顾及密集特征的遥感影像建筑物提取方法  

A Building Extraction Method from Remote Sensing Images Considering Dense Features

作　　者：华娟 陶于祥 罗舒月 罗小波[1] HUA Juan;TAO Yuxiang;LUO Shuyue;LUO Xiaobo(College of Computer Sciences and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065

出　　处：《遥感信息》2025年第1期87-94,共8页Remote Sensing Information

基　　金：国家重点研发计划政府间国际科技创新合作项目(2021YFEO194700);重庆市自然科学基金(CSTB2022NSCQ-MSX1234)。

摘　　要：如何快速、高精度提取建筑物信息对城市发展、灾害评估具有重要价值意义。针对DeeplabV3+模型提取建筑物存在错检漏检、边界识别粗糙等问题,文章提出一种顾及密集特征的建筑物提取网络DFE_DeeplabV3+。首先,使用MobiletNetV2作为骨干网络有效减少参数量;其次,引入密集特征金字塔模块和密集注意力模块以捕获更丰富的上下文信息和边界信息;最后,考虑低层特征包含更多原始信息,在解码区将两层低层特征融合,进一步提取建筑物的空间细节信息,增强模型性能。实验结果表明,该方法有效提高了模型的分割性能,在INRIA Aerial lmage数据集上取得了76.72%的IoU,明显改善了建筑物错检漏检、边缘信息模糊的分割效果,具有一定的实际应用价值。How to quickly and accurately extract building information is of great value for urban development and disaster assessment.This paper proposes a building extraction network DFE_DeeplabV3+that takes into account dense features to address issues such as false detections,missed detections,and rough boundary recognition when extracting buildings using the DeeplabV3+model.Firstly,use MobileNetV2 as the backbone network to effectively reduce the number of parameters.Secondly,dense feature pyramid modules and dense attention modules are introduced to capture richer contextual and boundary information.Finally,considering that the low-level features contain more original information,the two low-level features are fused in the decoding area to further extract spatial details of the building and enhance model performance.The experimental results show that the method can effectively improve the segmentation performance of the model,and 76.72%IoU can be obtained on the INRIA Aerial Image data set,which obviously improves the segmentation effect of building error detection and leakage detection and edge information ambiguity,and has certain practical application value.

关 键 词：深度学习 建筑物提取 多尺度特征 特征融合 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]