S-YOLO: An enhanced small object detection method based on adaptive gating strategy and dynamic multi-scale focus...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detecting small objects in drone aerial imagery is challenging due to real-time and computational constraints.
• Existing methods often struggle with diminutive targets and background interference.

Purpose of the Study:

• To propose S-YOLO, an efficient and streamlined small object detection framework based on YOLOv10.
• To enhance the detection of small objects in aerial imagery under computational constraints.

Main Methods:

• Developed S-YOLO with Enhanced Small Object Detection Layers for increased semantic richness.
• Incorporated the C2fGCU Module with Gated Convolutional Units (GCU) for adaptive feature modulation.
• Utilized a Dynamic Multi-Scale Fusion (DMSF) Module with SE-Norm for optimized feature integration.

Main Results:

• S-YOLO achieved significant mAP50:95 improvements: 5.3% (VisDrone2019), 4.4% (AI-TOD), and 1.4% (DOTA1.0) over YOLOv10-n.
• Maintained fewer parameters than YOLOv10-n.
• Processed 285 images per second, demonstrating high efficiency.

Conclusions:

• S-YOLO is a highly efficient solution for real-time small object detection in aerial imagery.
• The proposed framework effectively addresses the challenges of detecting diminutive targets.
• S-YOLO offers a superior balance of performance and computational efficiency.