YOLO-MARS for Infrared Target Detection: Towards near Space

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• Infrared target detection in near space faces challenges like scale variation, noise, and low contrast.
• Existing models struggle with small targets and complex backgrounds in these environments.

Purpose of the Study:

• To propose an efficient and accurate infrared target detection model for near space applications.
• To address limitations of traditional downsampling and feature fusion in infrared imagery.
• To create a more robust dataset for near space target detection training.

Main Methods:

• Developed YOLO-MARS, incorporating Space-to-Depth (SPD) convolution, Grouped Multi-Head Self-Attention (GMHSA), and Light Adaptive Spatial Feature Fusion (LASFF) modules.
• Integrated near space images with the HIT-UAV dataset to create the NS-HIT dataset.
• Evaluated model performance against YOLOv8, focusing on accuracy and computational complexity.

Main Results:

• YOLO-MARS achieved a 5.4% increase in mAP@0.5 compared to YOLOv8.
• Model complexity increased by only 10% with significant accuracy improvements.
• Enhanced capability in detecting small targets and suppressing background noise.

Conclusions:

• YOLO-MARS offers an efficient and reliable solution for near space infrared target detection.
• The proposed model effectively handles scale variations and complex background noise.
• The new NS-HIT dataset aids in training more robust detection models for this domain.