基于改进YOLOv5的陶瓷表面缺陷检测算法
Ceramic Surface Defect Detection Algorithm Based on Improved YOLOv5
潘金晶 1曾成 1张晶 1李再勇 1耿雪娜2
作者信息
- 1. 长园视觉科技(珠海)有限公司,广东 珠海 519085
- 2. 长春理工大学 计算机科学技术学院,吉林 长春 130013
- 折叠
摘要
提出一种陶瓷表面缺陷检测算法YOLOv5-G.该算法在YOLOv5 框架的基础上,将全局注意力机制(GAM)引入主干和颈部网络中,该机制能够在减少信息弥散的情况下放大全局交互特征,增强了网络的特征表达能力;使用α-CIoU作为改进算法的边界框回归损失函数,自适应地向上加权高IoU对象的损失和梯度,使得模型可以更加关注IoU高的目标,从而帮助提高定位精度.在工业相机成像的陶瓷表面缺陷数据集上进行测试,结果表明,与YOLOv5 模型相比,基于α-CIoU的YOLOv5 模型的平均精度均值(mAP50)和召回率(Recall)分别提升了 2.3%、4.6%;改进算法的平均精度均值(mAP50)、精确率(Precision)、召回率(Recall)分别提升了3.9%、1.7%、5.1%.
Abstract
It propose a ceramic surface defect detection algorithm YOLOv5-G.This algorithm introduces the Global Attention Mechanism(GAM)into the backbone and neck networks based on the YOLOv5 framework.This mechanism can amplify the global interaction features while reducing the information dispersion,and enhances the feature expression capability of the network.Taking α-CIoU as the bounding box regression loss function of the improved algorithm,the loss and gradient of high IoU objects are adaptively weighted upward,so that the model can pay more attention to the targets with high IoU,thus helping to improve the positioning accuracy.The test is carried out on the ceramic surface defect data set of the industrial camera imaging,and the results show that compared with the YOLOv5 model,the Mean Average Precision(mAP50)and Recall rate of the YOLOv5-G model based on α-CIoU are increased by 2.3%and 4.6%respectively.The Mean Average Precision(mAP50),Precision and Recall of the improved algorithm are all significantly improved by 3.9%,1.7%and 5.1%respectively.
关键词
陶瓷表面缺陷/全局注意力机制/α-IoU/YOLOv5Key words
ceramic surface defect/GAM/α-IoU/YOLOv5引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据