基于改进轻量级SE-Yolov4的热轧钢表面缺陷检测方法
Surface Defect Detection Method of Hot Rolled Steel Based on Improved Lightweight SE-Yolov4 Algorithm
黄晓红 1李静 2董诗琪 2王云阁3
作者信息
- 1. 华北理工大学 人工智能学院,河北 唐山 063210;华北理工大学 河北省工业智能感知重点实验室,河北 唐山 063210
- 2. 华北理工大学 人工智能学院,河北 唐山 063210
- 3. 唐山钢铁股份有限公司,河北 唐山 063016
- 折叠
摘要
针对传统热轧钢表面缺陷检测存在的检测精度较低、检测速度较慢,传统机器学习检测存在检测速度慢、鲁棒性差等问题,提出一种基于改进轻量级SE-Yolov4热轧钢表面缺陷检测方法.Yolov4 主干特征提取网络CSPDarknet53 的每一层残差网络中嵌入SENet结构,构成SE-Yolov4网络,有选择地聚集有效信息;同时在主干特征网络输出不同特征信息后和空间池化金字塔前后增加卷积层数,网络结构复杂化;SE-Yolov4算法中嵌入轻量化MobileNet v3 结构,减少模型参数量,提高检测速度.实验结果表明:该改进算法在测试集中的mAP值达到93.02%,较Yolov4算法检测精度提升7.2%,检测速度提升近3 倍.
Abstract
Aiming at the problems of low detection accuracy and slow detection speed of traditional hot rolled steel surface defect detection,and the problems of slow detection speed and poor robustness of traditional machine learning detection,a surface defect detection method based on improved lightweight SE-Yolov4 hot rolled steel is proposed.The SE-Yolov4 network is composed of SE-Yolov4 network,which is embedded in the residual network of each layer of the Yolov4 backbone feature extraction network CSPDarknet53 to selectively gather effective information.At the same time,after the backbone feature network outputs different feature information before and after the Spatial Pooling Pyramid,the number of convolution layers is increased,and the network structure is complex;SE-Yolov4 algorithm embeds lightweight mobileNet v3 structure to reduce the amount of model parameters and improve the detection speed.The experimental results show that the mAP value of the improved algorithm in the testset reaches 93.02%,which improves the detection accuracy by 7.2%and the detection speed by nearly three times compared with Yolov4 algorithm.
关键词
热轧钢/Yolov4/SENet/表面缺陷检测/卷积神经网络/MobileNet/v3Key words
hot rolled steel/Yolov4/SENet/surface defect detection/convolutional neural network/MobileNet v3引用本文复制引用
基金项目
河北省高等学校科学技术研究项目(ZD2020152)
华北理工大学技术转移基金平台及推广项目(TG2018004)
科技基础研究项目(JQN2019006)
出版年
2024
NETL
NSTL
维普
万方数据