AMSA-YOLO: Real-time object detection with adaptive multi-scale attention mechanism.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection is crucial for applications like autonomous driving and surveillance.
• The YOLO (You Only Look Once) algorithm series excels in real-time single-stage object detection.
• Existing YOLO models struggle with detecting small objects and in dense scenes.

Purpose of the Study:

• To improve object detection accuracy, especially for small and dense objects.
• To introduce an enhanced YOLO algorithm named AMSA-YOLO (Adaptive Multi-Scale Attention YOLO).

Main Methods:

• Developed AMSA-YOLO incorporating scale-aware modules.
• Integrated adaptive spatial attention and adaptive channel attention mechanisms.
• Evaluated performance on benchmark datasets like COCO, VisDrone, and CrowdHuman.

Main Results:

• AMSA-YOLO achieved a 2.3% mAP@0.5:0.95 improvement over YOLOv8s on COCO.
• Demonstrated a 3.6% improvement in small object detection AP.
• Maintained competitive inference speed with only a 10.3% decrease.

Conclusions:

• AMSA-YOLO significantly enhances object detection accuracy, particularly for small objects.
• The adaptive multi-scale attention mechanism proves effective in challenging detection scenarios.
• The proposed method offers a practical and effective solution for real-world object detection tasks.