Aerial small target detection algorithm based on cross-scale separated attention.

Area of Science:

• Computer Vision and Machine Learning
• Artificial Intelligence for Aerial Imaging

Background:

• UAV aerial photography presents challenges for target detection, including multi-scale targets, small object prevalence, occlusions, and background interference.
• Existing models like YOLOv11s have limitations in feature extraction, multi-scale target detection, and handling occluded targets.

Purpose of the Study:

• To develop an advanced UAV aerial small target detection algorithm, UAS-YOLO, that addresses the limitations of current models.
• To improve fine-grained feature extraction, cross-scale fusion, and occlusion resistance in UAV imagery.

Main Methods:

• Introduced an Adaptive Bidirectional Feature Pyramid Network (ABiFPN) for enhanced cross-scale feature fusion, dynamically adjusting weights for small targets.
• Integrated a Separated and Enhancement Attention Module (SEAM) to focus on key regions and compensate for information loss in occluded areas.
• Proposed a Universal Inverted Bottleneck (UIB) structure within the C3K2_UIB module for efficient feature selection and background noise suppression.

Main Results:

• UAS-YOLO achieved a 4.9 percentage point increase in mean Average Precision (mAP) on the VisDrone2019 dataset.
• The algorithm showed a 2.1 percentage point mAP improvement on the TinyPerson dataset.
• Demonstrated superior performance compared to existing advanced algorithms in complex UAV scenarios.

Conclusions:

• The proposed UAS-YOLO algorithm effectively addresses the challenges of small target detection in UAV aerial photography.
• The integration of ABiFPN, SEAM, and UIB modules significantly enhances detection accuracy and robustness against occlusion and background interference.