一种用于DSM局部缺失的深度学习修复算法  

A Deep Learning Repair Algorithm for DSM Local Missing

作　　者：官恺 刘智[1] 金飞[1] 韩佳容 芮杰[1] GUAN Kai;LIU Zhi;JIN Fei;HAN Jiarong;RUI Jie(Information Engineering University,Zhengzhou 450001,China;Xi’an Surveying and Mapping Station,Xi’an 710054,China)

机构地区：[1]信息工程大学,河南郑州450001 [2]西安测绘总站,陕西西安710054

出　　处：《测绘科学技术学报》2020年第3期281-286,共6页Journal of Geomatics Science and Technology

基　　金：国家自然科学基金项目(41601507)。

摘　　要：受地形、地物遮挡和反射率等影响,机载激光雷达获得的DSM数据值存在局部缺失。针对此问题,设计了一种基于U-Net改进的深度学习算法,成功地将深度学习图像修复方法应用于DSM修复。该算法通过在U-Net基础上结合部分卷积和注意力模块的方式能有效地减小修复误差,具有更好的鲁棒性。其中,部分卷积可增强不规则缺失边缘特征的提取能力;注意力模块能在通道和空间两个维度增加特征权重自适应学习机制。为了验证算法的有效性,采用多种方法对两个地区的DSM数据进行了实验。实验结果表明,改进后的算法有效地降低了修复误差,比传统U-Net网络方法降低了约30%;相比于传统方法误差更小,在缺失范围变化上具有更好的鲁棒性。Due to the influence of topography,ground objects shielding and reflectivity,DSM data obtained by airborne lidar is missing locally.In order to solve this problem,an improved deep learning algorithm based on U-Net was designed in this paper,and the deep learning image repair method is successfully applied to DSM reconstruction.The error is reduced effectively by the algorithm based on U-Net combined with partial convolution and attention module,and the algorithm has better robustness.Partial convolution can enhance the ability of anomalous edge feature extraction.Feature weight adaptive learning mechanism in channel and space dimensions can be increased by attention module.So as to verify the effectiveness of the algorithm,a variety of methods are used to conduct comparative experiments on DSM data from two regions.The results indicate that the improved algorithm effectively reduces the reconstruction error by about 30%compared with that of the based U-Net.The method based on deep learning has smaller error compared with the traditional method,and has better robustness in the change of the missing range.

关 键 词：DSM修复 深度学习 U型网络 部分卷积 注意力模块 

分 类 号：P237[天文地球—摄影测量与遥感]