DAN-YOLO-L: a lightweight YOLO approach for efficient infrared small object detection from UAV perspectives

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Small object detection in drone aerial photography is challenging due to small object scales, inconspicuous features, and complex backgrounds.
• Existing methods often struggle to achieve high accuracy and real-time performance simultaneously.

Purpose of the Study:

• To propose an improved YOLO-based detection model for enhanced small object detection in drone aerial imagery.
• To address limitations in feature extraction, multi-scale representation, and real-time processing for embedded applications.

Main Methods:

• A novel multi-channel feature extraction module (deconv-c3k2) combining Deconvolutional Network and C3 module was developed.
• An enhanced Auxiliary Head detection module was introduced to improve feature interaction across different network levels.
• The NWD-Inner-CIoU loss function was adopted to refine localization accuracy for small targets.
• An L1 pruning strategy was employed for model compression to meet real-time embedded deployment requirements.

Main Results:

• The proposed model significantly outperformed baseline models on the HIT-UAV dataset, achieving 81.9% mAP@0.5 and 51.4% mAP@0.5:0.95.
• Inference speed increased by 8.32%, while parameter count decreased by 20.8%, with a 79.2% reduction in model size.
• The model demonstrated strong generalization capabilities on the IRay Infrared Dataset.

Conclusions:

• The developed YOLO-based model offers significant improvements in detection accuracy, real-time performance, and lightweight design for small object detection in drone imagery.
• The approach exhibits practical value and stability for real-world applications, including embedded systems.