LWMNet:一种视野宽广轻量级立体匹配算法  

LWMNet:ALightweight Stereo Matching Algorithm with a Wide Field of View

作　　者：何涛 韩涛[1] HE Tao;HAN Tao(School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学电气与信息工程学院,安徽淮南232000

出　　处：《兰州工业学院学报》2025年第1期96-102,共7页Journal of Lanzhou Institute of Technology

摘　　要：针对现有立体匹配算法精度低以及网络参数量过大的问题,提出了LWMNet(Light Weight Matching Network)。引入Fused-MBConv结构代替特征提取中部分Conv3*3,应用改进的残差模块LWC(Light Weight Concat)替换原有残差模块,设计了一种新的金字塔结构SPPC(Spatial Pyramid Pooling Concat)提取多尺度的空间特征信息,并增加跳跃连接,有效地融合不同尺度的特征信息,以降低误差。在SceneFlow公开的数据集进行实验,结果表明:与PSMNet方法相比,在参数数量下降了50.037%的同时,减少资源占用,误差降低了12.57%,提高了立体匹配的精度,提升了运行效率。In order to solve the problems of low accuracy of existing stereo matching algorithms and excessive network parameters,LWMNet(Light Weight Matching Network)is proposed.First,the Fused-MBConv structure is introduced to replace part of Conv3*3 in feature extraction;Secondly,the original residual module is replaced with the improved residual module LWC(Light Weight Concat);Finally,a new pyramid structure SPPC(Spatial Pyramid Pooling Concat)is designed to extract multi-scale spatial feature information,and jump connections are added to effectively integrate feature information of different scales to reduce errors.Experiments are conducted on the SceneFlow public data set.The experimental results show that compared with the PSMNet method,the number of parameters is reduced by 50.037%,while reducing resource usage,the error is reduced by 12.57%,improving the accuracy of stereo matching and improving operating efficiency.

关 键 词：深度估计 立体匹配 立体匹配网络 卷积神经网络 金字塔模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]