电子测量与仪器学报

周友行,孟高磊,赵文杰,易倩.钢板表面缺陷深度主动学习高效分类方法[J].电子测量与仪器学报,2022,36(2):23-31

钢板表面缺陷深度主动学习高效分类方法

Efficient deep active learning for steel plate surface defects classification

DOI：

英文关键词:surface defects active learning convolutional neural network global pooling

基金项目:国家自然科学基金（52175254，51775468）、湖南省教育厅科学研究项目（20A505）、湘潭大学研究生科研创新项目（XDCX2021B174）资助

作者	单位
周友行	1. 湘潭大学机械工程学院，2. 复杂轨迹加工工艺及装备教育部工程研究中心
孟高磊	1. 湘潭大学机械工程学院
赵文杰	1. 湘潭大学机械工程学院
易倩	1. 湘潭大学机械工程学院

Author	Institution
Zhou Youhang	1. School of Mechanical Engineering, Xiangtan University,2. Engineering Research Center of Complex Tracks Processing Technology and Equipment of Ministry of Education
Meng Gaolei	1. School of Mechanical Engineering, Xiangtan University
Zhao Wenjie	1. School of Mechanical Engineering, Xiangtan University
Yi Qian	1. School of Mechanical Engineering, Xiangtan University

摘要点击次数: 685

全文下载次数: 1085

中文摘要:

针对钢板表面缺陷图像分类传统深度学习算法中需要大量标签数据的问题,提出一种基于主动学习的高效分类方法。该方法包含一个轻量级的卷积神经网络和一个基于不确定性的主动学习样本筛选策略。神经网络采用简化的 convolutional base 进行特征提取,然后用全局池化层替换掉传统密集连接分类器中的隐藏层来减轻过拟合。为了更好的衡量模型对未标签图像样本所属类别的不确定性,首先将未标签图像样本传入到用标签图像样本训练好的模型,得到模型对每一个未标签样本关于标签的概率分布(probability distribution over classes,PDC),然后用此模型对标签样本进行预测并得到模型对每个标签的平均 PDC。将两类分布的 KL-divergence 值作为不确定性指标来筛选未标签图像进行人工标注。根据在 NEU-CLS 开源缺陷数据集上的对比实验,该方法可以通过 44%的标签数据实现 97%的准确率,极大降低标注成本。

英文摘要:

Aiming at the problem that traditional deep learning strategies used in steel plate surface defect images classification rely on abundant labeled samples. This paper proposes an efficient deep active learning method with a lightweight convolutional neural network and a novel uncertainty based active learning strategy. The network adopts a simplified convolutional base to do feature extraction, and replaces the hidden layer in the final densely connected classifier with global pooling layer to mitigate overfitting. To better measure model uncertainty about unlabeled image samples, this method first passes unlabeled images through the model trained by labeled image samples to obtain the probability distribution over classes ( PDC) for every unlabeled sample, then uses the same model to make predictions on the labeled samples to get an average PDC for every class. The KL-divergence value between these two kinds of distributions can be used as a new uncertainty measure to select unlabeled images for annotation. According to the experiments on NEUCLS dataset, the proposed method can reach 97% accuracy with 44% labeled data, which can reduce annotation cost greatly.

查看全文查看/发表评论下载PDF阅读器