UAV Small Target Detection Model Based on Dual Branches and Adaptive Feature Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detecting small and dense targets in drone aerial imagery presents significant challenges.
• Existing models often struggle with scale inconsistency and limited feature representation for diminutive objects.

Purpose of the Study:

• To propose a novel small target detection model for drone aerial images.
• To enhance detection accuracy and efficiency while reducing computational overhead.

Main Methods:

• A dual-branch framework was developed to improve detection accuracy and parameter efficiency.
• Semantic and detail injection (SDI) and bidirectional adaptive feature fusion were integrated into the neck and head networks.
• A large separable kernel attention mechanism was incorporated into convolutional layers to enhance feature representation.

Main Results:

• The model demonstrated superior performance on the VisDrone2019 dataset.
• Achieved a 20.9% increase in mAP50 and a 23.7% increase in mAP50-95 compared to the original model.
• Reduced the total number of parameters by 61.3%, enhancing suitability for drone deployment.

Conclusions:

• The proposed model effectively addresses the challenges of small and dense target detection in drone imagery.
• The innovative feature fusion and attention mechanisms significantly improve detection capabilities across various scales.
• The model's efficiency and accuracy make it a practical solution for real-time aerial surveillance and analysis.