| 引用本文: | 成彬,王井浩,何博,雷华.基于FN-YOLOv5的连铸坯表面缺陷检测方法[J].材料科学与工艺,2025,33(3):57-66.DOI:10.11951/j.issn.1005-0299.20230309. |
| CHENG Bin,WANG Jinghao,HE Bo,LEI Hua.Surface defect detection method for continuous casting billets based on FN-YOLOv5[J].Materials Science and Technology,2025,33(3):57-66.DOI:10.11951/j.issn.1005-0299.20230309. |
|
| 摘要: |
| 为解决连铸坯表面缺陷类别多、特征差异大造成的检测精度低、误检漏检和难以实时检测等问题,提出一种基于改进YOLOv5s的面向连铸坯表面缺陷分形特征检测算法(FN-YOLOv5),以实现连铸坯表面缺陷的快速、精确、智能化检测。首先,在YOLOv5s模型多尺度特征融合网络引入压缩与激励注意力机制,实现特征权重自适应调整；其次,采用BiFPN网络替换原始网络结构,提高模型多尺度特征融合能力；最后,基于Swin Transformer引入C3STR模块,增强模型密集目标信息捕获能力。面向连铸坯表面缺陷数据集和经典开源热轧带钢缺陷数据集NEU-DET的实验结果表明:FN-YOLOv5算法在两个数据集的平均检测精确率分别达到0.786和0.784,较YOLOv5s算法分别提高5.4%和4.7%,检测速度分别为91.74和88.64帧每秒。在满足实际应用需求基础上,验证了检测精度和普适性能力的提升,与其他经典目标检测算法相比,FN-YOLOv5整体表现更出色均衡,为钢铁冶金领域智能化无损检测提供技术参考。 |
| 关键词: 连铸坯表面缺陷 目标检测 通道注意力机制 多尺度特征融合 滑动窗口变换 |
| DOI：10.11951/j.issn.1005-0299.20230309 |
| 分类号:TP183;TP391.4;TG245 |
| 文献标识码:A |
| 基金项目:陕西省自然科学基础研究计划项目(2021JM-360). |
|
| Surface defect detection method for continuous casting billets based on FN-YOLOv5 |
|
CHENG Bin1,WANG Jinghao1,HE Bo2,LEI Hua2
|
|
(1. School of Mechanical and Electrical Engineering, Xian University of Architecture and Technology, Xi’an 710055, China;2. China National Heavy Machinery Research Institute Co., Ltd., Xi’an 710016, China)
|
| Abstract: |
| To solve the problems of low detection accuracy, false positives, false negatives and the challenge of real-time detection caused by multiple types of surface defects and significant feature variations in continuous casting billets, an improved YOLOv5s based fractal nature detection algorithm for surface defects in continuous casting billets (FN-YOLOv5) is proposed to achieve fast, accurate, and intelligent detection of surface defects in continuous casting billets. Firstly, the SE attention mechanism is introduced into the multi-scale feature fusion network of the YOLOv5s model to achieve adaptive adjustment of feature weights. Secondly, the BiFPN network is used to replace the original network structure and improve the multi-scale feature fusion ability of the model. Finally, based on the Swin Transformer, the C3STR module is introduced to enhance the models ability to capture dense target information. The experimental results on both the continuous casting billet surface defect dataset and the classic open-source hot rolled strip defect dataset NEU-DET show that the FN-YOLOv5 algorithm has an average detection accuracy of 0.786 and 0.784, respectively in the two datasets, which shows an improvement of 5.4% and 4.7% compared to the YOLOv5s algorithm. The detection speedsare 91.74 frames per second and 88.64 frames per second, respectively. On the basis of meeting practical application requirements, the improvement of detection accuracy and universality ability are validated. Compared with other classic object detection algorithms, FN-YOLOv5 demonstrate superior overall performance,providing technological reference for intelligent non-destructive testing in the field of steel and metallurgy. |
| Key words: surface defects of continuous casting billets target detection channel attention mechanism multi scale feature fusion sliding window transform |
