石板材表面缺陷检测的无监督学习方法  

Unsupervised learning method for surface defect detection of slate materials

作　　者：王延春秋 葛泉波 刘华平[3] WANG Yanchunqiu;GE Quanbo;LIU Huaping(Logistics Engineering College,Shanghai Maritime University,Shanghai 201306,China;School of Automation,Nanjing University of Information Science&Technology,Nanjing 210044,China;Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China)

机构地区：[1]上海海事大学物流工程学院,上海201306 [2]南京信息工程大学自动化学院,江苏南京210044 [3]清华大学计算机科学与技术系,北京100084

出　　处：《智能系统学报》2023年第6期1344-1351,共8页CAAI Transactions on Intelligent Systems

摘　　要：石板材表面缺陷检测是一项具有挑战性的任务,尤其对于边缘磕碰、裂缝等细微缺陷,检测难度大。此外,冗余特征的存在会影响训练效果,多尺度特征学习需要进行多维计算,计算复杂度高。针对上述问题,本文提出一种基于无监督学习的石板材表面缺陷检测方法,它能够有效解决该任务存在的问题。首先,对预训练网络提取到的图像多尺度特征,采用正半交嵌入特征降维方式减少冗余特征的影响。然后,通过多过程特征学习降低计算中的时间复杂度,提高训练效率。最后,根据训练模型得出待测图像的局部马氏距离,提高检测性能。相关实验表明,本方法在石板材数据集上的结果优于当前几种先进方法,同时在石板材表面缺陷检测和定位方面证明本方法的有效性。The surface defect detection of slate is a challenging task,particularly for small defects like edge bumps and cracks.In addition,the existence of redundant features will influence the training effect,multiscale feature learning will require multidimensional calculation,and the calculation complexity will increase.Considering the above problems,this paper proposes a method for detecting surface defects in slate materials based on unsupervised learning to solve the problem in this task effectively.First,the semiorthogonal embedding feature dimension reduction is used on the multiscale features of the image extracted using a pretraining network to reduce the effect of redundant features.Further,the time complexity of calculation is reduced through multiprocess feature learning,increasing training efficiency.Finally,the local Markov distance of the image to be measured is obtained in accordance with the training model to improve the detection performance.Relevant experiments show that the results of this method on the slate data set are superior to several advanced methods at present,and the effectiveness of this method is verified by detecting and locating surface defects in slate materials.

关 键 词：石板材 表面缺陷 缺陷检测 无监督学习 多尺度特征 半正交嵌入特征 特征学习 马氏距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]