一种多尺度稀疏卷积的高分辨率遥感图像变化检测方法  

Change Detection Method for High-resolution Remote Sensing Image Based on Multi-scale and Sparse Convolution

作　　者：胡蕾[1] 江宇 李进 张永梅[2] HU Lei;JIANG Yu;LI Jin;ZHANG Yong-mei(School of Computer Information Engineering,Jiangxi Normal University,Nanchang 330022,China;School of Information,North China University of Technology,Beijing 100144,China)

机构地区：[1]江西师范大学计算机信息工程学院,南昌330022 [2]北方工业大学信息学院,北京100144

出　　处：《小型微型计算机系统》2020年第11期2365-2370,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61662033,61262036)资助.

摘　　要：高分辨率遥感图像中地物越来越清晰,变化检测不仅要能检测出大目标的变化,也要能检测出小目标的变化,还要兼顾干扰因素对变化性质的判断.本文针对高分辨率遥感图像变化检测,提出一种多尺度稀疏卷积模型,利用不同数量不同尺度的卷积层提取多尺度的特征,通过1×1卷积层实现跨通道信息整合,把不同通道中相关性高、同一空间位置的特征聚合在一起,有效减少了通道数量和参数数量,使得模型呈现稀疏性,大幅度削减参数的相互依存关系,一定程度上缓解了过拟合问题,使模型具有高效的学习能力和高容量的表达能力.同时,本文探讨了孪生网络和多通道网络对变化检测精度的影响.通过对不同场景的高分辨率遥感图像数据进行实验,表明所提方法能有效检测大目标和小目标的变化情况.The ground objects in high-resolution remote sensing images are becoming clearer.With interference factors are involved,change detection not only need to find the changes of large targets,but also the changes of small targets.This paper proposes a multi-scale sparse convolution for high-resolution remote sensing image change detection.Multi-scale features are extracted by different numbers of multi-scale convolution layers.And 1×1 convolution layers are used to aggregate features with high correlation in the same spatial position of different channels.The convolutional layer effectively reduces the number of channels and parameters and makes the model sparser.It greatly weakens the interdependence of parameters,and alleviates the model of overfitting in certain degree,which enables the model to learn efficiently and express with high capacity.Meanwhile,this paper explores the impact of Siamese networks and Multi-channel networks on the accuracy of change detection.Experiments on high-resolution remote sensing images of different scenes show that the proposed method can effectively detect the changes of large targets and small targets.

关 键 词：高分辨率遥感图像 变化检测 深度学习 稀疏卷积 多尺度 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]