基于改进YOLOv5的X射线图像铸件缺陷实时检测  

Real-time detection of casting defect in X-ray images based on improved YOLOv5

作　　者：胡哲 徐承志[1] 雷光波[1] 徐丽[1] HU Zhe;XU Chengzhi;LEI Guangbo;XU Li(School of Computer Science,Hubei University of Technology,Wuhan 430068,China)

机构地区：[1]湖北工业大学计算机学院,武汉430068

出　　处：《激光杂志》2022年第5期54-59,共6页Laser Journal

基　　金：湖北省教育厅科学技术研究计划项目(No.B2019049)。

摘　　要：铸件缺陷检测是一项重要的质量管理程序。为了避免人为因素的影响,提高检测精度,对YOLOv5s6的目标检测算法进行改进,用于X射线图像的铸件缺陷检测。首先设计了一种C3CA模块用于捕获跨通道、方向感知和位置感知的信息。然后通过在骨干网络中融合多头自注意力机制来捕获局部与全局信息。最后采用Focal-EIoU Loss损失函数。实验结果表明:在相同训练条件下,改进后YOLOv5s6的AP50值达到了90.2%,F1因子达到了87.8%,相较原始模型分别提高了3.4%和2.3%,具有检测准确率高、实时性强等特点。Casting defect detection is a significant quality management procedure.In order to avoid the influence of human factors and improve the detection accuracy,an improved YOLOv5s6 target detection algorithm is proposed for X-ray image casting defect detection.Firstly,a C3CA module is designed to capture cross-channel,direction-aware,and location-aware information.Then,the multi-head self-attention mechanism is integrated in the backbone network to capture local and global information.Finally,the Focal-EIoU Loss function is adopted.The experimental results show that under the same training conditions,the AP50 value of the improved YOLOv5s6 reaches 90.2%,and the F1factor reaches 87.8%,which are 3.4%and 2.3%higher than the original model,respectively.It has the characteristics of high detection accuracy and strong real-time performance.

关 键 词：铸件缺陷检测 多头自注意力机制 YOLOv5s6 X射线图像 

分 类 号：TN911[电子电信—通信与信息系统]