基于轻量化NDFEDet-SOLOv2的遥感图像建筑物提取方法

A Method for Extracting Buildings from Remote Sensing Images Based on Lightweight NDFEDet-SOLOv2

目的在地籍测绘和灾害管理等领域中,建筑物轮廓和位置的自动提取是至关重要的一环。为了解决高分辨率遥感图像建筑物因环境因素导致分割精度不准确等问题,提出了一种改进的轻量化SOLOv2实例分割模型——NDFEDet-SOLOv2。方法该模型选用双向特征金字塔网络(BiFPN)特征融合方式的轻量级EfficientDet网络,其中将骨干网络部分的EfficientNet升级为EfficientNetv2,EfficientNetv2中的三层MBConv模块SE注意力更换为含有DropBlock正则化的轻量级标准化注意力机制(NAM),构成NAD-MBConv模块。BiFPN特征融合部分,向其尾端各特征层并入双水平路由注意视觉变压器(BiFormer),形成双向水平路由注意特征金字塔网络结构(Bi-FPN-Former),从而聚焦微小建筑物轮廓信息,以实现更高层次的特征融合。结果NDFEDet-SOLOv2模型相较于传统轻量级SOLOv2实例分割算法,平均精度mAP、mAP 50和mAP 75分别提高了3.9%、3.7%和2.5%,检测帧率(FPS)提高了2.7帧/s。结论轻量化NDFEDet-SOLOv2实例分割算模型消除了建筑物边角的图像畸变,在地理环境空间不均等复杂情况下也能准确提取出遥感图像建筑物的基本轮廓,从而为城市布局更新和建筑变化检测提供理论参考。

Objective In the fields of cadastral surveying and disaster management,automatic extraction of building contours and positions is crucial.To address the problem of inaccurate segmentation accuracy of buildings in high-resolution remote sensing images due to environmental factors,an improved lightweight SOLOv2 instance segmentation model called NDFEDet-SOLOv2 was proposed.Methods This model adopted a lightweight EfficientDet network with a bidirectional feature pyramid network(BiFPN) feature fusion method.The EfficientNet in the backbone network was upgraded to EfficientNetv2,and the three-layer MBConv module SE attention in EfficientNetv2 was replaced with a lightweight normalized attention mechanism(NAM) containing DropBlock regularization,forming the NAD-MBConv module.The feature fusion part of BiFPN incorporated a bi-level routing attention visual transformer(BiFormer) into each feature layer at its tail end to form a bi-directional horizontally routing attention feature pyramid network structure(Bi-FPN-Former),which focused on the contour information of small buildings and achieved higher-level feature fusion.Results Compared with traditional lightweight SOLOv2 instance segmentation algorithms,the NDFEDet-SOLOv2 model has improved average accuracies by 3.9%,3.7%,and 2.5% for mAP,mAP_(50),and mAP_(75),respectively,and improved detection frame rate(FPS) by 4.7 frames/s.Conclusion The lightweight NDFEDet-SOLOV2 instance segmentation algorithm model eliminates image distortion of building edges and corners,and can accurately extract the basic contours of buildings in remote sensing images even in complex and uneven geographical environments.This provides a theoretical reference for the update of urban layouts and the detection of building changes.