.png){kind=icon}
Return
ASMNet: Axis-decoupled sparse mining for efficient remote sensing object detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.134210.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote sensing object detection (RSOD) requires effective long-range context modeling while maintaining computational efficiency. Existing large-kernel backbones often enlarge the receptive field through dense spatial aggregation, which may introduce redundant spatial interactions and non-negligible computational overhead, especially for high-resolution remote sensing images. To address this issue, we propose a lightweight and deployment-friendly backbone, termed Axis-Decoupled Sparse Mining Network (ASMNet), for efficient RSOD. The ASM module in ASMNet decomposes two-dimensional contextual modeling into sparse mining along the height and width axes, thereby reducing unnecessary dense spatial interactions while preserving long-range representation capability. Within the ASM module, Target-Aware Sparse Mining (TASM) serves as the core mechanism for selectively mining target-consistent sparse relations from compact axial candidates. By introducing axis-aware target priors and lightweight coordinate-wise offset selection, TASM enables ASMNet to capture useful long-range dependencies in a target-aware, localization-friendly, and computationally efficient manner. Extensive experiments on the DOTA series datasets and DIOR-R demonstrate that ASMNet achieves a favorable accuracy-efficiency trade-off across multiple detection frameworks. Notably, ASMNet-T attains 72.43% mAP on the DOTA-v1.5 single-scale setting while using only 29.8% of the FLOPs of PKINet-S and delivering substantially faster inference. On-device deployment experiments further validate the practical efficiency of ASMNet in edge-oriented inference scenarios. These results indicate that ASMNet effectively satisfies the long-range modeling requirement of RSOD while remaining lightweight, efficient, and deployment-friendly. Code: https://github.com/LarryInD/ASMNet .
Keywords:
Lightweight backbone
Remote sensing object detection
Target-aware sparse mining
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available
Cited Papers
No cited papers available