融合金字塔差分特征的新增建筑物检测网络  

A Newly Added Building Detection Network Fusing Pyramid Difference Features

作　　者：李普庆 丁海勇 于加东 LI Puqing;DING Haiyong;YU Jiadong(School of Remote Sensing and Geomatics Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;College of Medical Information Engineering,Shandong First Medical University and Shandong Academy of Medical Sciences,Taian,Shandong 271000,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,南京210044 [2]山东第一医科大学(山东省医学科学院)医学信息工程学院,山东泰安271000

出　　处：《遥感信息》2023年第1期146-154,共9页Remote Sensing Information

基　　金：国家自然科学基金项目(41571350)。

摘　　要：针对当前使用孪生网络检测新增建筑物时,简单的通道合并不能有效突出影像变化特征这一问题,提出一种融合金字塔差分特征的网络,将孪生网络提取的特征图作差分来突出变化特征。为了减少参数量,孪生网络使用深度可分离卷积;在编码层深处,使用不同感受野的空洞卷积提取多尺度特征;解码阶段使用DUpsampling上采样来减少影像信息的丢失。利用武汉大学建筑物数据集进行实验,结果表明,网络在提取建筑物新增区域时可以有效抑制噪声的干扰和解决边界粗糙问题。相比于经典的变化检测网络,可以获得更高的检测精度,准确率达到91.33%,召回率达到88.31%,F1分数达到89.79%,总体精度达到96.64%。For the current use of the Siamese network to detect new buildings,simple channel merging cannot effectively highlight the changing characteristics of the image.This research proposes a network combined with pyramid difference feature extraction.The feature map extracted by the Siamese network is differentiated to highlight the changing features,and the Siamese network uses depth wise separable convolution to reduce the amount of network parameters.In the deepest part of the encoding layer,the dilated convolution of different receptive fields is used to extract multi-scale features.The data-dependent up-sampling method is used in the decoding stage to reduce the loss of image information.Experiments with the building dataset of Wuhan university show that the proposed network can effectively suppress noise interference and solve the problem of rough boundaries when extracting the increasing areas of buildings.Compared with the classic change detection network,higher accuracy can be obtained.The accuracy,the recall rate,the F1 score and the overall accuracy are 91.33%,88.31%,89.79%and 96.64%,respectively.

关 键 词：神经网络 金字塔差分 建筑物检测 深度可分离卷积 DUpsampling 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]