摘要
针对酒瓶盖瑕疵会影响产品质量的问题,提出了一种酒瓶盖瑕疵YOLOv3-MRHA检测算法,基于YOLOv3算法,对其主干网络和特征提取层进行改进。为减少主干网络特征丢失现象,提出了多级特征融合(multilevel feature fusion,MFF)模块;为提高检测的准确率,增加了尺度为104×104的特征层,并构造了一种增强特征信息的残差特征增强(residual feature enhancement,RFE)模块;为提高深层特征层的检测能力,引入了空洞卷积,使浅层信息向下融合,在特征提取层使用通道注意力机制。结果表明,所提YOLOv3-MRHA算法的检测精度比YOLOv3算法提高近6%,可有效地提高瑕疵检测的准确率,满足工业质检的要求。
Aiming at the problem that wine bottle cap defect would affect product quality,a wine bottle cap defect detection algorithm,YOLOv3-MRHA,was proposed.Based on YOLOv3 algorithm,its backbone and feature extraction layer were improved.Firstly,multilevel feature fusion(MFF)module was proposed to reduce the feature loss in backbone.Secondly,in order to improve the accuracy of detection,the scale was increased to 104×104,and the residual feature enhancement(RFE)module was used for enhancing feature information.Finally,in order to improve the detection ability of deep feature layer,the dilated convolution was introduced to fuse the shallow information downward,and the channel attention mechanism was used in the feature extraction layer.The result shows that the detection accuracy of YOLOv3-MRHA algorithm is nearly 6%higher than that of YOLOv3 algorithm.The algorithm effectively improves the accuracy of detection and meets the requirements of industrial quality inspection.
作者
李玉洁
韩进
刘恩爽
LI Yujie;HAN Jin;LIU Enshuang(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao,Shandong 266590,China)
出处
《中国科技论文》
CAS
北大核心
2022年第11期1236-1244,共9页
China Sciencepaper
基金
山东省自然科学基金资助项目(ZR2020KE023,ZR2021MD057)。
关键词
酒瓶盖瑕疵检测
多级特征融合
残差特征增强
空洞卷积
通道注意力机制
wine bottle cap defect detection
multilevel feature fusion(MFF)
residual feature enhancement(RFE)
dilated convolution
channel attention mechanism
