摘要
针对工业上手机屏幕缺陷检测准确度不高,分割精度低等问题,提出一种基于无监督网络的方法,实现手机屏幕的缺陷分割。首先,通过无监督的卷积去噪自编码器构建多尺度特征的图像重构网络,实现从缺陷图像中重构出多层背景纹理图像。然后,将缺陷图像与多层背景重构图像分别进行减法运算,消除背景纹理的影响。最后,通过自适应阈值策略进行分割处理,再将多层分割结果进行融合,提升缺陷分割准确度。为提升重构性能,在网络中结合一种改进的损失函数进行训练。在分割处理中,根据残差图像的像素直方图是单峰的特点,使用三角法进行自适应阈值分割,提升分割精确度。经实验验证,通过本文方法进行手机屏幕的缺陷分割,分割精度达到90.30%,准确度和实时性满足工业要求,并具有实用性。
Based on an unsupervised network,a method for cell phone screen defect segmentation is proposed to solve the problem of low accuracy in cell phone screen defect detection.First,an image reconstruction network with multiscale features is constructed through an unsupervised convolutional denoising autoencoder,which reconstructs the multilayer background texture image from the defect image.Then,the defect and multilayer-reconstructed images are subtracted separately to eliminate the influence of the background texture.Finally,adaptive threshold strategy is used for segmentation and the segmentation results are fused to improve the accuracy of defect segmentation.To improve the reconstruction performance,an improved loss function is proposed to train the reconstruction network.Based on an image pixel histogram,the triangle method is used for global adaptive threshold segmentation to improve the segmentation accuracy.The experimental result shows that the proposed method can predict the cell phone screen defect area,reaching 90.30% accuracy.The accuracy and real time of the proposed method meet industrial requirements and it is practical.
作者
代朝东
许国良
毛骄
顾桐
雒江涛
Dai Chaodong;Xu Guoliang;Mao Jiao;Gu Tong;Luo Jiangtao(College of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing,400065,China;Institute of Electronic Information and Net work Engineering,Chongqing University of Posts and Telecommunications,Chongqing,400065,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第20期367-376,共10页
Laser & Optoelectronics Progress
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
机器视觉
缺陷分割
手机屏幕
无监督网络
图像重构网络
阈值分割
machine vision
defect segmentation
cell phone screen
unsupervised network
image reconstruction network
threshold segmentation
