ED-DETR: An edge-guided dual-branch feature optimization network for enhanced small object detection in UAV images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Object detection in unmanned aerial vehicle (UAV) imagery faces challenges with low accuracy and poor detection of small objects.
• Existing methods struggle to effectively extract and utilize features for small targets in complex environments.

Purpose of the Study:

• To develop an improved object detection network for UAV imagery, specifically addressing the detection of small objects.
• To enhance feature extraction and localization accuracy for small targets in challenging aerial datasets.

Main Methods:

• Proposed the edge-guided dual-branch feature optimization network (ED-DETR), built upon the real-time detection transformer (RT-DETR).
• Introduced a dual-branch feature extraction and aggregation unit (DFEA) for separate high-frequency and low-frequency feature extraction, utilizing reparameterization for efficiency.
• Developed an edge-guided dual-branch feature extraction and aggregation (EDFEA) module to leverage edge information for enhanced small-target perception.
• Implemented a hybrid loss function (Mal-Shape) combining Mal loss and ShapeIoU loss for robust bounding box matching and precise localization.

Main Results:

• ED-DETR achieved performance improvements of 3.7%, 1.2%, and 1.6% in mAP50 on the VisDrone, RSOD, and UAVDT datasets, respectively, compared to the baseline.
• The proposed method demonstrated enhanced small-object detection capabilities without significantly increasing computational complexity.
• The edge-guided approach and hybrid loss function proved effective in complex scenarios and for precise localization.

Conclusions:

• The ED-DETR model effectively addresses the limitations of low detection accuracy and poor small-object detectability in UAV imagery.
• The integration of edge features and a hybrid loss function significantly boosts performance for small object detection tasks.
• ED-DETR offers a promising solution for real-time object detection in UAV applications requiring high accuracy for small targets.