计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (23): 265-269.DOI: 10.3778/j.issn.1002-8331.1909-0299

• 工程与应用 • 上一篇    下一篇

结合随机子空间和级联残差网络的缺陷检测

金闳奇,陈新度,吴磊   

  1. 1.广东工业大学 机电工程学院,广州 510006
    2.广东工业大学 广东省计算机集成制造重点实验室,广州 510006
  • 出版日期:2020-12-01 发布日期:2020-11-30

Defect Detection Combining Random Subspace and Cascade Residual Network

JIN Hongqi, CHEN Xindu, WU Lei   

  1. 1.School of Electromechanical Engineering, Guangdong University of Technology, Guangzhou 510006, China
    2.Guangdong Provincial Key Laboratory of Computer Integrated Manufacturing, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2020-12-01 Published:2020-11-30

摘要:

针对基于深度学习的表面缺陷检测方法中的小样本问题,提出一种结合随机子空间和级联残差网络的缺陷检测方法(RSM-MTResNet)。该方法将缺陷数据分解为多个随机子空间,在每个子空间上构建残差网络,通过级联多个残差网络得到融合特征。在NEU表面缺陷数据集上进行实验,运用了混淆矩阵和[F1]值来评估模型性能。结果表明该方法的分类准确率为97.66%,比传统CNN方法的准确率高了14.5%,[F1]值均提高10.0%以上,这证明了该方法不仅能在一定程度解决小样本问题,同时能获得较高的识别性能。

关键词: 缺陷检测, 深度学习, 随机子空间, 级联残差网络, 小样本

Abstract:

Aiming at the small sample problem of surface defect detection method based on deep learning, a defect detection method combining random subspace and cascade residual network(RSM-MTResNet) is proposed. Firstly, the defect data is decomposed into multiple random subspaces, then the residual network is constructed on each subspace, and then the fusion features are obtained by cascades of multiple residual networks. Experiments are carried out on the NEU surface defect data set, and confusion matrix and F1 values are used to evaluate the model performance. The result shows that the classification accuracy of the method in this paper is 97.66%, 14.5% higher than that of the traditional CNN method, and F1 values are all improved by more than 10.0%, which proves that the method can not only solve the problem of small samples to a certain extent, but also achieve high recognition performance.

Key words: defect detection, deep learning, random subspace, cascade residual network, small sample