LKD-YOLOv8: A Lightweight Knowledge Distillation-Based Method for Infrared Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Infrared object detection is crucial for military, security, and aerospace applications.
• Edge devices face challenges balancing accuracy and computational efficiency in infrared object detection.
• Robust models are needed for rapid, real-time inference on resource-constrained platforms.

Purpose of the Study:

• To enhance infrared target detection accuracy and model robustness for edge devices.
• To develop an efficient infrared object detection method suitable for real-time inference.
• To present LKD-YOLOv8, integrating YOLOv8 with knowledge distillation and lightweight components.

Main Methods:

• Implemented LKD-YOLOv8, combining YOLOv8 architecture with masked generative distillation (MGD).
• Incorporated lightweight convolution (Linear deformable convolution - LDConv) for spatial feature extraction.
• Integrated coordinate attention (CA) for refined feature alignment and employed a teacher-student model (YOLOv8s to YOLOv8n).

Main Results:

• LKD-YOLOv8 achieved a 1.18% improvement in mAP@0.5:0.95 compared to baseline methods.
• Reduced model parameter size by 7.9%, enhancing computational efficiency.
• Demonstrated effective balance between detection accuracy and efficiency for edge deployment.

Conclusions:

• LKD-YOLOv8 offers a viable solution for accurate and efficient infrared object detection on edge devices.
• The method's performance and reduced size make it suitable for real-time, resource-constrained infrared scenarios.
• Knowledge distillation combined with architectural improvements enhances model adaptability and performance.