BDK-YOLOv8: An Enhanced Algorithm for UAV Infrared Image Object Detection

Abstract

This paper presents an infrared small object detection algorithm based on YOLOv8n to address challenges like large model size, complex backgrounds, poor small object detection, and scale variations. First, a new C2f-DCNv3 module is introduced to reduce parameter redundancy and enhance feature extraction. A Bidirectional Feature Pyramid Network (BiFPN) is added to the neck structure for improved detection of very small objects, enabling better multi-scale feature fusion. An improved SIOU loss function is also proposed, prioritizing small object samples and those with average-quality annotations. Finally, channel pruning is applied to reduce model parameters and computational complexity, improving detection efficiency.Experimental results show that the proposed algorithm achieves 94.3% mAP50 on the HIT-UAV dataset, a 1.6% improvement over the original YOLOv8n, with a 3% increase in recall. Model parameters and computational load are reduced by 55.1% and 1.2%, respectively, while the model size decreases by 1.77MB. Overall, the improved model offers a strong balance between accuracy and efficiency, making it well-suited for embedded devices and industrial drone detection in various scenarios.

Keywords

• Infrared target detection
• DCNv3
• channel pruning
• YOLOv8n