EABI-DETR: An Efficient Aerial Small Object Detection Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Small object detection in aerial imagery is crucial for tasks like remote sensing and UAV surveillance.
• Existing models face challenges with small object size, scale variations, and complex backgrounds, limiting performance.
• Capturing fine-grained semantics and high-resolution textures in aerial scenes remains difficult for current detectors.

Purpose of the Study:

• To propose an efficient aerial small object detection model, EABI-DETR (Efficient Attention and Bi-level Integration DETR), based on RT-DETR.
• To enhance the perception of small objects by integrating lightweight attention mechanisms and multi-scale feature fusion.
• To improve localization robustness for better handling of challenging aerial detection scenarios.

Main Methods:

• Developed a lightweight backbone network (C2f-EMA) combining C2f structure with an efficient multi-scale attention (EMA) mechanism.
• Designed a P2-BiFPN bi-directional multi-scale fusion module to incorporate shallow, high-resolution features and enhance cross-scale information flow.
• Introduced a Focaler-MPDIoU loss function to address hard samples during regression optimization.

Main Results:

• EABI-DETR achieved 53.4% mAP@0.5 and 34.1% mAP@0.5:0.95 on the VisDrone2019 dataset.
• The proposed model outperformed the baseline RT-DETR by 6.2% and 5.1% in respective metrics.
• EABI-DETR maintained high inference efficiency while demonstrating significant performance gains.

Conclusions:

• The integration of lightweight attention mechanisms and shallow feature fusion is effective for aerial small object detection.
• EABI-DETR offers a novel and efficient approach for UAV-based visual perception tasks.
• The proposed enhancements provide a new paradigm for improving small object detection in complex aerial scenes.