电子测量与仪器学报

马燕婷,赵红东,阎超,封慧杰,于快快,刘赫.改进 YOLOv5 网络的带钢表面缺陷检测方法[J].电子测量与仪器学报,2022,36(8):150-157

改进 YOLOv5 网络的带钢表面缺陷检测方法

Strip steel surface defect detection method by improved YOLOv5 network

DOI：

中文关键词: 缺陷检测多尺度特征融合 Transformer RepVGG

英文关键词:defect detection multi-scale feature fusion Transformer RepVGG

基金项目:天津市科技计划项目(企业科技特派员项目)(21YDTPJC00050)、光电信息控制和安全技术重点实验室基金项目(2021JCJQLB055008)资助

作者	单位
马燕婷	1. 河北工业大学电子信息工程学院
赵红东	1. 河北工业大学电子信息工程学院，2. 光电信息控制和安全技术重点实验室
阎超	1. 河北工业大学电子信息工程学院
封慧杰	1. 河北工业大学电子信息工程学院
于快快	3. 中国电子科技集团公司第五十三研究所
刘赫	4. 天津金沃能源科技有限公司

Author	Institution
Ma Yanting	1. School of Electronic and Information Engineering, Hebei University of Technology
Zhao Hongdong	1. School of Electronic and Information Engineering, Hebei University of Technology,2. Key Laboratory of Optoelectronic Information Control and Security Technology
Yan Chao	1. School of Electronic and Information Engineering, Hebei University of Technology
Feng Huijie	1. School of Electronic and Information Engineering, Hebei University of Technology
Yu Kuaikuai	3. The 53rd Research Institute of China Electronics Technology Group Corporation
Liu He	4. Tianjin Jinwo Energy Technology Company Limited

摘要点击次数: 959

全文下载次数: 1310

中文摘要:

带钢表面缺陷检测已成为保证带钢生产质量的重要环节之一。针对当前带钢缺陷检测算法精度有待提高等问题,提出了一种基于 YOLOv5 网络改进的算法模型 MT-YOLOv5。首先在主干网络中引入 Transformer 自注意力机制,使主干网络更聚焦于图像全局特征信息的提取;其次采用 T-BiFPN 网络结构,将 Transformer 层与 BiFPN 网络结构相结合,进一步增强了图像浅层特征信息与深层特征信息的融合;然后引入改进后的轻量化网络 RepVGG 替换主干网络中的部分卷积层,增强主干网络的特征提取能力;最后增加预测层,检测不同尺度的目标。实验结果表明,MT-YOLOv5 算法在 NEU-DET 数据集上的均值平均精度 (mAP)达到了 82. 4%,较原 YOLOv5s 算法提高了 5. 3%,检测速度为 65. 4 fps,更好地均衡了检测速度与检测精度。

英文摘要:

Strip steel surface defect detection has become one of the important links to guarantee the quality of strip steel production. Aiming at the problem of improving the detection accuracy of current strip steel defect detection algorithm, an improved MT-YOLOv5 algorithm based on YOLOv5 is proposed. Firstly, introducing Transformer self-attention mechanism in the backbone network to make the network more focused on the extraction of global image feature information. Secondly, combining the Transformer layer with the BiFPN structure, and the T-BiFPN network is used to further enhance the fusion of image shallow feature information and deep feature information. Then, an improved lightweight network RepVGG is introduced to replace part of the convolutional layers in the backbone network, which can enhance the feature extraction capability of backbone network. Finally, adding a prediction layer to detect objects of different scales. The experimental results show that the value of mean average precision (mAP) of the MT-YOLOv5 algorithm is 82. 4% on the NEU-DET dataset, which is 5. 3% higher than the original YOLOv5s algorithm, and the detection speed reaches 65. 4 fps, which achieves a better balance between detection speed and detection accuracy.

查看全文查看/发表评论下载PDF阅读器