融合边缘信息的立体匹配网络研究  

Research on Stereoscopic Matching Network Integrating Edge Information

作　　者：黄家熙 周晓聪[1] HUANG Jiaxi;ZHOU Xiaocong(School of Computer Science and Engineering,Sun Yat-Sen University,Guangzhou Guangdong 511400,China)

机构地区：[1]中山大学计算机学院,广东广州511400

出　　处：《信息与电脑》2024年第22期95-97,共3页Information & Computer

摘　　要：在计算机视觉领域,深度信息是一个重要的基础信息,在三维重建、自动驾驶等应用中发挥着至关重要的作用。立体匹配是获取深度信息的常用方法,神经网络的应用大大提高了深度信息的精度,但在遮挡或无纹理区域会造成精度下降。文章提出了一种融合边缘特征的立体匹配网络,引入一个边缘检测子网络,并通过联合学习视差和边缘特征,在边缘信息的引导下估计视差图。实验结果表明,该方法在Scene Flow和KITTI数据集上的精度有所提升,并且训练效率也得到了提高。In the field of computer vision,depth information is an important foundational information that plays a crucial role in applications such as 3D reconstruction and autonomous driving.Stereo matching is a common method for obtaining depth information,and the application of neural networks greatly improves the accuracy of depth.However,it still causes a decrease in accuracy in occluded or textureless areas.This article proposes a stereo matching network that integrates edge features,introduces an edge detection sub network,and jointly learns disparity and edge features,and estimates disparity maps under the guidance of edge information.The experimental results show that the accuracy of the method is improved on SceneFlow and KITTI datasets,and the training efficiency is also improved.

关 键 词：立体匹配 视差估计 边缘检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]