融合多重多尺度特征的高分辨率遥感影像建筑物提取网络

A Building Extraction Network for High-resolution Remote Sensing Images Using Multiple Multiscale Features

针对高分辨率遥感影像因复杂背景信息导致的建筑物边界、角点以及内部信息出现的错分、漏分问题,提出了一种融合多重多尺度目标特征的DPRS-Net深度学习网络。DPRS-Net采用Resnet50与Swin-T(Tiny)的并行编码结构以结合两种编码优势,进而获取特征图的大范围深层信息;利用跳跃连接降低建筑物的边界特征损失;引入特征金字塔注意力模块和密集空洞空间特征金字塔池化模块,使采样过程中的建筑物细节特征损失减弱。为验证模型的优势性、分析性能提升原因,在WHU和自建Changchun3建筑物数据集上进行对比和消融实验。结果表明,DPRS-Net在两种数据集上均取得更高的精度,提取的建筑物信息更为完整,且模型各结构均能有效提升预测效果。

Aiming at the problem of misclassification and omission on building boundaries,corners,and internal information in high-resolution remote sensing images due to complex background information,this paper proposes a DPRS-Net deep learning network that integrates multiple multi-scale target features.DPRS-Net adopts the parallel coding structure of Resnet50 and Swin-T(Tiny)to combine the advantages of the two coding and to obtain large-scale deep information of the feature map.It reduces the loss of boundary features of buildings by using jump connections,and introduces the feature pyramid attention module and the dense void space feature pyramid pooling module,so that the loss of detailed features of buildings during the sampling process is weakened.To verify the superiority of the proposed model and analyze the reasons for the performance enhancement,this paper conducts comparison and ablation experiments on WHU and self-built Changchun3 building datasets.The results show that DPRS-Net achieves higher accuracy on both datasets,the extracted building information is more complete,and each structure of the model can effectively improve the prediction effect.