RGB-D图像多尺度融合的深度图像增强算法  

Depth Image Enhancement Algorithm with RGB‑D Image Multi‑scale Fusion

作　　者：赖水长 过洁[1] 李悫炜 郭延文[1] LAI Shuichang;GUO Jie;LI Quewei;GUO Yanwen(State Key Laboratory for Novel Software Technology,Nanjing University,Nanjing 210023,China)

机构地区：[1]南京大学计算机软件新技术国家重点实验室,南京210023

出　　处：《指挥信息系统与技术》2022年第3期78-84,共7页Command Information System and Technology

基　　金：国家自然科学基金(62032011和61972194);“十三五”全军共用信息系统装备预研课题(31511040202)资助项目。

摘　　要：随着消费级RGB及深度(RGB⁃D)图像相机技术的不断进步,场景的深度信息也在场景理解领域得以广泛应用。为了解决消费级相机获取的深度图像存在的分辨率低以及带有明显空洞与噪声等问题,提出了一种RGB⁃D图像多尺度融合的深度图像增强算法。RGB图像和深度图像分别经独立分支处理得到多尺度高维特征,而这2类特征在多个尺度上逐步融合,从而使空洞和噪声从粗尺度到细尺度逐步减少。此外,设计了一种混合多尺度损失函数,以确保全局物体结构清晰,并保留物体边界的深度不连续性。试验结果表明,该算法及其损失函数均可提高深度图像的质量,算法可行有效。With the continuous progress of consumer RGB and depth(RGB-D)images camera tech⁃nology,the depth information of scenes has been widely used in the field of scene understanding.In or⁃der to solve the problems of low resolution and obvious voids and noises in depth images acquired by consumer cameras,a depth image enhancement algorithm with RGB-D image multi-scale fusion is proposed.RGB image and depth image are processed by independent branch processing to obtain multi-scale high-dimensional features,and then two kinds of features are gradually fused at multiple scales,so that the voids and noises are gradually reduced from coarse scale to fine scale.In addition,a hybrid multi-scale loss function is designed to ensure that the global object structure is clear and the depth discontinuity of the object boundary is preserved.Experimental results show that both the algo⁃rithm and its loss function can improve the quality of depth image,and the algorithm is feasible and ef⁃fective.

关 键 词：深度图像增强 特征融合 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]