-
摘要: 为了解决传统人工方法对废钢分类评级人为因素干扰大且效率低下等问题，提出基于挤压−激励（Squeeze−Excitation，SE）注意力机制构建废钢分类评级的深度学习网络模型，并对采集到的废钢卸载过程图像进行模型训练和验证。首先，搭建物理尺寸比例为1∶3废钢质量查验物理模型，采用高分辨率视觉传感器模拟采集货车卸载废钢作业场景下不同废钢的形貌特征；然后，对采集到的废钢图像使用跨阶段局部网络进行特征提取，利用空间金字塔结构解决特征丢失问题，采用注意力机制关注通道间的相关性；最后，在包含7个标签分类的两个数据集进行模型训练与验证。实验表明：该模型能够有效地对不同级别的废钢进行自动评级判定，全类别准确率达到83.7%，全类别平均精度为88.8%，在准确性方面相比于传统人工验质方法具有显著优势，解决了废钢入库过程中质量评价的公正性难题。Abstract: Not only is scrap steel an indispensable ferritic raw material for the modern steel industry, but it is also the only green raw material that can replace iron ore in large quantities. The quality of the scrap steel directly affects the quality of molten steel, which makes it necessary to sort and grade scrap steel before it enters the furnace. Most iron and steel enterprises determine the grade of scrap steel mainly by visual inspection and caliper-based measurements by quality management personnel. As a result, this process is prone to human errors and low efficiency. Therefore, given that the major challenges of scrap inspection include the many categories of scrap, complex actual detection scenarios, and challenges in manual system connection, a deep learning network model CSSNet was proposed for scrap classification and rating based on the Squeeze-Excitation (SE) attention mechanism, and images of the scrap unloading process were collected for model training and validation. First, a 1∶3 physical model of scrap steel quality inspection was built to simulate this process. High-resolution visual sensors were used to collect images of diverse types of scrap steel in the scene of trucks unloading scrap steel. Then, a cross-stage local network was used to extract the features of the collected scrap images, the spatial pyramid structure was used to solve the problem of feature loss, and the attention mechanism was used to focus on the correlation between channels and retain the channel with the most feature information. Finally, model training and validation were done using two datasets containing seven labels for classification. In the model prediction stage, the constructed scrap steel quality inspection model CSSNet was used to judge the scrap steel category and quality to verify the accuracy and detection efficiency of the model. Based on the self-made scrap steel validation dataset, its performance was compared with mainstream single-stage object detection packages such as YOLOv4, YOLOv5s, and the two-stage object detection model Faster R-CNN. The model was found to be able to effectively rate different grades of scrap steel, with the classification accuracy rate of all categories has reached 83.7% and an mAP value of 88.8%. The performance of the CSSNet model is better than the other three target detection models. CSSNet can not only fully meet the needs of the actual production applications in terms of accuracy, real-time performance, and identification and rating efficiency but also surpass the traditional manual scrap quality inspection method, address multiple issues in the evaluation of scrap steel quality, and realize automated scrap steel quality testing.
-
图 5 模型在HK_S和HK_L数据集上loss值变化图. (a) HK_S数据集，batch-size=16; (b)为HK_L数据集，batch-size=16; (c) HK_S数据集（添加SE注意力）; (d) HK_L数据集（添加SE注意力）
Figure 5. Changes in the loss value of the model on the HK_S and HK_L datasets: (a) HK_S dataset, batch-size=16; (b) HK_L dataset, batch-size=16; (c) HK_S dataset (add SE attention); (d) HK_L dataset (add SE attention)
图 6 模型加入SE注意力机制前后的表现效果对比. (a)未加SE注意力机制; (b)添加SE注意力机制
Figure 6. Comparison of performance effects before and after adding the SE attention mechanism into the model: (a) no SE attention mechanism; (b) add SE attention mechanism
图 7 模型检测效果图. (a)为未检测废钢图像; (b)检测后的废钢图像
Figure 7. Model detection renderings: (a) the undetected scrap image; (b) the detected scrap image
图 8 各类别评价指标曲线. (a) P−R曲线; (b) F1曲线; (c) R曲线; (d) P曲线
Figure 8. Evaluation index curve of each category: (a) P–R curve; (b) F1 curve; (c) R curve; (d) P curve
表 1 HK_S、HK_L数据集
Table 1. HK_S and HK_L datasets
Datasets Images Labels Training images Training labels Validation images Validation labels HK_S 139 6297 125 5750 14 547 HK_L 278 12592 250 11388 28 1204 
下载: 导出CSV
表 2 HK_S和HK_L数据集各类别标签数量
Table 2. Number of labels for each category in HK_S and HK_L datasets
Label category HK_S labels HK_S training labels HK_L labels HK_L training labels <3 mm 92 57 184 164 3−6 mm 327 319 654 598 >6 mm 4799 4393 9596 8668 Galvanized 365 337 730 663 Greasy dirt 196 173 392 359 Paint 126 89 252 233 Inclusion 392 382 784 703
下载: 导出CSV
表 3 正例与负例
Table 3. Positive and negative
Type P (Positive) N (Negative) T (True) True positive (TP) True negative (TN) F (False) False positive (FP) False negative (FN)
下载: 导出CSV
表 4 HK_S数据集模型评价指数
Table 4. HK_S dataset model evaluation index
Model Batch-size Epoch F1 mAP Yolov5s 8 200 0.48 0.506 CSP+SPP 8 200 0.48 0.552 CSP+SPP+SE 8 200 0.60 0.642 Yolov5s 16 200 0.64 0.646 CSP+SPP 16 200 0.63 0.665 CSP+SPP+SE 16 200 0.71 0.719 Yolov5s 32 200 0.68 0.684 CSP+SPP 32 200 0.66 0.709 CSP+SPP+SE 32 200 0.70 0.720 CSP+SPP+SE 32 300 0.75 0.754
下载: 导出CSV
表 5 HK_L数据集模型评价指数
Table 5. HK_L dataset model evaluation index
Model Batch-size Epoch F1 mAP Yolov5s 8 200 0.61 0.641 CSP+SPP 8 200 0.69 0.699 CSP+SPP+SE 8 200 0.75 0.755 Yolov5s 16 200 0.79 0.792 CSP+SPP 16 200 0.79 0.802 CSP+SPP+SE 16 200 0.83 0.833 Yolov5s 32 200 0.82 0.805 CSP+SPP 32 200 0.84 0.839 CSP+SPP+SE 32 200 0.87 0.868 CSP+SPP+SE 32 400 0.87 0.888
下载: 导出CSV
表 6 不同网络模型检测结果比较
Table 6. Comparison of detection results of different network models
Model Datasets mAP/% YOLOv4 HK_S 60.0 YOLOv5s HK_S 50.6 Faster R-CNN HK_S 50.9 CSSNet HK_S 64.2 YOLOv4 HK_L 68.1 YOLOv5s HK_L 64.1 Faster R-CNN HK_L 64.1 CSSNet HK_L 75.5
下载: 导出CSV
表 7 各类别验证集的表现情况
Table 7. Performance under each category of the validation set
Class Images Labels P/% R/% AP/% <3 mm 28 20 100 79.8 86.6 3–6 mm 28 56 89.5 91.2 93.7 >6 mm 28 928 91.2 83.7 88.8 Galvanized 28 67 87.3 94.0 94.0 Paint 28 81 92.0 85.2 91.9 Greasy dirt 28 33 83.1 69.7 76.7 Inclusion 28 19 100 81.6 89.8
下载: 导出CSV
-
参考文献
[1] Wu Y G, Xiao B Q, Zhu L G, et al. Current situation analysis and prospect of iron and steel raw material for electric arc furnace steelmaking. Iron Steel, 2021, 56(11): 55 doi: 10.13228/j.boyuan.issn0449-749x.20210170吴耀光, 肖步庆, 朱立光, 等. 电炉炼钢钢铁原料的现状分析与展望. 钢铁, 2021, 56(11):55 doi: 10.13228/j.boyuan.issn0449-749x.20210170 [2] Zhu R, Wu X T, Wei G S, et al. Development of green and intelligent technologies in electric arc furnace steelmaking processes. Iron Steel, 2019, 54(8): 9 doi: 10.13228/j.boyuan.issn0449-749x.20190188朱荣, 吴学涛, 魏光升, 等. 电弧炉炼钢绿色及智能化技术进展. 钢铁, 2019, 54(8):9 doi: 10.13228/j.boyuan.issn0449-749x.20190188 [3] Shangguan F Q, Yin R Y, Li Y, et al. Dissussion on strategic significance of developing full scrap EAF process in China. Iron Steel, 2021, 56(8): 86 doi: 10.13228/j.boyuan.issn0449-749x.20200401上官方钦, 殷瑞钰, 李煜, 等. 论中国发展全废钢电炉流程的战略意义. 钢铁, 2021, 56(8):86 doi: 10.13228/j.boyuan.issn0449-749x.20200401 [4] Wang Z L, Gu C, Wang M, et al. Research progress and application status of deep learning in steelmaking process. Chin J Eng, http://kns.cnki.net/kcms/detail/10.1297.TF.20220506.1903.003.html王仲亮, 顾超, 王敏, 等. 深度学习在炼钢过程中的研究进展及应用现状. 工程科学学报, http://kns.cnki.net/kcms/detail/10.1297.TF.20220506.1903.003.html [5] Shangguan F Q, Li X P, Zhou J C, et al. Strategic research on development of steel scrap resources in China. Iron Steel, 2020, 55(6): 8 doi: 10.13228/j.boyuan.issn0449-749x.20190427上官方钦, 郦秀萍, 周继程, 等. 中国废钢资源发展战略研究. 钢铁, 2020, 55(6):8 doi: 10.13228/j.boyuan.issn0449-749x.20190427 [6] Ma Y Q. Do iron ore, scrap steel, carbon emission allowance, and seaborne transportation prices drive steel price fluctuations? Resour Policy, 2021, 72: 102115 doi: 10.1016/j.resourpol.2021.102115 [7] Smirnov N V, Rybin E I. Machine learning methods for solving scrap metal classification task // 2020 International Russian Automation Conference (RusAutoCon). Sochi, 2020: 1020 [8] Gao Z J, Sridhar S, Spiller D E, et al. Applying improved optical recognition with machine learning on sorting Cu impurities in steel scrap. J Sustain Metall, 2020, 6(4): 785 doi: 10.1007/s40831-020-00300-8 [9] Koyanaka S, Kobayashi K. Automatic sorting of lightweight metal scrap by sensing apparent density and three-dimensional shape. Resour Conserv Recycl, 2010, 54(9): 571 doi: 10.1016/j.resconrec.2009.10.014 [10] Díaz-Romero D, Sterkens W, Van den Eynde S, et al. Deep learning computer vision for the separation of Cast-and Wrought-Aluminum scrap. Resour Conserv Recycl, 2021, 172: 105685 doi: 10.1016/j.resconrec.2021.105685 [11] Li Y F, Qin X P, Zhang Z Y, et al. A robust identification method for nonferrous metal scraps based on deep learning and superpixel optimization. Waste Manag Res, 2021, 39(4): 573 doi: 10.1177/0734242X20987884 [12] Chen S, Hu Z L, Wang C, et al. Research on the process of small sample non-ferrous metal recognition and separation based on deep learning. Waste Manag, 2021, 126: 266 doi: 10.1016/j.wasman.2021.03.019 [13] Ramsurrun N, Suddul G, Armoogum S, et al. Recyclable waste classification using computer vision and deep learning // 2021 Zooming Innovation in Consumer Technologies Conference (ZINC). Novi Sad, 2021: 11 [14] Mei Y G. Research on Intelligent Classification of Scrap Based on Machine Vision and LIBS Technology [Dissertation]. Beijing: University of Science and Technology Beijing, 2021梅亚光. 基于机器视觉与LIBS技术的废钢智能分类研究[学位论文]. 北京: 北京科技大学, 2021 [15] Xu G, Li M, Xu J W. Application of machine learning in automatic discrimination of product quality of deep drawn steel. Chin J Eng, 2022, 44(6): 1062 doi: 10.3321/j.issn.1001-053X.2022.6.bjkjdxxb202206011徐钢, 黎敏, 徐金梧. 机器学习在深冲钢质量自动判级中的应用. 工程科学学报, 2022, 44(6):1062 doi: 10.3321/j.issn.1001-053X.2022.6.bjkjdxxb202206011 [16] Li J Y, Yang Z F, Zheng J F, et al. Applications of iron and steel industry with deep learning technologies. Iron Steel, 2021, 56(9): 43 doi: 10.13228/j.boyuan.issn0449-749x.20210296李江昀, 杨志方, 郑俊锋, 等. 深度学习技术在钢铁工业中的应用. 钢铁, 2021, 56(9):43 doi: 10.13228/j.boyuan.issn0449-749x.20210296 [17] Wang D L, Zhang Y Q, Pan Y, et al. An automated inspection method for the steel box girder bottom of long-span bridges based on deep learning. IEEE Access, 2020, 8: 94010 doi: 10.1109/ACCESS.2020.2994275 [18] Zhang L, Yang F, Zhang Y D, et al. Road crack detection using deep convolutional neural network // 2016 IEEE International Conference on Image Processing (ICIP). Phoenix, 2016: 3708 [19] Liu Z D, Li W X, Wei Z H. Qualitative classification of waste textiles based on near infrared spectroscopy and the convolutional network. Text Res J, 2020, 90(9-10): 1057 doi: 10.1177/0040517519886032 [20] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need // Advances in Neural Information Processing Systems. Long Beach, 2017: 30 [21] Wang C Y, Liao H Y M, Wu Y H, et al. CSPNet: A new backbone that can enhance learning capability of CNN // Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Seattle, 2020: 1571 [22] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. Commun ACM, 2017, 60(6): 84 doi: 10.1145/3065386 [23] He K M, Zhang X Y, Ren S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell, 2015, 37(9): 1904 doi: 10.1109/TPAMI.2015.2389824 [24] Hu J, Shen L, Sun G. Squeeze-and-excitation networks // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, 2018: 7132 [25] Zhu X K, Lyu S C, Wang X, et al. TPH-YOLOv5: Improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios // Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, 2021: 2778 [26] He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image recognition // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, 2016: 770 [27] Alzubaidi L, Zhang J L, Humaidi A J, et al. Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions. J Big Data, 2021, 8(1): 53 doi: 10.1186/s40537-021-00444-8 -
点击查看大图
计量
- 文章访问数: 298
- HTML全文浏览量: 104
- PDF下载量: 74
- 被引次数: 0
