In the steel manufacturing process, surface repetitive defects are easily caused by equipment wear or process fluctuations, and intelligent detection and clustering methods are urgently required to achieve efficient recognition and fault source tracing. However, in practical production, significant morphological differences exist in the same type of defects. Meanwhile, affected by missed detection, false detection, spatial drift, and localization error, defects do not always present a strict or stable periodic distribution. Therefore, a two-stage generalized category discovery method based on hierarchical distribution alignment regularization was proposed to adapt to the new defect morphology appearing in production. In this method, the known cluster supervised model was utilized as a knowledge anchor. By implicitly balancing the marginal probabilities of known and unknown class groups in a batch and explicitly promoting the prediction uniformity within respective groups, a more structured prediction distribution control was achieved, and the defect clustering accuracy was effectively improved. The results show that compared with that of the method, the clustering accuracy of the proposed method on all categories of the steel surface repetitive mixed defect dataset is improved by 4.54%, and that on unknown categories is improved by 5.87%, which demonstrates a significant advantage in effective separation between categories.
针对上述缺陷聚类的难题,本文提出了一种基于层级分布对齐正则化(hierarchical distribution alignment regularization,HDAR)的两阶段广义类别发现方法,以图像特征为基础对钢材表面重复性缺陷进行聚类.该方法不依赖缺陷的时空位置信息,而是利用已知聚类簇监督模型作为知识锚点,通过隐式地平衡批次中已知类与未知类群组的边缘概率、显式地促进各自群组内部预测的均匀性,辅助对已知簇和未知簇的发现,有效维持了全局类别平衡,在缓解旧类遗忘的同时强化了未知类辨识度,进而显著提升了缺陷聚类精度.本文的代码与示例数据集已公开至https://github.com/zhifeng-d/HDAR.
WuKun-peng, ShiJie. Periodic defect detection method for strip steel surface based on Siamese network[J]. Metallurgical Industry Automation, 2020, 44(6): 93-98.
LiFeng-fan, KuangJian-long, JiJia-hao, et al. Application of machine learning in predicting the service performance of metallic materials[J]. Chinese Journal of Engineering, 2024, 46(1): 120-136.
[5]
JiangY, LiJ S, YangX, et al. Applications of generative adversarial networks in materials science[J]. Materials Genome Engineering Advances, 2024, 2(1): e30.
[6]
XieY F, HuW T, XieS W, et al. Surface defect detection algorithm based on feature-enhanced YOLO[J]. Cognitive Computation, 2023, 15(2): 565-579.
MaGe, LiHong-wei, YanZi-wei, et al. Improved small target detection algorithm based on multi-attention and YOLOv5s for traffic sign recognition[J]. Chinese Journal of Engineering, 2024, 46(9): 1647-1658.
[9]
VazeS, HanK, VedaldiA, et al. Generalized category discovery[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). New Orleans,2022: 7492-7501.
[10]
ZhengM K, YouS, HuangL, et al. SimMatch: semi-supervised learning with similarity matching[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). New Orleans, 2022: 14451-14461.
[11]
SohnK, BerthelotD, CarliniN, et al. FixMatch: simplifying semi-supervised learning with consistency and confidence[J]. Advances in Neural Information Processing Systems, 2020, 33: 596-608.
[12]
HanK, RebuffiS A, EhrhardtS, et al. AutoNovel: automatically discovering and learning novel visual categories[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(10): 6767-6781.
[13]
ZhongZ, FiniE, RoyS, et al. Neighborhood contrastive learning for novel class discovery[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR). Nashville, 2021: 10862-10870.
[14]
MaS J, ZhuF, ZhongZ, et al. Happy: a debiased learning framework for continual generalized category discovery[J]. Advances in Neural Information Processing Systems, 2024, 37: 50850-50875.
[15]
GuP Y, ZhangC Y, XuR J, et al. Class-relation knowledge distillation for novel class discovery[C] //2023 IEEE/CVF International Conference on Computer Vision (ICCV). Paris,2024: 16428-16437.
[16]
WangY Z, ChenZ Y, YangD K, et al. Self-cooperation knowledge distillation for novel class discovery[C]//Computer Vision—ECCV 2024. Cham: Springer, 2025: 459-476.
[17]
WenX, ZhaoB C, QiX J. Parametric classification for generalized category discovery: a baseline study[C]//2023 IEEE/CVF International Conference on Computer Vision(ICCV). Paris,2024: 16544-16554.
[18]
HuZ Y, DuanY, ZhangY M, et al. Prototypical classifier with distribution consistency regularization for generalized category discovery: a strong baseline[J]. Neural Networks, 2025, 182: 106908.
[19]
CaoX Z, ZhengX Y, WangG H, et al. Solving the catastrophic forgetting problem in generalized category discovery[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).Seattle, 2024: 16880-16889.
[20]
WangY, ZhaoB C, LuY C, et al. Debiased prototypical learning improves generalized category discovery[C]//2024 IEEE International Conference on Multimedia and Expo (ICME).Niagara Falls, 2024: 1-6.
[21]
FiniE, SanginetoE, LathuilièreS, et al. A unified objective for novel class discovery[C]//2021 IEEE/CVF International Conference on Computer Vision(ICCV).Montreal, 2021: 9284-9292.
[22]
HanK, VedaldiA, ZissermanA. Learning to discover novel visual categories via deep transfer clustering[C]// IEEE/CVF International Conference on Computer Vision(ICCV).Seoul,2019: 8401-8409.
[23]
CaronM, TouvronH, MisraI, et al. Emerging properties in self-supervised vision Transformers[C]//2021 IEEE/CVF International Conference on Computer Vision(ICCV).Montreal,2021: 9630-9640.
[24]
DosovitskiyA, BeyerL, KolesnikovA, et al. An image is worth 16x16 words: Transformers for image recognition at scale[EB/OL]. (2020-10-22)[2025-11-30].
BaiYun-kun, ZhangHao-yu, XingYu-xiang. A review of industrial image anomaly detection based on unsupervised deep learning [J]. Chinese Journal of Stereology and Image Analysis, 2025,30(1):102-125.
WuJun-fei, GuXin-fu. Mechanism of abnormal rolling microstructure in Ti2AINb alloy hot-rolled plate[J]. Chinese Journal of Stereology and Image Analysis, 2024,29(1):19-28.
[29]
BaoY Q, SongK C, LiuJ, et al. Triplet-graph reasoning network for few-shot metal generic surface defect segmentation[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 5011111.
[30]
LyuX M, DuanF J, JiangJ J, et al. Deep metallic surface defect detection: the new benchmark and detection network[J]. Sensors, 2020, 20(6): 1562.
[31]
DalalN, TriggsB. Histograms of oriented gradients for human detection[C]//2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR). San Diego, 2005: 886-893.
[32]
ZhaoB C, WenX, HanK. Learning semi-supervised gaussian mixture models for generalized category discovery[C]//2023 IEEE/CVF International Conference on Computer Vision(ICCV).Paris, 2023: 16577-16587.
[33]
PuN, ZhongZ, SebeN. Dynamic conceptional contrastive learning for generalized category discovery[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).Vancouver, 2023: 7579-7588.
[34]
ChoiS, KangD, ChoM. Contrastive mean-shift learning for generalized category discovery[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).Seattle,2024: 23094-23104.
[35]
Van der MaatenL, HintonG. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(11): 2579-2605.