SenseLite: A YOLO-Based Lightweight Model for Small Object Detection in Aerial Imagery

Area of Science:

• Aerial remote sensing
• Computer vision
• Deep learning for object detection

Background:

• Detecting small objects in aerial imagery is difficult due to subtle features, complex environments, and low resolution.
• Existing deep learning models often require substantial computational resources, limiting real-time applications in areas like urban planning and traffic monitoring.

Purpose of the Study:

• To develop a lightweight and efficient model for accurate and real-time small object detection in aerial images.
• To improve upon existing object detection frameworks like YOLOv5 for aerial remote sensing applications.

Main Methods:

• Introduced SenseLite, a streamlined YOLOv5 architecture incorporating Involution in the backbone for enhanced semantics and GSConv/slim-Neck in the neck for reduced complexity.
• Integrated a squeeze-and-excitation (SE) mechanism to boost channel communication and improve detection accuracy.
• Utilized Soft-NMS to effectively handle overlapping detections for precise concurrent identification.

Main Results:

• SenseLite reduced parameters by 30.5% (7.05M to 4.9M) and computational load (GFLOPs from 15.9 to 11.2).
• Achieved a 5.5% mAP0.5 improvement, 0.9% higher precision, and 1.4% better recall on the DOTA dataset compared to YOLOv5.
• Demonstrated superior performance against other leading object detection methods in aerial imagery.

Conclusions:

• SenseLite offers a significant advancement in lightweight and efficient small object detection for aerial remote sensing.
• The model's improved accuracy and reduced computational demands enable practical real-time applications.
• SenseLite provides a competitive solution for critical tasks requiring precise aerial object identification.