浮空器主缆绳表面的小样本学习缺陷检测研究

Research on Surface Defect Detection of Aerostat Main Cable Based on Few-shot Metric Learning

针对浮空器主缆绳表面缺陷样本获取困难的特点,提出一种小样本度量学习方法来检测主缆绳的表面缺陷。小样本学习由特征编码和度量两模块组成,特征编码模块采用预训练的卷积神经网络在通用图像集上提取样本特征,度量模块用来度量与未知类别样本最相似的样本从而完成缺陷种类的分类;由于辅助数据和检测数据差异较大,将微调策略引入小样本学习方法。实验结果表明:本文方法的缺陷检测准确率达93.17%,相比传统机器学习和深度学习方法,准确率大幅提升;微调策略可进一步提升本文方法缆绳表面缺陷检测性能,准确率由93.17%提升至93.85%;在钢材缺陷分类数据集NEU上本文方法也可获得91.22%的准确率。

In view of the difficulty in obtaining the surface defect samples of the main cable of the aerostat,a few-shot learning method was proposed to detect the surface defects of the main cable.Few-shot metric learning consists of feature encoder module and metric module.The feature encoder module was used to extract features from a common used image dataset by a pre-training convolutional neural network(CNN),and the metric module was used to measure the samples that were the most similar to the unknown class samples,so as to complete the classification of defect.The experimental results show that the defect detection accuracy of this method is93.17%,which is greatly improved compared with traditional machine learning and deep learning methods.The fine tuning strategy can further improve the detection performance of cable surface defects by this method,and the accuracy can be improved from 93.17%to 93.85%.On the steel defect classification dataset NEU,the accuracy of this method can also be 91.22%.