基于结构重参数化的抗遮挡光场深度估计网络  

Anti-occlusion light field depth estimation network based on structural re-parameterization

作　　者：廖万 张倩[1] 高莹 王斌[1] 严涛 LIAO Wan;ZHANG Qian;GAO Ying;WANG Bin;YAN Tao(College of Information,Mechanical and Electrical Engineering,Shanghai Normal University,Shanghai 201418,China;School of Mechanical,Electrical and Information Engineering,Putian University,Putian 351100,Fujian,China)

机构地区：[1]上海师范大学信息与机电工程学院,上海201418 [2]莆田学院机电与信息工程学院,福建莆田351100

出　　处：《上海师范大学学报（自然科学版）》2023年第2期183-188,共6页Journal of Shanghai Normal University(Natural Sciences)

基　　金：福建省自然科学基金(2019J01816);莆田市科技局资助项目(2021G2001-8);福建省高等学校新世纪优秀人才项目(2018JYTRC(PU));莆田学院引进人才科研启动资助项目(2019003)。

摘　　要：针对光场深度估计网络结构中运算时间较长的问题,设计了一种能够被重参数化的多分支串联残差块结构(Res-DBLB),加快了网络运算速度,同时引入复合卷积块(RepConv)和卷积注意力模块(CBAM),优化网络性能.对于复杂的遮挡场景,利用深度图生成遮挡掩码,计算遮挡感知成本的构造函数,消除遮挡的负面影响.实验结果表明:与传统方法相比,该算法的均方误差和坏像素率更低,推理速度更快,同时在复杂遮挡场景中表现出较高的稳健性.A diverse branch residual block(Res-DBLB)which could be re-parameterized was designed to accelerate the network operation speed notably.Meanwhile a RepConv and convolutional block attention module(CBAM)were introduced to optimize the network performance.This was done to address the issue that the optical field depth estimation network structure bringed about long operation time by connecting a large number of multi-branch residual blocks in series to improve the network performance.The depth map was employed to produce occlusion masks and calculate the occlusion-aware cost constructor to eliminate the negative effects of occlusion in complex occlusion situations.The experimental results demonstrated that the method performed more robustly in complicated occlusion settings and had lower mean square error,less bad pixel,and faster inference compared to the conventional method.

关 键 词：光场深度估计 重参数化 卷积注意力模块(CBAM) 遮挡 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]