摘要
在手机屏幕工业化生产过程中,缺陷检测的好坏直接影响手机屏幕的合格率。少量的缺陷样本不足以完成数据驱动的分割网络的训练,因此如何利用少量的缺陷图像完成缺陷分割成为关键问题。该文针对此问题提出一种基于协同注意力的小样本手机屏幕缺陷分割网络(Co-ASNet)。该网络使用交叉注意力块在特征提取时获取更加丰富的上下文缺陷特征信息,同时引入了协同注意力的方式来加强支持图像与查询图像相同缺陷目标之间的特征信息交互,增强缺陷特征表示,另外,使用了改进的联合损失函数来完成网络的训练。该文采用手机屏幕缺陷数据集进行实验,实验结果表明,Co-ASNet能够使用少量的缺陷样本完成良好的缺陷分割效果。
In the commercial process of mobile phone screens,the quality of defect detection affects directly the qualified rate of mobile phone screens.A few defect samples are not enough to complete the training of datadriven segmentation networks,so how to use a few defect samples to complete the defect segmentation is a key problem.In view of this problem,a Co-Attention Segmentation Network(Co-ASNet)is proposed.This network uses Criss-cross attention blocks to capture contextual defect feature information during feature extraction.At the same time,the Co-attention method is applied to enhance the defect feature information interaction between the same defect target in the support image and query image,and then the defect feature representation is reinforced.Also,the improved joint loss function is used to complete the network training.The experimental results show that Co-ASNet can use a few defect samples to achieve an excellent effect of defect segmentation.
作者
许国良
毛骄
XU Guoliang;MAO Jiao(Institute of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Electronic Information and Networking Research Institute,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)
出处
《电子与信息学报》
EI
CSCD
北大核心
2022年第4期1476-1483,共8页
Journal of Electronics & Information Technology
基金
重庆市技术创新与应用示范(产业类重点研发)项目(cstc2018jszx-cyzdX0124)。
关键词
手机屏幕缺陷
小样本分割
协同注意力
交叉注意力
Mobile phone screen defects
Few-shot segmentation
Co-attention
Criss-cross attention
