融合SA注意力机制的YOLO5s在石油油管表面缺陷检测的应用  

The Application of YOLOv5s with SA Attention Mechanism in Surface Defect Detection of Oil Pipes

作　　者：郭桂标 邢雪[1] 刘宇琦 王超 孙明革[1] GUO Guibiao;XING Xue;LIU Yuqi;WANG Chao;SUN Mingge(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin Jilin 132022,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林吉林132022

出　　处：《机床与液压》2024年第24期228-235,共8页Machine Tool & Hydraulics

基　　金：吉林省教育厅科学技术研究项目(产业化培育项目JJKH20230306CY)。

摘　　要：针对石油厂油管表面缺陷检测存在检测精度低、速度慢和模型复杂等问题,提出一种SA-YOLO算法。以YOLOv5s模型为基础,对原数据集进行预处理,采用BoTNet Transformer结构代替Backbone特征主干的部分卷积,并用multi-head self-attention(MHSA)替换卷积层,以减少网络层,同时提高获取全局信息的能力;最后,将Shuffle Attention(SA)注意力机制融合到C3结构中,根据每个位置的重要性得到注意力权重,从而提高模型的泛化能力和计算效率,减少运行时间。实验结果表明:SA-YOLO算法在石油厂采集的数据集上的均值平均精度(mAP)达到了93%,较原YOLOv5s算法提高了3.3%,检测速度以及检测精度均明显提高。In view of the problems of low detection accuracy,slow speed,and complex models in the surface defect detection of oil pipes in oil plants,a SA-YOLO algorithm was proposed.Based on the YOLOv5s model,the original dataset was preprocessed,then the partial convolution of the Backbone feature backbone was replaced with the BoTNet Transformer structure,and directly the convolutional layer was replaced with multi-head self-attention(MHSA),to reduce network layers and improve the ability of obtaining global information.Finally,the Shuffle Attention mechanism was integrated into the C3 structure,and attention weights were obtained by the importance of each position,thereby improving the generalization ability and computational efficiency of the model,and reducing runtime.The experimental results show that the mean average precision(mAP)of the SA-YOLO algorithm on the dataset collected by the oil plant reaches 93%,which is 3.3%higher than the original YOLOv5s algorithm,the detection speed and accuracy are significantly improved.

关 键 词：缺陷检测 BoTNet Transformer结构 Shuffle Attention(SA)注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TG155[自动化与计算机技术—控制科学与工程]