重建结果存在图像模糊、缺陷残留和纹理不一致等问题。针对以上问题，主要工作如下：

（1）提出了一种面向时间步长自适应扩散模型的织物表面图像重建方法，解决了现有模型的重建结果存在图像模糊和缺陷残留的问题。首先构建使用单纯形噪声控制扩散过程的去噪扩散概率模型(Simplex noise-Denoising Diffusion Probabilistic Model, SN-DDPM)，通过递归式地优化后验潜在向量分布，使模型逐渐逼近无缺陷样本表面特征的概率分布；其次利用时间步长自适应模块动态调整最佳时间步长，使模型能够灵活适应不同数据分布；最后通过与七种无监督缺陷检测方法对比，SN-DDPM获得最佳重建结果，在结构相似度指标和峰值信噪比指标的平均值上分别提高了0.01和0.16dB。

（2）提出了一种基于频率调谐显著性检测的缺陷分割方法，解决了已有检测方法精度低，存在细小目标难以检出的问题。首先使用SN-DDPM将缺陷图像重建为无缺陷图像；其次通过图像差分、频率调谐显著性检测以及阈值二值化等图像处理手段分割缺陷；最后进行实验并将所提方法与当前应用的七种无监督缺陷检测方法进行对比，实验结果表明：所提方法在F1和IoU的平均值上分别提高了5.42%和7.61%。

（3）提出了一种面向掩码修复模型的织物表面缺陷检测方法，解决了重建结果存在纹理不一致的问题。首先利用以Transformer为架构的掩码修复模型(Mask repair model, MRM)捕获全图的特征关系，引入结构相似度作为约束有效关注织物的纹理特性；然后以SN-DDPM的重建结果作为引导获取目标掩码，使用掩码的方式修复缺陷区域；最后利用基于频率调谐显著性检测的缺陷分割方法分割缺陷。实验结果表明：相比于SN-DDPM，该方法在结构相似度指标和峰值信噪比指标的平均值上分别提高了0.05和3.35dB，在F1和IoU的平均值上分别提高了4.52%和2.35%。

（4）设计并实现了织物表面缺陷无监督检测系统，验证了所提表面缺陷检测方法的有效性。首先对MRM知识蒸馏并优化掩码策略，参数量减少91.30%，平均单张图像检测时间缩短至0.08s；其次将知识蒸馏后的掩码修复模型(Knowledge Distillation- Progressive Mask Repair Model, KD-PMRM)部署至硬件装置，并搭建缺陷管理软件；最后通过实验测试，结果表明：该系统可有效分割缺陷并将缺陷图像上传至管理软件，系统响应平均时间41ms。

综上所述，提出的织物表面缺陷无监督检测方法可以检测未定义缺陷，具有较高的检测精度和较强的泛化能力，能够满足织物表面缺陷检测场景的需求。

Surface defect detection is a key means of quality control in fabric production. Unsupervised deep learning reconstruction methods can effectively solve problems such as scarcity of fabric defect samples, high labeling costs, and lack of prior knowledge. However, currently unsupervised reconstruction models still face challenges in reconstructing defect images into defect free images with high quality, resulting in image blurring, residual defects, and inconsistent textures in the reconstruction results. In response to the above issues, the main tasks are as follows:

（1）A fabric surface image reconstruction method oriented to the time-step adaptive diffusion model is proposed, which solves the problems of image blurring and defect residue in the reconstruction results of existing models. Firstly, Simplex Noise -- Denoising Diffusion Probabilistic Model (SN-DDPM) is constructed to gradually approach to the probability distribution of surface features of the defect-free samples through multiple iterative diffusions. Secondly, the timestep adaptive module is utilized to dynamically adjust the optimal timestep, enabling the model to flexibly adapt to different data distributions. Finally, by comparing with seven unsupervised defect detection methods, SN-DDPM obtains the best reconstruction results, which improves the average values of structural similarity index and peak signal-to-noise ratio index by 0.01 and 0.16 dB, respectively.

（2）A defect segmentation method based on frequency-tuned saliency detection is proposed to solve the problem of low accuracy of existing detection methods and the existence of fine targets that are difficult to detect. Firstly, the defect image is reconstructed into a defect-free image using SN-DDPM. Secondly, the defects are segmented by image processing means such as image differencing, frequency-tuned saliency detection, and threshold binarization. Finally, experiments are carried out and an exhaustive comparison is made between the proposed method and the seven currently widely used unsupervised defect detection methods, and the results of the experiments show that: the proposed method improves the average value of F1 and IoU by 5.42% and 7.61%, respectively.

（3）A fabric surface defect detection method oriented to the mask repair model is proposed to solve the problem of texture inconsistency in the reconstruction results. Firstly, we utilize the Mask repair model (MRM) with Transformer as the architecture to capture the feature relationship of the whole map, and introduce the structural similarity as a constraint to pay attention to the texture characteristics of the fabric effectively. Then we use the reconstruction result of SN-DDPM as a guide to obtain the target mask, and repair the defective region using the mask. Finally, we utilize the defect segmentation method based on frequency tuning and significance detection. Finally, the defect segmentation method based on frequency tuning and significance detection is utilized to segment the defects. The experimental results show that compared with SN-DDPM, this method improves the structural similarity index and peak signal-to-noise ratio index by 0.05 and 3.35 dB on average, and the F1 value and IoU value by 4.52% and 2.35% on average, respectively.

（4）An unsupervised inspection system for fabric surface defects is designed and implemented, and the effectiveness of the proposed surface defect detection method is verified. Firstly, the knowledge distillation of MRM and the optimization of masking strategy reduce the amount of parameters by 91.30%, and the average single image detection time is shortened to 0.08s. Secondly, the knowledge distillation- Progressive Mask Repair Model (KD-PMRM) is deployed to the hardware device, and the defect management software is built. Finally, the experimental test shows that the system can effectively segment the defects and upload the defects images to the fabric surface unsupervised detection system. and build the defect management software. Through the experimental test, the results show that the system can effectively segment the defects and upload the defect images to the management software, and the average system response time is 41ms.

In summary, the proposed unsupervised method for detecting defects on fabric surfaces demonstrates a notable capacity to identify uncharacterized anomalies with high precision and robust generalization, thereby fulfilling the requisite criteria for fabric surface defect detection applications.

[1] Li C, Li J, Li Y, et al. Fabric defect detection in textile manufacturing: a survey of the state of the art[J]. Security and Communication Networks, 2021, 2021: 1-13.

[2] Bhatt P M, Malhan R K, Rajendran P, et al. Image-based surface defect detection using deep learning: A review[J]. Journal of Computing and Information Science in Engineering, 2021, 21(4): 1-15.

[3] Wong W K, Jiang J L. Computer vision techniques for detecting fabric defects[M]//Applications of computer vision in fashion and textiles. Woodhead Publishing, 2018: 47-60.

[4] 邵家鑫, 都东, 石涵, 等. 基于厚壁工件X射线实时成像的焊缝缺陷自动检测[J]. 清华大学学报(自然科学版), 2013, 53(02): 150-154.

[5] Shibata T, Hashizume H, Kitajima S, et al. Experimental study on NDT method using electromagnetic waves[J]. Journal of materials processing technology, 2005, 161(1-2): 348-352.

[6] 杨海瑞, 谈志杰, 喻虹, 等. 基于单幅散斑的X射线强度关联缺陷检测方法[J]. 光学学报, 2023, 43(10): 301-308.

[7] Zhang J, Drinkwater B W, Wilcox P D, et al. Defect detection using ultrasonic arrays: The multi-mode total focusing method[J]. NDT & e International, 2010, 43(2): 123-133.

[8] D’orazio T, Leo M, Distante A, et al. Automatic ultrasonic inspection for internal defect detection in composite materials[J]. NDT & e International, 2008, 41(2): 145-154.

[9] 张翠平, 许小婷. 牙齿无损探伤有限元分析[J]. 医用生物力学, 2023, 38(05):899-904.

[10] 武麒阳, 杜红秀, 王聪聪. 自密实混凝土钢管柱超声检测内部缺陷研究[J]. 太原理工大学学报, 2023, 54(06): 1144-1152.

[11] 夏庆国. 光电测量的新能源汽车车轮踏面缺陷检测方法[J]. 激光杂志, 2021, 42(05): 161-165.

[12] 张静, 叶玉堂, 谢煜, 等. 金属圆柱工件缺陷的光电检测[J]. 光学精密工程, 2014, 22(07): 1871-1876.

[13] 王娜, 刘立拓, 宋晓娇, 等. 基于过焦扫描光学显微镜的光学元件亚表面缺陷检测方法[J]. 光学学报, 2023, 43(21): 120-127.

[14] Rashidi A, Olfatbakhsh T, Crawford B, et al. A review of current challenges and case study toward optimizing micro-computed X-ray tomography of carbon fabric composites[J]. Materials, 2020, 13(16): 1-30.

[15] Rasheed A, Zafar B, Rasheed A, et al. Fabric defect detection using computer vision techniques: a comprehensive review[J]. Mathematical Problems in Engineering, 2020, 2020: 1-24.

[16] 朱刘盅, 李竞择, 吴磊, 等. 基于Gabor滤波器的光纤缺陷检测方法[J]. 南京理工大学学报, 2022, 46(01): 119-126.

[17] 仰梓淮, 黄海鸿, 刘贺, 等. 基于颜色直方图的电路板表面缺陷检测[J]. 计算机集成制造系统, 2024, 1(09): 1-15.

[18] Hao Z, Wang Z, Bai D, et al. Towards the steel plate defect detection: Multidimensional feature information extraction and fusion[J]. Concurrency and Computation: Practice and Experience, 2021, 33(21): 1-8.

[19] Gao Y, Li X, Wang X V, et al. A review on recent advances in vision-based defect recognition towards industrial intelligence[J]. Journal of Manufacturing Systems, 2022, 62: 753-766.

[20] Chen M, Yu L, Zhi C, et al. Improved faster R-CNN for fabric defect detection based on Gabor filter with Genetic Algorithm optimization[J]. Computers in Industry, 2022, 134: 1-18.

[21] Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[J]. Advances in neural information processing systems, 2015, 28:1-9.

[22] Suryarasmi A, Chang C C, Akhmalia R, et al. FN-Net: A lightweight CNN-based architecture for fabric defect detection with adaptive threshold-based class determination[J]. Displays, 2022, 73: 1-17.

[23] Hu B, Wang J. Detection of PCB surface defects with improved faster-RCNN and feature pyramid network[J]. Ieee Access, 2020, 8: 108335-108345.

[24] Yue X, Wang Q, He L, et al. Research on tiny target detection technology of fabric defects based on improved Yolo[J]. Applied Sciences, 2022, 12(13): 1-16.

[25] 蒋尚俊, 易辉, 李红涛, 等. 基于迁移学习与ResNet的太阳电池缺陷检测方法[J]. 太阳能学报, 2023, 44(07): 116-121.

[26] Xu Y, Zhang K, Wang L. Metal surface defect detection using modified YOLO[J]. Algorithms, 2021, 14(9): 1-14.

[27] Lei J, Hu X, Wang Y, et al. PyramidFlow: High-Resolution Defect Contrastive Localization using Pyramid Normalizing Flow[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 14143-14152.

[28] Li Y, Zhao W, Pan J. Deformable patterned fabric defect detection with fisher criterion-based deep learning[J]. IEEE Transactions on Automation Science and Engineering, 2016, 14(2): 1256-1264.

[29] Yan Y, Wang D, Zhou G, et al. Unsupervised anomaly segmentation via multilevel image reconstruction and adaptive attention-level transition[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1-12.

[30] Zhang H, Liu S, Tan Q, et al. Colour‐patterned fabric defect detection based on an unsupervised multi‐scale U‐shaped denoising convolutional autoencoder model[J]. Coloration Technology, 2022, 138(5): 522-537.

[31] Li X, Zheng Y, Chen B, et al. Dual Attention-Based Industrial Surface Defect Detection with Consistency Loss[J]. Sensors, 2022, 22(14): 1-18.

[32] Zhang H, Qiao G, Lu S, et al. Attention-based Feature Fusion Generative Adversarial Network for yarn-dyed fabric defect detection[J]. Textile Research Journal, 2023, 93(5-6): 1178-1195.

[33] Zhang L, Dai Y, Fan F, et al. Anomaly Detection of GAN Industrial Image Based on Attention Feature Fusion[J]. Sensors, 2022, 23(1): 1-16.

[34] Murase H, Fukumizu K. Algan: Anomaly detection by generating pseudo anomalous data via latent variables[J]. IEEE Access, 2022, 10: 44259-44270.

[35] Wei C, Liang J, Liu H, et al. Multi-stage unsupervised fabric defect detection based on DCGAN[J]. The Visual Computer, 2023, 39(12): 6655-6671.

[36] Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11): 139-144.

[37] Zhai J, Zhang S, Chen J, et al. Autoencoder and its various variants[C]//2018 IEEE international conference on systems, man, and cybernetics (SMC). IEEE, 2018: 415-419.

[38] Kahraman Y, Durmuşoğlu A. Deep learning-based fabric defect detection: A review[J]. Textile Research Journal, 2023, 93(5-6): 1485-1503.

[39] Dong G, Liao G, Liu H, et al. A review of the autoencoder and its variants: A comparative perspective from target recognition in synthetic-aperture radar images[J]. IEEE Geoscience and Remote Sensing Magazine, 2018, 6(3): 44-68.

[40] Yang L, Zhang Z, Song Y, et al. Diffusion models: A comprehensive survey of methods and applications[J]. ACM Computing Surveys, 2023, 56(4): 1-39.

[41] Croitoru F A, Hondru V, Ionescu R T, et al. Diffusion models in vision: A survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 1-25.

[42] Vincent P, Larochelle H, Bengio Y, et al. Extracting and composing robust features with denoising autoencoders[C]//Proceedings of the 25th international conference on Machine learning. 2008: 1096-1103.

[43] Bengio Y, Courville A, Vincent P. Representation learning: A review and new perspectives[J]. IEEE transactions on pattern analysis and machine intelligence, 2013, 35(8): 1798-1828.

[44] Perlin K. Improving noise[C]//Proceedings of the 29th annual conference on Computer graphics and interactive techniques. 2002: 681-682.

[45] Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models[J]. Advances in neural information processing systems, 2020, 33: 6840-6851.

[46] Graham M S, Pinaya W H L, Tudosiu P D, et al. Denoising diffusion models for out-of-distribution detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 2947-2956.

[47] Wang Z, Bovik A C, Sheikh H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE transactions on image processing, 2004, 13(4): 600-612.

[48] Zhang H. Yarn-dyed Fabric Image Dataset Version 1, 2021. http://github.com/ZHW-AI/YDFID-1

[49] Huynh-Thu Q, Ghanbari M. Scope of validity of PSNR in image/video quality assessment[J]. Electronics letters, 2008, 44(13): 800-801.

[50] Zhang H, Tang W, Zhang L, et al. Defect detection of yarn-dyed shirts based on denoising convolutional self-encoder[C]//2019 IEEE 8th Data Driven Control and Learning Systems Conference (DDCLS). IEEE, 2019: 1263-1268.

[51] Hu G, Huang J, Wang Q, et al. Unsupervised fabric defect detection based on a deep convolutional generative adversarial network[J]. Textile Research Journal, 2020, 90(3-4): 247-270.

[52] Bansal A, Ma S, Ramanan D, et al. Recycle-gan: Unsupervised video retargeting[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 119-135.

[53] Mei S, Yang H, Yin Z. An unsupervised-learning-based approach for automated defect inspection on textured surfaces[J]. IEEE Transactions on Instrumentation and Measurement, 2018, 67(6): 1266-1277.

[54] Zhang H, Tan Q, Lu S. Yarn-dyed shirt piece defect detection based on an unsupervised reconstruction model of the U-shaped denoising convolutional auto-encoder[J]. Journal of Xidian University, 2021, 48(3): 123-130.

[55] Wei W, Deng D, Zeng L, et al. Real-time implementation of fabric defect detection based on variational automatic encoder with structure similarity[J]. Journal of Real-Time Image Processing, 2021, 18: 807-823.

[56] Cross G R, Jain A K. Markov random field texture models[J]. IEEE Transactions on pattern analysis and machine intelligence, 1983 (1): 25-39.

[57] Fogel I, Sagi D. Gabor filters as texture discriminator[J]. Biological cybernetics, 1989, 61(2): 103-113.

[58] Chen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 801-818.

[59] Glenn Jocher, Ayush Chaurasia, Jing Qiu. Yolo by ultralytics Version 8.0.0, 2023. Available online: https://github.com/ultralytics/ultralytics.

[60] He K, Chen X, Xie S, et al. Masked autoencoders are scalable vision learners[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 16000-16009.

[61] Parmar N, Vaswani A, Uszkoreit J, et al. Image transformer[C]//International conference on machine learning. PMLR, 2018: 4055-4064.

[62] Wang Z, Bovik A C. A universal image quality index[J]. IEEE signal processing letters, 2002, 9(3): 81-84.

[63] Gou J, Yu B, Maybank S J, et al. Knowledge distillation: A survey[J]. International Journal of Computer Vision, 2021, 129: 1789-1819.

[64] Bai Y, Wang Z, Xiao J, et al. Masked autoencoders enable efficient knowledge distillers[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 24256-24265.

[65] 王睿, 齐建鹏, 陈亮, 等. 面向边缘智能的协同推理综述[J]. 计算机研究与发展, 2023, 60(02): 398-414.