SFC-YOLO: An Accuracy-Enhanced and Parameter-Efficient YOLO Framework for Small Vehicle Detection in Aerial Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Small vehicle detection in aerial imagery is crucial for intelligent transportation and urban monitoring.
• Challenges include small object size, complex backgrounds, and viewpoint variations, leading to detection errors.

Purpose of the Study:

• To propose SFC-YOLO, an accuracy-enhanced and parameter-efficient framework for small vehicle detection in aerial images.
• To improve detection performance by addressing challenges like missed detections and false alarms.

Main Methods:

• The SFC-YOLO framework integrates a Feature Complementary Block (FCB) for local-detail and semantic-context compensation.
• A Dynamic Feature Alignment Upsampling (DFAU) unit is used for content-adaptive feature alignment.
• A P5-only Hidden-State Attention (HSA) module enhances global semantic interaction.

Main Results:

• SFC-YOLO reduced parameter count from 2.584 M to 2.466 M.
• Achieved improved mean Precision (0.602 to 0.665), Recall (0.599 to 0.608), and mAP50 (0.614 to 0.652) on the VEDAI dataset.
• The framework shows a trade-off of increased GFLOPs (6.3 to 8.4) for enhanced accuracy with fewer parameters.

Conclusions:

• SFC-YOLO offers improved detection accuracy for small vehicles in aerial images with a reduced parameter count.
• The framework represents a parameter-efficient approach, albeit with higher theoretical computation.
• Results demonstrate the effectiveness of the proposed modules in enhancing small vehicle detection capabilities.