SWL-YOLO: a synergistic feature fusion strategy for small object detection in remote sensing images based on YOLOv11
To address the challenges of detail loss, feature extraction difficulties, densely distributed small objects, and insufficient feature information in degraded remote sensing images, we introduce SWL-YOLO, a lightweight model built upon YOLOv11. SWL-YOLO incorporates Spatial Adaptive Feature Module (...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
Institute of Electrical and Electronics Engineers Inc.
2025
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/123429/1/123429.pdf http://psasir.upm.edu.my/id/eprint/123429/ https://ieeexplore.ieee.org/document/11308109/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | To address the challenges of detail loss, feature extraction difficulties, densely distributed small objects, and insufficient feature information in degraded remote sensing images, we introduce SWL-YOLO, a lightweight model built upon YOLOv11. SWL-YOLO incorporates Spatial Adaptive Feature Module (SAFM), Wavelet Downsampling (WDown), and a Large Selective Kernel (LSK) mechanism to adaptively enhance both spatial and contextual representations. Specifically, the SAFM improves sensitivity to fine-grained spatial features, thereby improving its ability to perceive small targets and edges. The wavelet downsampling module performs wavelet decomposition and subsampling, preserving high-frequency detail information while reducing computational complexity. The LSK mechanism dynamically adjusts receptive fields, enabling the model to better handle small objects, complex backgrounds, and multi-category targets through spatially adaptive feature enhancement and context-aware scale selection. While SAFM ensures enhanced local feature modulation, LSK complements it by providing global context awareness, together forming a synergistic spatial feature fusion mechanism. Furthermore, building upon the CIoU of YOLOv11, we develop an improved GeoCIoU loss, which employs a dual-penalty mechanism for loss calculation to achieve more accurate training feedback. Experiments on the VisDrone and NWPU VHR-10 datasets indicate that SWL-YOLO outperforms the baseline models, with mAP50 improvements of 5.1 % and 4.2 %, respectively, showcasing its superior performance in remote sensing target detection. |
|---|