A Feature-Enhanced Small Object Detection Algorithm Based on Attention Mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Object recognition using deep learning is crucial for Unmanned Aerial Vehicle (UAV) applications.
• UAV imagery presents challenges like small object size, high density, and background interference, hindering detection accuracy.

Purpose of the Study:

• To improve object detection performance in UAV images, specifically addressing challenges of small objects and inaccurate localization.
• To develop an enhanced deep learning framework for robust object recognition in complex aerial visual data.

Main Methods:

• Utilized YOLOv8s as the base framework, incorporating a multi-level feature fusion algorithm.
• Introduced an attention mechanism for improved small object feature extraction and a dynamic head for accurate localization.
• Implemented Slideloss and ShapeIoU to enhance learning from difficult samples and better handle bounding box variations.

Main Results:

• Achieved significant improvements on VisDrone2019 dataset, with nearly 10% increase in accuracy and 11% in recall compared to the baseline.
• On the AI-TODv1.5 dataset, the mean Average Precision at Intersection over Union (mAP50) improved from 38.8 to 45.2.

Conclusions:

• The proposed method effectively addresses key challenges in UAV-based object detection, outperforming baseline models.
• The integration of advanced techniques like attention mechanisms and specialized loss functions leads to superior performance in complex scenarios.