基于改进YOLOv5 的轴类零件表面缺陷检测算法
Surface Defect Detection Algorithm of Shaft Parts Based on Improved YOLOv5
张昕楠 1李颖 1郝涌汀 1付靖凯 1于鹏1
作者信息
- 1. 沈阳理工大学机械工程学院,沈阳 110159
- 折叠
摘要
针对当前轴件表面缺陷种类繁多、缺陷形态复杂等原因导致的检测精度低,提出了一种改进的YOLOv5 的目标轴件表面缺陷检测方法.为解决在日常生产中经常出现轴件表面小目标缺陷被漏检、错检的问题,在原YOLOv5 基础上,添加一个新的小目标检测层,并将较浅特征图与深特征图拼接,使得整个网络更加关注小目标缺陷.同时为解决多目标缺陷和不完整轴件检测精度低与漏检,自建数据集中添加多目标缺陷与遮挡处理的轴件图像数据,经对比实验可知,改进的YOLOv5模型的检测性能优于FasterRCNN、SSD、原始YOLOv5 三种主流算法模型,测试的平均值精度分别高出7%、9%、4%.证明了该方法对轴件的表面缺陷检测精度更高,且对多目标缺陷与不完整的轴件的检测效果有显著提升.
Abstract
In view of the low detection accuracy caused by various types of surface defects of shaft parts and complex defect forms,an improved YOLOv5 surface defect detection method for target shaft parts was pro-posed in this paper.In order to solve the problem of often missing or failing to detect small target defects on the surface of shaft parts in daily production,a new small target detection layer has been added on the basis of the original YOLOv5,combining shallow feature maps with deep feature maps,making the entire network pay more attention to small target defects.At the same time,in order to solve multi-object defects and low detec-tion accuracy and missing detection of incomplete shaft parts,the shaft image data of multi-object defects and occlusion processing is added to the self-built data set in this paper.According to the comparison experi-ments,the detection performance of the improved YOLOV5 model is better than that of FasterRCNN,SSD and original YOLOV5,and the average accuracy of the test is higher than 7%,9%and 4%respectively.The conclusion proves that this method has higher accuracy in detecting surface defects of shaft components,and significantly improves the detection effect on multi-objective defects and incomplete shaft components.
关键词
轴件/缺陷检测/改进的YOLOV5/模型/小目标/多目标Key words
shaft/defect detection/improved YOLOV5 model/small target/multiple target引用本文复制引用
基金项目
辽宁省教育厅科学研究经费项目青年科技人才"育苗"项目(LG202031)
沈阳理工大学引进高层次人才科研支持计划(101014700081)
出版年
2024
NETL
NSTL
万方数据