Consistent and comprehensive scale aggregation network for drone-view small object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Object detection in Unmanned Aerial Vehicle (UAV) scenarios is hindered by small object sizes.
• Small objects present challenges due to limited visual cues and information loss during feature extraction.
• Positional sensitivity of small objects complicates accurate bounding box prediction.

Purpose of the Study:

• To introduce a novel network, the Consistent and Comprehensive Scale Aggregation Network (C²SANet), for enhanced small object detection in UAVs.
• To improve feature representation and localization accuracy for small objects.

Main Methods:

• Developed a Multi-Scale Interactive Feature Pyramid Network (MSI-FPN) with Deformable-Based Spatial Calibration (DSC) and Scale Feature Enhancement (SFE) modules.
• DSC enhances semantic propagation consistency by calibrating features across scales.
• SFE unifies spatial sizes and facilitates comprehensive information exchange using an encoder-decoder structure.
• Implemented a Coarse-to-Fine Detection Head (CFDH) for iterative refinement of bounding box predictions.

Main Results:

• C²SANet demonstrates significant improvements in small object detection performance.
• The MSI-FPN effectively generates high-quality feature representations for small objects.
• The CFDH accurately refines localization predictions, addressing positional sensitivity.

Conclusions:

• C²SANet offers an effective solution for accurate small object detection in UAV imagery.
• The proposed network architecture and modules show strong generalizability and effectiveness.
• This work advances the capabilities of computer vision systems in challenging aerial surveillance and monitoring applications.