北京测绘2025，Vol.39Issue(11)：1594-1601,8.DOI:10.19580/j.cnki.1007-3000.2025.11.006
结合卷积与移动曼巴的轻量级遥感目标检测方法
Lightweight remote sensing target detection method combining convolution and Mobile Mamba
摘要
Abstract
To address the problem of real-time target detection in remote sensing images on low-power hardware,this paper proposed a lightweight target detection model that combined local and global feature extraction.A lightweight feature extrac-tion network was built by combining convolutional neural network and mobile mamba unit(MMB),which jointly captured both local and global contextual features of the target.By incorporating"cross-scale fusion"and"feature selection"mecha-nisms,a multi-scale feature fusion network was constructed to combine feature maps from different levels and filter negative samples.During the detection stage,the target box localization and classification losses were computed using a combination of enhanced intersection-over-union ratio(IoU)and variational focal loss,while a consistent dual allocation detection head was used to output the detection results,avoiding the generation of excessive candidate boxes and reducing computational redundancy.The experimental results show that the constructed model achieves an average precision of 92.34%and 91.25%on two test datasets,improving by 5.82%and 6.88%,respectively,compared to real-time detection converter.On the embedded test platform,the detection of a single image takes only 16.42 ms,outperforming mainstream lightweight detec-tion models.关键词
遥感影像/轻量级目标检测/移动曼巴单元(MMB)/跨尺度特征融合/一致双重分配策略(DLA)Key words
remote sensing image/lightweight target detection/mobile mamba unit(MMB)/cross-scale feature fusion/dual allocation strategy(DLA)分类
测绘与仪器引用本文复制引用
梁开伦,肖泽标..结合卷积与移动曼巴的轻量级遥感目标检测方法[J].北京测绘,2025,39(11):1594-1601,8.基金项目
广东省科技项目(2024A1111120008) （2024A1111120008）
