基于改进U^(2)-Net的透明件划痕检测方法  

Scratch Detection Method of Transparent Parts Based on Improved U^(2)-Net

作　　者：陈其浩 孙林[1] 张倩[1] CHEN Qi-hao;SUN Lin;ZHANG Qian(College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao 266590, China)

机构地区：[1]山东科技大学测绘与空间信息学院,青岛266590

出　　处：《科学技术与工程》2022年第2期620-627,共8页Science Technology and Engineering

基　　金：国家自然科学基金(41171408);山东省自然科学基金(ZR201702210379)。

摘　　要：为了满足透明件表面质量和市场竞争的需求,实现产品表面缺陷的自动化检测至关重要。针对透明件表面划痕快速检测问题,提出了一种基于改进U^(2)-Net的缺陷检测方法。首先,直接应用U^(2)-Net网络进行透明件表面划痕检测的数据集准备、网络搭建、损失函数、评估指标;其次,初始化网络进行训练,分析产生误检漏检及低效的原因;最后,优化损失函数,加入正则化技术,并给出在输入数据前加入Mosaic数据增强,解码阶段融入深层可分离卷积以及加入Attention机制的改进方案。结果表明:本文提出的改进方案能够有效分割出不同情况下的划痕,准确率达到0.987,漏检率为0.006,并在检测速度上有19%的提升。可见改进U^(2)-Net的透明件划痕检测方法能够很好满足工业流水线准确检测缺陷的实际需求。In order to meet the requirements of surface quality and market competition for transparent parts,it is of great importance to realize the automatic detection of surface defects in products.A defect detection method based on improved U^(2)-net was proposed according to the problems in rapid detection of surface scratch on transparent parts.First,dataset preparation,network construction,loss function and evaluation index in direct application of U^(2)-net to perform detection of surface scratch in transparent parts was discussed.Second,the network for training was initialized,and the causes of false and missing detection as well as for low efficiency was analyzed.Finally,the loss function was optimized,the regularization technique and Mosaic data enhancement was added before inputting data.The Depthwise Separable Convolution was integrated at the decode stage,and the Attention mechanism was added.The results show as follows.The scratches in different situations can effectively be segmented by the improvement scheme proposed.The accuracy rate is 0.987,the missing detection rate is 0.006,and the detection speed is increased by 19%.The actual demand of accurate defect detection in industrial production lines can be fairly satisfy by the scratch detection method for transparent parts based on improved U^(2)-net.

关 键 词：透明件 划痕检测 神经网络 U^(2)-Net 语义分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]