Small Object Detection Algorithm Based on Partial Convolution and Attention Fusion Detection Head
With the increasing utilization of unmanned aerial vehicles (UAVs), enhancing the detection perfor-mance of UAV aerial images has become increasingly crucial. This paper proposes a small object detection algorithm based on partial convolution and attention fusion detection head, aiming to address th...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Editorial Office of Aero Weaponry
2025-06-01
|
| Series: | Hangkong bingqi |
| Subjects: | |
| Online Access: | https://www.aeroweaponry.avic.com/fileup/1673-5048/PDF/2024-0168.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the increasing utilization of unmanned aerial vehicles (UAVs), enhancing the detection perfor-mance of UAV aerial images has become increasingly crucial. This paper proposes a small object detection algorithm based on partial convolution and attention fusion detection head, aiming to address the limitations of current mainstream object detection algorithms in detecting small objects in aerial images. To improve spatial feature extraction and control network computing time, a more efficient FasterNet backbone network is introduced along with partial convolution (PConv) to reduce memory access and redundant calculations during deep convolution. The feature extraction network is optimized to enhance the detection effectiveness for small-sized targets. Additionally, a Dynamic Head is incorporated into the detection head, effectively applying attention mechanism to improve overall detection performance. Finally, the bounding box loss function is optimized as Inner-ShapeIoU, focusing on shape and scale of the bounding box to improve the accuracy for bounding box regression calculation while utilizing auxiliary bounding boxes to expedite convergence speed. Experimental evaluations are conducted using the public dataset VisDrone2019. Compared with the original YOLOv8n algorithm, the proposed method achieves an 11.9% increase in accuracy P and a 13.4% increase in mAP50, indicating significant improvement in small object detection accuracy. |
|---|---|
| ISSN: | 1673-5048 |