北京测绘2025，Vol.39Issue(12)：1733-1740,8.DOI:10.19580/j.cnki.1007-3000.2025.12.004
结合ARConv与VIM的遥感影像目标检测模型
Remote sensing image object detection method combining ARConv and VIM
摘要
Abstract
To address the issue of low detection accuracy and weak generalization performance of existing models for object detection in remote sensing images,which is caused by complex backgrounds and significant scale differences among various ground objects,this paper proposes a remote sensing image object detection model that combines adaptive rectangular convo-lution(ARConv)and visual mamba(VIM)units.In the backbone network,the ARConv unit is used to construct a detailed feature extraction layer,and the VIM unit is used to build a global context extraction layer.A reparameterization mechanism is introduced in the bidirectional weighted feature pyramid to efficiently extract and fuse multi-granularity features.During the training phase,internal weighted intersection-over-union(IoU)and polygon loss are used to calculate the target regres-sion and classification losses.The model is guided to learn complex sample features through auxiliary weighted bounding boxes.Experimental results show that the proposed model achieves average precision values of 93.82%and 90.46%on two public datasets,outperforming current mainstream models,and is capable of outputting detection results in real-time under test conditions.关键词
遥感影像/目标检测/适应矩形卷积/视觉曼巴/重参数-双向加权特征金字塔Key words
remote sensing image/object detection/adaptive rectangular convolution/visual mamba/reparameterization-bidirectional weighted feature pyramid分类
天文与地球科学引用本文复制引用
何神佑,陈勇明..结合ARConv与VIM的遥感影像目标检测模型[J].北京测绘,2025,39(12):1733-1740,8.基金项目
广东省科技计划(2021B1212100003) （2021B1212100003）
