.png){kind=icon}
Return
MSGHNet: multi-scale gradient-aware hypergraph network for remote sensing small object detection
.png){kind=icon}
PRE
AI
DOI：10.1088/1361-6501/ae599d.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the application scenarios of drone vision technology, accurate and efficient small object detection is a core demand driving technological advancement. Despite progress in existing methods, key challenges persist, including feature loss during subsampling, insufficient multi-scale fusion, and weak modeling of object relationships in complex scenes. To address these issues, this paper proposes an improved high-resolution small object detection model MSGHNet: it integrates a small-object-oriented subsampling module PRDown to preserve critical feature integrity, designs a PMSCF module for enhanced cross-scale feature expression, and innovatively embeds an adaptive hypergraph computation module into the neck network. Collaborating with PMSCF, the edge-aware hypergraph module leverages its strong capability in modeling complex relationships to adaptively capture spatial topologies, semantic similarities, and dynamic interactions between objects, effectively enhancing detection performance and generalization for small objects in complex scenarios. Experimental results demonstrate significant performance improvements on both Dior and VisDrone2019 mainstream drone image datasets. Compared to baseline model YOLOv11s, the proposed method achieves 3.04% and 2.48% gains in mAP@50, respectively, while maintaining real-time inference speed, providing an effective solution for precise small object detection in drone scenarios.
Keywords:
small object detection
drone vision
multi-scale feature fusion
hypergraph network
remote sensing
Journal
IF:
3.4
Papers:
2.7K
Citations:
2.3W
Organization
Cited Papers
LAYN: Lightweight Multi-Scale Attention YOLOv8 Network for Small Object Detection.png)
IEEE ACCESS
OAAI
IEEE ACCESSIF3.6
OAAIShareSaveShareSave
View more