基于深度学习视野自选择的密集匹配网路  

Dense matching network based on deep learning vision self selection

作　　者：张昊 官恺 金飞 ZHANG Hao;GUAN Kai;JIN Fei

机构地区：[1]中国人民解放军战略支援部队信息工程大学,河南郑州450000 [2]西安测绘总站,陕西西安710000

出　　处：《智能城市》2023年第8期35-38,共4页Intelligent City

摘　　要：当前以PSMNet为代表的密集匹配网络结构在匹配代价计算过程中采用了多尺度技术,通过视差软回归的方式寻找最佳视差点,但多尺度参与计算会导致不同尺度相互干扰,视差不连续,区域的匹配误差增加。文章设计了不同视野的多支路网络结构AFSNet,针对每个像素点选出最佳视野。AFSNet能够有效减少视差非连续边缘误差,提升匹配精度,EPE和3PE在4个数据集上比采用多尺度的PSMNetS平均减少了1.79%和6.76%;以偏移损失和置信网络为选择依据,可以有效选择最佳尺度,实现合并精度优于任意支路精度;在测试的网络中,AFSNet模型泛化性整体优于其他参与测试的网络。The current dense matching network structure represented by PSMNet adopts multi-scale technology in the matching cost calculation process,searching for the optimal disparity point through disparity soft regression.However,multi-scale participation in the calculation will cause interference between different scales,discontinuous disparity,and increased matching error in the region.The article designs a multi branch network structure AFSNet with different fields of view,selecting the best field of view for each pixel.AFSNet can effectively reduce non continuous edge errors in disparity and improve matching accuracy.EPE and 3PE on four datasets have an average reduction of 1.79%and 6.76%compared to PSMNets using multi-scale methods;Based on offset loss and confidence network selection,the optimal scale can be effectively selected,achieving merging accuracy superior to any branch accuracy;In the tested network,the generalization of the AFSNet model is generally superior to other participating networks.

关 键 词：密集匹配 深度学习 视野自选择 多尺度 迁移学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]