基于改进YOLOv5的微小柱状零件表面缺陷检测系统
Surface Defect Detection System for Tiny Columnar Parts Based on Improved YOLOv5
朱永利 1刘斌 2曹微 1孙佳晨 1李陈 1王国海 1王宇 3张昌华2
作者信息
- 1. 中核北方核燃料元件有限公司
- 2. 电子科技大学机械与电气工程学院
- 3. 西华大学机械工程学院
- 折叠
摘要
针对"细小曲面"的零件表面缺陷检测技术,目前尚不多见.以一种直径在mm级、表面缺陷尺寸在μm级的柱状金属零件为研究对象,设计了一套微小柱状零件表面缺陷检测系统.首先分析了检测对象的技术特征与"全、细、微、多、高"检测难点.然后针对"全、细、微"的技术要求,给出了图像采集装置的设计方案,实现了在微小零件柱状曲面上的全表面逐一成像.随后,针对"多、高"两大技术难点,提出了一种基于YOLOv5 与注意力机制的检测算法,给出了实验数据集的准备方法并开展了和现有方法的多组对比实验与消融实验,实验结果表明所提算法能够辨识零件表面多种缺陷.相关研究已经应用于工业现场,显著提高了微小柱状零件表面缺陷的检测效率和准确率,对其他类似零件的表面缺陷检测也有借鉴价值.
Abstract
There are not many techniques for detecting surface defects on"fine surfaces".Taking a cylindrical metal part with a diameter of millimeters and a surface defect size of micrometers as a research object,a surface defect detection system for tiny cylindrical parts was designed.First,the technical features of detecting objects and the challenging aspects of"comprehen-sive,detailed,subtle,diverse,and high"detection were analyzed.In accordance with the technical requirements of"full,fine,and micro",a design scheme for the image acquisition device was proposed.This implementation achieved continuous full surface ima-ging on the cylindrical surface of small part.Additionally,a detection algorithm was brought out,based on YOLOv5 and an atten-tion mechanism,in order to tackle the two major technical obstacles of"multiple,high".A method for preparing the experimental dataset was presented,followed by several comparative and ablation experiments that were carried out using existing methods.The experiments demonstrate that the proposed algorithm can effectively detect different defects present on the surface of the parts.The research is implemented in industrial application,resulting in a notable enhancement of effectiveness and precision in surface de-fect detection on small cylindrical parts.The related research is also valuable for the detection of surface defects on other similar parts.
关键词
微小柱状零件/表面缺陷检测/机器视觉/目标检测/选择注意力机制/随机注意力机制Key words
tiny columnar parts/surface defect detection/machine vision/target detection/selective attention mechanism/stochas-tic attention mechanism引用本文复制引用
出版年
2024
NETL
NSTL
万方数据