摘要
为了实现换向器表面缺陷自动检测,针对实际缺陷样本数量少、标注成本高等问题,设计一种基于特征空间的无监督换向器表面缺陷检测方法。利用预训练的EfficientNet网络提高对于特征的提取能力;同时提出一种平衡利用不同深度特征信息建立异常图的方法,加强对于缺陷轮廓的识别精度;通过对每个像素分别建立高斯分布,实现像素级缺陷分割。实验表明,提出的方法对缺陷样本的识别精度达到98.7%,对缺陷区域的分割精度达到98.0%,具有较好的缺陷检测能力,在工业应用上具有研究价值和实践意义。
In order to realize the automatic detection of the surface defects of the commutator,an unsupervised commutator surface defect detection method based on feature space is designed for the problem that the actual number of defect samples is small,the labeling cost is high.Improve the feature extraction ability by using the pre-trained EfficientNet network;propose a method to balance the use of different depth feature information to build the anomaly map to strengthen the recognition accuracy of the defect contour;realize pixel-level defects segmentation by establishing a Gaussian distribution for each pixel.Experiments show that the proposed method has a recognition accuracy of 98.7%for defect samples and a segmentation accuracy of 98.0%for defect regions.It has good defect detection capabilities and has research value and practical significance in industrial applications.
作者
董泽宇
叶芸
姚剑敏
严群
李德财
林坚普
DONG Zeyu;YE Yun;YAO Jianmin;YAN Qun;LI Decai;LIN Jianpu(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China;Jinjiang RichSence Electronic Technology Company Limited,Jinjiang 362200,China;College of Advanced Manufacturing,Fuzhou University,Jinjiang 362200,China)
出处
《电视技术》
2021年第12期104-109,119,共7页
Video Engineering
基金
国家重点研发计划课题(No.2017YFB0404604)
福建光电信息科学与技术创新实验室重大项目(No.2020ZZ112)
关键词
换向器
缺陷检测
缺陷分割
无监督学习
commutator
defect detection
defect segmentation
unsupervised learning
