基于通道注意力机制的室内场景深度图补全  

Depth map completion for indoor scenes based on the channel attention mechanism

作　　者：任瀚实 周志宇[1] 孙树森[2] REN Hanshi;ZHOU Zhiyu;SUN Shusen(School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China;School of Computer Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学信息科学与工程学院,杭州310018 [2]浙江理工大学计算机科学与技术学院,杭州310018

出　　处：《浙江理工大学学报（自然科学版）》2023年第3期344-352,共9页Journal of Zhejiang Sci-Tech University(Natural Sciences)

摘　　要：在深度相机获取的室内场景深度图中,部分像素点缺失深度信息;为补全深度信息,设计了端对端的场景深度图补全网络,在此基础上提出了一种基于通道注意力机制的室内场景深度图补全方法。该方法将场景彩色图与缺失部分深度信息的场景深度图作为场景深度图补全网络的输入,首先提取场景彩色图和深度图的联合特征,并根据通道注意力机制将提取到的联合特征进行解码,得到初始预测深度图;然后借助非局部区域上的传播算法逐步优化场景深度的预测信息,得到完整的场景深度图;最后在Matterport3D等数据集上进行实验,并将该方法与典型方法进行比较分析。实验结果表明,该方法融合了场景彩色图和深度图特征信息,通过注意力机制提高了深度图补全网络的性能,有效补全了深度相机拍摄室内场景时缺失的深度信息。When the depth camera scans the indoor scene,the depth information of some pixels is missing.To solve this problem,we proposed a depth map completion method for indoor scenes based on a channel attention mechani sm and an endto-end scene depth map completion network.With the scene color map and incomplete scene depth map as input,the network first extracted the scene color map features and depth map features,and decoded the extracted combined features based on the channel attention mechanism to obtain the initially predicted depth map.Then the predicted scene depth information was gradually optimized with the help of non-local spatial propagation algorithm to finally obtain the complete scene depth map.Finally,the proposed method was compared with typical methods using datasets such as Matterport3D.Experimental results show that this method,integrating the feature information of the scene color map and the depth map,improves the performance of depth map completion network through the attention mechanism,and effectively complements the lack of depth information when shooting indoor scenes by depth camera.

关 键 词：深度图 深度补全 深度学习 注意力机制 室内场景 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]