摘要
传统单视图三维重建方法不能满足人工关节磨屑三维信息提取要求,为提高重建精度,提出一种基于SA-UNet网络的人工关节磨屑显微单视图深度估计方法,实现从单一视图下磨屑三维信息的快速获取。该方法首先构建一个融合自注意力机制的U-Net深度估计网络,然后使用光学显微镜和激光共聚焦显微镜分别收集磨屑的二维图像和深度图,再通过图像语义分割提取目标磨屑区域,消除图像背景的干扰,最后利用透视变换使二维图像和深度图相对应,获得训练样本。采用该方法对TC4材料磨屑显微单视图进行深度估计,以激光共聚焦显微镜的三维信息作为参考。结果表明,该方法预测深度的平均相对误差为7.35%,均方根相对误差为3.93,效果优于U-Net、BTS和ACAN。
Traditional single-view 3D reconstruction method cannot meet the requirement of 3D feature extraction of artificial joint wear particles.To improve the reconstruction accuracy,a SA-UNet network was proposed to estimate the depth information from a micro-single-view image of artificial joint wear particles.For this method,a U-Net depth estimation network was firstly developed by integrating with a self-attention module,and then 2D and 3D images of artificial joint wear particles were respectively collected using a conventional optical microscope and a laser scanning confocal microscope(LSCM),meanwhile,the target particle regions were segmented to eliminate the influence of image background,and finally the perspective transformation was used to build the relationships between the 2D and 3D images,so as to obtain sample dataset.The microscopic images of TC4 wear particles were used to test the network model,and the 3D information of laser scanning confocal microscope(LSCM)was used as a reference.The results show that the square relative error and the root mean square error of the prediction depth obtained by the proposed method are 7.35%and 3.93,respectively,which are both better than that of U-Net,BTS and ACAN.
作者
伍锐斌
彭业萍
曹广忠
王松
曹树鹏
WU Ruibin;PENG Yeping;CAO Guangzhong;WANG Song;CAO Shupeng(College of Mechatronics and Control Engineering,Shenzhen University,Shenzhen Guangdong 518060,China;Guangdong Key Laboratory of Electromagnetic Control and Intelligent Robots,Shenzhen Guangdong 518060,China;Biomechanics and Biotechnology Lab,Research Institute of Tsinghua University in Shenzhen,Shenzhen Guangdong 518057,China)
出处
《润滑与密封》
CAS
CSCD
北大核心
2022年第7期40-48,共9页
Lubrication Engineering
基金
国家自然科学基金项目(51905351,U1813212)
深圳市科技计划项目(JCYJ20190808113413430,JCYJ20190807144001746)。
关键词
人工关节
磨屑
显微图像
单视图
深度估计
artificial joint
wear particle
microscopic image
single-view
depth estimation
