基于改进YOLOv5的热轧带钢缺陷检测
Improved YOLOv5-Based Defect Detection for Hot-Rolled Steel Strips
李顺 1杨颖1
作者信息
- 1. 广西大学计算机与电子信息学院,广西 南宁 530004
- 折叠
摘要
针对热轧带钢的缺陷检测,存在目标尺寸过小、特征不清晰和错测漏检等问题,提出了一种基于改进YOLOv5 的热轧带钢缺陷检测方法.首先基于K-means++算法在聚合网络上加入了一种超参数锚框算法,提高了锚框的准确率.其次重新设计了新的特征提取模块,增加了检测的尺度并且加入了非线性卷积模块强化了目标缺陷的语义信息.最后针对置信度损失函数,使用较为平滑的相对熵来取代交叉熵,提高了模型收敛时的稳定性.与基准算法的实验结果显示,使用改进YOLOv5 的平均检测精度比原版YOLOv5 提高了8.2%,泛化能力更强,检测速度更快,错误和漏检率更低.
Abstract
For the defect detection of hot-rolled strip steel,problems include too small target size,unclear fea-tures,and wrong measurement and omission detection.This paper proposes a defect detection method for hot-rolled strip steel based on improved YOLOv5.Firstly,a hyperparametric anchor frame algorithm is added to the aggregation network based on the K-means++algorithm to improve the accuracy of the anchor frame.Secondly,a new feature ex-traction module is redesigned to increase the detection scale,and a nonlinear convolution module is added to enhance the semantic information of the target defects.Finally,for the confidence loss function,a smoother relative entropy is used instead of cross-entropy to improve the stability of the model when converging.Experimental results with the benchmark algorithm show that the average detection accuracy using the improved YOLOv5 is 8.2%better than the o-riginal YOLOv5,with more vital generalization ability,faster detection,and lower error and miss detection rates.
关键词
神经网络/缺陷检测/超参数锚框算法/特征提取模块/相对熵Key words
Neural network/Defect detection/Hyperparametric anchor frame algorithm/Feature extraction mod-ule/L scatter引用本文复制引用
基金项目
国家自然科学基金(61763002)
广西创新驱动发展专项(桂科AA20302002)
广西科技基地和人才专项(桂科AD21076002)
出版年
2024
NETL
NSTL
万方数据