摘要
针对航空发动机内部检测叶片凸台缺陷的问题,提出了一种基于YOLOv4(You Only Look Once)的目标检测算法。算法使用迁移学习加载了在coco公开数据集上训练的预训练模型权重,为了更好的适应对凸台检测中小目标、结构复杂的特点,通过聚类分析的方法调整先验框尺寸,同时对原始数据集使用Mosaic方法进行数据增强。实验结果表明,改进后的YOLOv4模型在检测精度上提高了15.85%,召回率提高了21%,平均交并比可达0.75,检测性能优于在同一数据集中使用的SSD目标检测算法。
Aiming at the defect detection of the blade boss inside the aero engine,an object detection algorithm based on YOLOv4(You Only Look Once) is proposed.The algorithm used transfer learning to load the weights of the pre-trained model trained on the coco public data set.In order to adapt to the characteristics of small targets and complex structures in the detection of bosses,the size of the bounding boxes was adjusted through cluster analysis.Then the original data were enhanced by the Mosaic method.The experimental results show that the improved YOLOv4 model increases the detection accuracy by 15.85%,the recall rate by 21%,with an average intersection ratio of 0.75,and the detection performance is significantly better than the SSD object detection algorithm.
作者
陈为
钟欣童
张婧
李泽辰
CHEN Wei;ZHONG Xin-tong;ZHANG Jing;LI Ze-chen(College of Automation and Electronic Engineering,Qingdao University of Science and Technology,Qingdao Shandong 266000,China;Qingdao Beihai Ship Building Heavy Industry CO.,LTD.,Qingdao Shandong 266000,China)
出处
《计算机仿真》
北大核心
2022年第7期17-21,共5页
Computer Simulation
基金
国家自然科学基金(61304093)。
关键词
目标检测
叶片凸台检测
聚类分析
数据增强
Object detection
Blade boss detection
Cluster analysis
Data enhancement
作者简介
陈为(1974-),男(汉族),山东省青岛市人,博士,副教授,主要研究领域为计算机控制、嵌入式系统开发、机器视觉等;钟欣童(1996-),女(汉族),山东省烟台市人,硕士研究生,主要研究领域为目标检测、神经网络等;张婧((1978-),女(汉族),山东省青岛市人,博士,工程师,主要研究领域为自动化;李泽辰(1994-),男(汉族),山东省临沂市人,硕士研究生,主要研究领域为机器视觉。
