摘要
为了实现焊缝缺陷的准确分类,提出一种多特征融合的焊缝图像多标签分类算法。首先,通过残差神经网络(ResNet-50)提取图像的特征信息,根据得到的特征图构建图结构,提出关联度引导邻域传播(RDGNP)算法优化图结构;然后,使用图卷积神经网络(GCN)提取图结构的特征信息,并设计特征融合模块实现图像特征和图结构特征的结合;最后,得到多标签分类结果。实验结果表明:文中算法能够有效地实现焊缝缺陷的多标签分类,在X射线焊缝缺陷数据集上的性能有明显提升。
In order to achieve accurate classification of welding defects,a weld image multi-label classification algorithm based on multi-feature fusion is proposed.Firstly,feature information of images is extracted by a residual neural network(ResNet-50),and the graph structure is constructed based on the obtained feature maps.An algorithm named relation degree guided neighborhood propagation(RDGNP)is proposed to optimize the graph structure.Then,the feature information of the graph structure is extracted using graph convolutional neural network(GCN),and a feature fusion module is designed to achieve the combination of image features and graph structure features.Finally,multi-label classification results are obtained.Experimental results show that the proposed method can effectively realize the multi-label classification of welding defects,and its performance on the X-ray welding defects dataset is significantly improved.
作者
牛顿
林宁
林振超
黄凯
王合佳
郑力新
NIU Dun;LIN Ning;LIN Zhenchao;HUANG Kai;WANG Hejia;ZHENG Lixin(College of Engineering,Huaqiao University,Quanzhou 362021,China;Fujian Special Equipment Inspection and Research Institute,Quanzhou 362021,China)
出处
《华侨大学学报(自然科学版)》
CAS
2024年第4期514-523,共10页
Journal of Huaqiao University(Natural Science)
基金
福建省科技计划项目(2020Y0039)
福建省泉州市科技计划项目(2020C042R)。
关键词
多标签分类
全局相关性
图像特征
图结构特征
特征融合
multi-label classification
global relevance
image feature
graph structure feature
feature fusion