An improved UAV image object detection algorithm combining multi-scale feature fusion and receptive-field...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Unmanned Aerial Vehicle (UAV) image object detection faces challenges like scale variation, complex backgrounds, and dense small targets.
• Existing algorithms struggle with the unique characteristics of UAV imagery, limiting their effectiveness.

Purpose of the Study:

• To develop an improved object detection algorithm, MFRA-YOLO, for enhanced performance on UAV imagery.
• To address challenges including scale variation, dense small targets, and complex backgrounds in UAV images.

Main Methods:

• Proposed MFRA-YOLO algorithm built upon YOLOv8n, incorporating Monte Carlo attention into receptive-field attention-based convolution for enhanced cross-scale interaction.
• Implemented a multi-scale selective fusion module and scale sequence feature fusion for adaptive feature integration, improving small target detection.
• Introduced Focaler-PIoUv2 loss function to balance hard and easy samples and boost detection accuracy.

Main Results:

• MFRA-YOLO demonstrated superior accuracy-efficiency tradeoffs on the VisDrone2019 dataset compared to YOLOv8n and other YOLO variants.
• Achieved a 3.5% increase in mAP50 and a 2.3% increase in mAP50:95 over YOLOv8n with minimal parameter and computational cost increases.
• Maintained 143 FPS, satisfying real-time deployment requirements for UAVs, and showed strong generalization on the RSOD dataset.

Conclusions:

• MFRA-YOLO significantly enhances object detection performance for UAV imagery.
• The algorithm achieves an excellent balance between detection accuracy and computational efficiency.
• MFRA-YOLO offers distinct advantages over state-of-the-art algorithms for UAV scenarios and demonstrates robust generalization.