| 无监督密集匹配特征提取网络性能分析 |
| |
| 引用本文： | 金飞,官恺,刘智,韩佳容,芮杰,李庆高.无监督密集匹配特征提取网络性能分析[J].测绘学报,2022,51(3):426-436. |
| |
| 作者姓名： | 金飞 官恺 刘智 韩佳容 芮杰 李庆高 |
| |
| 作者单位： | 1. 信息工程大学地理空间信息学院, 河南 郑州 450001;2. 西安测绘总站, 陕西 西安 710054 |
| |
| 基金项目： | 国家自然科学基金(41601507)~~； |
| |
| 摘 要： | 随着人工智能的发展，基于深度学习的有监督密集匹配方法在虚拟、室内及驾驶等近景数据集上取得了不错的表现。针对航空影像密集匹配标签数据获取困难的问题，本文在无监督密集匹配框架下，借鉴多个有监督网络结构，分别在航空影像数据集和作为参照的近景数据集上测试了匹配精度，实现了网络结构模块与精度关系的定性分析，为进一步探索深度学习在测绘领域的实用化提供了重要的参考。试验在相同损失函数条件下，分别采用DispNetS、DispNetC、iResNet、GCNet、PSMNetB及PSMNetS网络结构进行测试。经分析，得出如下结论：①测试的网络结构中，PSMNetS在航空影像数据集和近景数据集上表现稳定，且精度最高，训练整体耗时少，具有实用化的潜力；②在监督方法中效果更好的网络结构在无监督方法中效果不一定更好，其精度不仅取决于网络自身的匹配能力，同时也依赖于网络与损失函数的兼容性；③孪生网络模块、相关信息融合模块、金字塔池化模块和堆叠沙漏模块与无监督损失函数兼容性良好，可提升网络精度，而iResNet的图像重构迭代精化模块与重构损失函数重复约束，会产生“负优化”的作用。
|
| 关 键 词： | 密集匹配 深度学习 无监督 特征提取 航空影像 |
| 收稿时间： | 2020-10-14 |
| 修稿时间： | 2021-11-21 |
A performing analysis of unsupervised dense matching feature extraction networks |
| |
| JIN Fe,GUAN Kai,LIU Zhi,HAN Jiarong,RUI Jie,LI Qinggao.A performing analysis of unsupervised dense matching feature extraction networks[J].Acta Geodaetica et Cartographica Sinica,2022,51(3):426-436. |
| |
| Authors: | JIN Fe GUAN Kai LIU Zhi HAN Jiarong RUI Jie LI Qinggao |
| |
| Institution: | 1. Institute of Geospatial Information, Information Engineering University, Zhengzhou 450001, China;2. The Technical Division of Surveying and Mapping of Xi'an, Xi'an 710054, China |
| |
| Abstract: | With the development of artificial intelligence, supervised dense matching method based on deep learning has achieved good performance in virtual, indoor and driving data sets. In view of the difficulty in obtaining aerial image dense matching tag data, we use unsupervised dense matching framework for reference, and test the matching accuracy on aerial image data set and referential close range data set respectively, and realize the qualitative analysis of the relationship between network structure module and precision, which provides a further exploration of the practical application of deep learning in the field of surveying and mapping, and has important reference value. Under the same loss function condition, DispNetS, DispNetC, iResNet, GCNet, PSMNetB and PSMNetS network structures are used to test. Through analysis, the following conclusions are obtained:① Among the tested network structures, PSMNetS has the highest accuracy in aerial image data set and close range data set, and has the potential of practical application; ② The network with better performance in the supervised method may not have better performance in the unsupervised method. Its accuracy depends not only on the matching ability of the network itself, but also on the compatibility between the network and the loss function; ③ The twin network module, related information fusion module, pyramid pooling module and stacked hourglass module have good compatibility with unsupervised loss function, which can improve the network accuracy, while iResNet's image reconstruction iterative refinement module and reconstruction loss function repeat constraints, which will produce a "negative optimization" effect. |
| |
| Keywords: | |
| 本文献已被 万方数据 等数据库收录！ |
| 点击此处可从《测绘学报》浏览原始摘要信息 |
| 点击此处可从《测绘学报》下载免费的PDF全文 |
|
