基于组合零样本学习的接触网吊弦线缺陷识别

Defect identification for catenary dropper line based on compositional zero-shot learning

目前现场接触网吊弦缺陷图像严重不足,导致模型特征学习不充分,识别准确率难以得到有效提高,为此提出基于组合零样本学习的接触网吊弦线缺陷识别方法.采用以ResNet-50作为主干网络的视觉特征提取模块提取图像视觉特征;使用预训练的Word2Vec词向量对标签组合图中的节点特征进行初始化,并通过2层图卷积网络学习标签组合图中各节点之间的依赖关系,从而优化组合标签节点的语义特征,改善最终的识别效果;将提取到的视觉特征和优化后的组合标签节点的语义特征相对齐,构建相似度函数计算图像视觉特征与组合标签语义特征之间的相似度得分,并通过交叉熵损失完成图像组合标签的预测.仿真实验结果表明:所提方法对可见类样本的类平均检测准确率为93.5%,对不可见类样本的类平均检测准确率为86.5%.

Defect identification method for catenary dropper line based on compositional zero-shot learning was proposed,aiming at the problem of insufficient learning of model features and difficulty in effectively improving the recognition accuracy caused by the serious lack of image of catenary defects on site.The visual feature extraction module using ResNet-50 as the backbone network was used to extract image visual features.The pre-trained Word2Vec word vector was used to initialize the node features in the label combination graph.The dependence relationship between the nodes in the label combination graph was learned through the 2-layer graph convolutional networks,thereby optimizing the semantic features of the combined label nodes and improving the final recognition effect.The extracted visual features were matched with the semantic features of the optimized combined label nodes,and the similarity function was constructed to calculate the similarity score between the visual features of the image and the semantic features of the combined label.The prediction of the combined label was completed through the cross-entropy loss.The simulation results show that the proposed method has an average class detection accuracy of 93.5%for seen samples and 86.5%for unseen samples.