DHI-Net: A Novel Detail-Preserving and Hierarchical Interaction Network For Building Extraction

Accurate building extraction holds immense importance for many applications, such as urban planning and disaster rescue. Deep learning (DL)-based methods can automatically and efficiently extract buildings from remote sensing (RS) images. U-Net is one of the most widely used foundational models for...

Full description

Saved in:
Bibliographic Details
Published in:IEEE geoscience and remote sensing letters 2024, Vol.21, p.1-5
Main Authors: Song, Baogui, Shao, Wen, Shao, Pan, Wang, Jianming, Xiong, Jing, Qi, Chenwei
Format: Article
Language:English
Subjects:
Building codes
building extraction
Buildings
Convolution
Data mining
Decoding
Deep learning
detail-preserving
Feature extraction
hierarchical interaction
Image edge detection
Information processing
Modules
Public buildings
Remote sensing
Remote sensing (RS) images
Training
Urban planning
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate building extraction holds immense importance for many applications, such as urban planning and disaster rescue. Deep learning (DL)-based methods can automatically and efficiently extract buildings from remote sensing (RS) images. U-Net is one of the most widely used foundational models for constructing DL building extraction methods. However, the existing U-Net-based DL building extraction methods have the following limitations: The successive downsampling may cause the loss of details, and the interaction between the features from different layers is insufficient. To overcome the above limitations, this letter proposes a novel detail-preserving and hierarchical interaction network (DHI-Net) for building extraction, i.e., DHI-Net. Firstly, a detail-preserving module (DPM) is proposed based on SPD-Conv to preserve more detail information during the encoding process. Then, a hierarchical interaction module (HIM) is presented to enhance the interaction between different layers in the decoding process. HIM can realize the adaptive selection of the features of different layers and obtain multi-scale building features more effectively. Experimental results on two public building extraction datasets confirmed the effectiveness of the proposed DHI-Net. The code is available at https://github.com/HaH-1/building-extraction/tree/main/DHI-Net .
ISSN:1545-598X
1558-0571
DOI:10.1109/LGRS.2024.3439100