利用改进CNN的无人机遥感影像特征描述符的学习方法  

A Learning Method Using UAV Remote Sensing Image Feature Descriptor with Improved CNN

作　　者：王瑞瑞[1] 冯伍法[1] 张艳[1] 王涛[1] 刘冰[1] 

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《测绘科学与工程》2017年第5期51-56,共6页Geomatics Science and Engineering

基　　金：河南省科技攻关计划基金资助项目（152102210014）.

摘　　要：由于人工设计的特征描述符往往通过增加描述符构造的复杂度提高鲁棒性，导致描述符的实时性降低。为了提高描述符鲁棒性与实时性的兼容能力，本文对基于卷积神经网络（CNN）的特征描述符学习方法进行改进。通过交换三重样本间基准样本挖掘“难样本对”，解决了现有基于CNN特征描述符学习方法中“难样本对”挖掘计算成本较高的问题。同时，结合无人机影像的特点，采用多分辨率训练样本代替原有单一分辨率训练样本，使基于CNN的特征描述符学习方法更好地适用于无人机影像匹配。试验表明，本文改进的基于CNN的特征描述符学习方法获得的描述符可以较好地兼顾鲁棒性和实时性，且对无人机影像匹配有较好的应用效果。Since the robustness of artificially-designed feature descriptor is often improved by increasing the complexity of descriptor construction, the real-time performance of descriptor is reduced. In order to improve the compatibility of descriptor ro- bustness and real-time performance, an improved learning method of feature descriptor based on convolutional neural networks （CNN） is put forward in this paper. According to the method, ＂hard samples＂ are extravated by exchanging the reference sam- ples of triple samples, and thus the high computational cost of current method based on CNN is brought down. At the same time, according to characteristics of the UAV image, the original single resolution training samples are replaced by muhi-resolution training samples, making the CNN-based feature descriptor learning method more suitable for UAV image matching. The test re- suits show that the descriptor obtained from the improved CNN-based feature descriptor learning method delivers a good perform- ance from both the robust and real-time aspects, and it can also work well in UAV image matching.

关 键 词：特征描述符 卷积神经网路 难样本对 鲁棒性 实时性 

分 类 号：P237[天文地球—摄影测量与遥感]