基于自蒸馏的热轧带钢表面缺陷识别

doi:10.13228/j.boyuan.issn1006-9356.20230527

摘要
图/表
参考文献
相关文章 (15)

全文: PDF (0 KB) HTML (0 KB)
输出: BibTeX | EndNote (RIS)

摘要表面缺陷检测是热轧带钢生产过程中十分重要的环节,是提高热轧带钢产品质量的关键。首先,针对当前带钢表面缺陷图像形貌多样、干扰多以及检测算法精度有待提高的问题,设计了一种自蒸馏框架(Self-knowledge Distillation,SD)来提高模型的缺陷检测精度,通过采用几何变换、色彩增强和图像线性插值的数据增强方法,增加样本数据的表征信息;从标签的概率分布和接近程度两方面提出了新的损失函数(LCK Loss),使模型能更好地学习样本提供的知识,实现相同数据不同表征间的信息传递,进而提高模型的泛化能力。其次,将Poolformer网络应用到热轧带钢表面缺陷检测中,针对Poolformer12网络参数量多的问题,设计了轻量化网络LRAM-Poolformer8,通过减少网络层数和计算量,实现模型加速的目的。最后,在武钢CSP机组的8类带钢表面缺陷数据集上进行了试验,结果表明,所提SD-LRAM-Poolformer8模型的平均识别精度为98.20%,相较于Poolformer12,检测精度提高了1.62个百分点,且计算量仅为原来的56.4%,这证明了新模型在热轧带钢表面缺陷检测中的可行性与有效性。

	服务

	把本文推荐给朋友
	加入我的书架
	加入引用管理器
	E-mail Alert
	作者相关文章
	李秋雨
	李维刚
	田志强

关键词 ：热轧带钢, 表面缺陷检测, 自蒸馏, 轻量化网络, 深度学习

Abstract：Surface defect detection is critical aspect in the production process of hot-rolled strip steel, serving as a key factor in improving the quality of these steel products. Firstly, to address the challenges posed by the diverse morphologies of surface defect images, various interferences, and the need for improved detection accuracy, a self-knowledge distillation framework (SD) was designed to enhance the defect detection accuracy of the model. The representational information of sample data was increased by employing data augmentation methods such as geometric transformations, color enhancement, and linear interpolation. Additionally, a novel loss function (LCK Loss) was proposed, considering both the label's probability distribution and closeness, enabling the model to better learn the knowledge provided by the samples and facilitating the transfer of information between different representations of the same data, thereby improving the model's generalization ability. Secondly, the Poolformer network was applied to the surface defect detection of hot-rolled strip steel. To address the issue of a large number of parameters in the Poolformer12 network, a lightweight network called LRAM-Poolformer8 was designed, achieving model acceleration by reducing network depth and computational complexity. Finally, experiments were conducted on 8-class surface defect dataset from WISCO's CSP unit. The results demonstrate that the proposed SD-LRAM-Poolformer8 model achieves average recognition accuracy of 98.20%. Compared to Poolformer12, it shows an improvement of 1.62 percent points in detection accuracy while reducing the computational complexity to only 56.4% of the original model. These findings highlight the feasibility and effectiveness of the new model in surface defect detection of hot-rolled strip steel.

Key words： hot-rolled strip steel surface defect detection self-knowledge distillation lightweight network deep learning

收稿日期: 2023-09-07

基金资助:国家自然科学基金资助项目(51774219)

通讯作者: 李维刚(1977—), 男, 博士, 教授; E-mail: liweigang.luck@foxmail.com

作者简介: 李秋雨(1998—), 男, 硕士生; E-mail: 1173547041@qq.com

引用本文:

李秋雨, 李维刚, 田志强. 基于自蒸馏的热轧带钢表面缺陷识别[J]. 中国冶金, 2024, 34(2): 126-133. LI Qiuyu, LI Weigang, TIAN Zhiqiang. Surface defect recognition of hot-rolled strip steel based on self-knowledge distillation[J]. China Metallurgy, 2024, 34(2): 126-133.

链接本文:

http://www.zgyj.ac.cn/CN/10.13228/j.boyuan.issn1006-9356.20230527 或 http://www.zgyj.ac.cn/CN/Y2024/V34/I2/126

[1]	王文广, 徐芳, 李兴波, 等. 热连轧高速钢轧辊对产品表面质量和板形控制的影响[J]. 中国冶金, 2022, 32(2):97.(WANG W G, XU F, LI X B, et al. Influence of HSS roll on product surface quality and shape control in hot strip rolling[J]. China Metallurgy, 2022, 32 (2): 97.)
[2]	焦安杰, 张盛攀, 李燕春, 等. 轧制工艺对高强耐候钢板形控制的影响[J]. 中国冶金, 2021, 31(6):82.(JIAO A J, ZHANG S P, LI Y C, et al. Effect of rolling process on shape control of high strength weathering resistant steel scale[J]. China Metallurgy, 2021, 31 (6): 82.)
[3]	刘坤, 王晓晨, 张天明, 等. 基于有限元仿真的热连轧精轧带钢楔形控制分析[J]. 中国冶金, 2023, 33(9):104.(LIU K, WANG X C, ZHANG T M, et al. Wedge control analysis of finishing strip based on finite element simulation in hot continuous rolling[J]. China Metallurgy, 2023, 33 (9):104.)
[4]	孙建亮, 孙孟乾, 郭贺松, 等. 热轧带材多缺陷和单缺陷表面质量综合预报[J]. 钢铁, 2021, 56(1):51.(SUN J L, SUN M Q, GUO H S, et al. Comprehensive prediction of surface quality of hot-rolled strip with multiple and single defects[J]. Iron and Steel, 2021, 56 (1): 51.)
[5]	张创举, 申文君, 潘时松, 等. Q355热轧板山峰状表面裂纹形貌特征和形成原因[J]. 中国冶金, 2023, 33(1):55.(ZHANG C J, SHEN W J, PAN S J, et al. Morphology characteristic and formation cause of mountain-like surface crack in Q355 hot-rolled plate[J]. China Metallurgy, 2023, 33 (1): 55.)
[6]	任英, 李鑫哲, 袁天祥, 等. 热轧卷板表面夹渣缺陷来源分析及控制现状[J]. 钢铁, 2022, 57(1):13. (REN Y, LI X Z, YUAN T X, et al. Source analysis and control status of slag defects on surface of hot-rolled coil[J]. Iron and Steel, 2022, 57 (1): 13.)
[7]	张文军, 刘格非, 王孝建, 等. 热镀铝锌带钢表面缺陷形成机理及工艺优化[J]. 钢铁, 2023, 58(4):87. (ZHANG W J, LIU G F, WANG X J, et al. Formation mechanism and process optimization of surface defects on hot-dip Al-Zn strip steel[J]. Iron and Steel, 2023, 58 (4): 87.)
[8]	王章印, 姜敏, 王新华. Q345D钢精炼过程夹杂物生成及演变行为[J]. 钢铁, 2022, 57(2):63.(WANG Z Y, JIANG M, WANG X H. Formation and evolution of inclusions in Q345D steel during secondary refining process[J]. Iron and Steel, 2022, 57 (2): 63.)
[9]	徐科, 周鹏, 贺笛, 等. 机器视觉技术及在钢铁生产中的应用[C]//第十二届中国钢铁年会论文集——大会特邀报告&分会场特邀报告. 北京:中国金属学会, 2019: 137.(XU K, ZHOU P, HE D, et al. Machine vision technology and its application in steel production[C]//12th China Steel Annual Conference-Invited Report of the Conference and Branch Venues. Beijing:The Chinese Society for Metals, 2019: 137.)
[10]	康永林. “十三五”中国轧钢技术进步及展望[J]. 钢铁, 2021, 56(10):1.(KANG Y L. China steel rolling technology progress in the 13th five-year plan and prospection[J]. Iron and Steel, 2021, 56 (10): 1.)
[11]	郭贺松, 孙建亮, 叶春林, 等. 基于SAE-DBN混合深度网络的热轧带钢表面缺陷预报[J]. 钢铁, 2023, 58(8):120.(GUO H S, SUN J L, YE C L, et al. Surface defect prediction of hot rolled strip based on SAE-DBN hybrid depth network[J]. Iron and Steel, 2023, 58 (8): 120.)
[12]	邓能辉, 侯睿, 叶俊明. 基于深度学习的圆钢表面缺陷检测系统[J]. 中国冶金, 2022, 32(12):113.(DENG N H, HOU R, YE J M. Round steel surface defect detection system based on deep learning[J]. China Metallurgy, 2022, 32 (12): 113.)
[13]	李金灵, 李维刚, 陈燕才, 等. 基于改进YOLOv5算法的带钢表面缺陷检测[J]. 钢铁研究学报, 2023, 35(6): 767.(LI J L, LI W G, CHEN Y C, et al. Strip surface defect detection based on improved YOLOv5 algorithm[J]. Journal of Iron and Steel Research, 2023, 35(6): 767.)
[14]	周鹏, 徐科, 杨朝霖. 基于SIFT的中厚板表面缺陷识别方法[J]. 清华大学学报(自然科学版), 2018, 58(10): 881.(ZHOU P, XU K, YANG C L. Surface defect recognition method for medium and thick plates based on SIFT[J]. Journal of Tsinghua University (Natural Science Edition), 2018, 58(10): 881.)
[15]	寇旭鹏, 刘帅君, 麻之润. 基于Faster-RCNN的钢带缺陷检测方法[J]. 中国冶金, 2021, 31(4): 77.(KOU X P, LIU S J, MA Z R. Defect detection method of steel strip based on Faster-RCNN[J]. China Metallurgy, 2021, 31(4): 77.)
[16]	沈春光, 李虎威, 荆涛, 等. 基于深度学习的带钢表面缺陷检测在小样本数据集的应用[J]. 轧钢, 2022, 39(2):82.(SHEN C G, LI H W, JING T, et al. Application of deep learning-based surface defect recognition of plate in small sample dataset[J]. Rolling Steel, 2022, 39 (2): 82.)
[17]	贺笛. 深度学习在钢板表面缺陷与字符识别中的应用[D]. 北京:北京科技大学, 2021.(HE D. Application of Deep Learning Method in Detecting Steel Surface Defects and Chars[D]. Beijing: University of Science and Technology Beijing,2021.)
[18]	李维刚, 徐康, 李金灵, 等. 热轧带钢表面缺陷识别算法研究与应用[J]. 钢铁, 2022, 57(10): 139.(LI W G, XU K, LI J L, et al. Research and application of surface defect identification algorithm for hot rolled strip steel[J]. Iron and Steel, 2022, 57(10): 139.
[19]	MAO A, MOHRI M, ZHONG Y. Cross-entropy loss functions: Theoretical analysis and applications[J/OL]. (2023-06-01). http://arxiv.org/pdf/2304.07288.pdf.
[20]	YIN S, ZHU P, WU X, et al. Integrating information by Kullback-Leibler constraint for text classification[J]. Neural Computing and Applications, 2023, 35: 17521.
[21]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J/OL]. (2023-06-01). http://arxiv.org/pdf/2010.11929.pdf.
[22]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]// CVF conference on computer vision and pattern recognition. Las Vegas:IEEE, 2016: 770.
[23]	YU W, LUO M, ZHOU P, et al. Metaformer is actually what you need for vision[C]// CVF Conference on Computer Vision and Pattern Recognition. New Orleans:IEEE,2022: 10819.
[24]	HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[J/OL]. (2023-06-01). http://arxiv.org/pdf/1053.02531.pdf.
[25]	BHAT P, ARANI E, ZONOOZ B. Distill on the go: online knowledge distillation in self-supervised learning[C]// /CVF Conference on Computer Vision and Pattern Recognition. Nashville:IEEE,2021: 2678.
[26]	TOLSTIKHIN I O, HOULSBY N, KOLESNIKOV A, et al. MLP-mixer: An all-mlp architecture for vision[J]. Advances in neural information processing systems, 2021, 34: 24261.
[27]	GUIBAS J, MARDANI M, LI Z, et al. Adaptive fourier neural operators: Efficient token mixers for transformers[J/OL]. (2023-06-01). http://arxiv.org/pdf/2111.13587.pdf.
[28]	TOUVRON H, CORD M, DOUZE M, et al. Training data-efficient image transformers & distillation through attention[C]//International conference on machine learning.[s.l.]: PMLR, 2021: 10347.
[29]	DEVRIES T, TAYLOR G W. Improved regularization of convolutional neural networks with cutout[J/OL]. (2023-06-01). http://arxiv.org/pdf/1708.04552.pdf.
[30]	ZHANG H, CISSE M, DAUPHIN Y N, et al. mixup: Beyond empirical risk minimization[J/OL]. (2023-06-01). http://arxiv.org/pdf/1710.09412.pdf.