集成小波变换与全局感知的轻量建筑提取网络

打开文本图片集

Lightweight building extraction network integrating wavelet transform and global awareness

SHAO Wen1，2， SHAO Pan1，2*， SONG Baogui²， XIONG Biao1.2

（1.Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering， China Three Gorges University，Yichang 443OO2，China;

2. College of Computer and Information Technology， China Three Gorges University， Yichang 443002，China;

3.College of Physics and Information Engineering，Fuzhou University，Fuzhou 35OlO8， China）

Abstract：Building extraction based on deep learning is an important research direction in the field of remote sensing.To effctively balance computational effciency and extraction accuracy，a lightweight building extraction network integrating wavelet transform and global awareness is proposed. First， by integrating parameter sharing，star-shaped operations，and depthwise separable convolution，a starshared depthwise convolution block is proposed to efficiently and accurately extract building features. Secondly，wavelet transform and frequency-domain spatial attention are introduced to propose an efficient boundary enhancement module that enhances the network’sability to characterize building boundary features.Finally，employing a lightweight non-local attention mechanism anda hierarchical feature interaction strategy，a global context-aware module is proposed. This module significantly improves the fusioneffectiveness of hierarchical features and enhances the overall perception capability during the network’s decoding stage. Experimental results demonstrate that the proposed network achieves Intersection over Union（IoU） scores of 90% and 79.16% on the publicly available WHU and Inria building extraction datasets，respectively. Concurrently，the model maintains a low parameter count （Params）of 3.O9 million，FLOPs of 4.93 billion，and an inference speed of 30.24 frames per second （FPS）. Compared to existing methods such as Swin Transformer，BuildFormer， SDSCUNet，EasyNet， HDNet，and CaSaFormerNet，the proposed method achieves higher extraction accuracy while maintaining low computational complexity， achieving a superior balance between computational eficiency and extraction accuracy .

Key words： building extraction; lightweight; boundary enhancement；wavelet transform； global context

1引言

随着遥感技术的快速发展，基于高分辨率遥感影像的建筑物提取技术在城市规划、环境监测及灾害评估等领域发挥着重要作用，但是也带来精度与效率的双重挑战。（剩余21444字）