MicroSight-DETR: spatial-preserving real-time transformer with multi-domain fusion for UAV micro-object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• UAV monitoring faces challenges with low detection accuracy for small targets.
• Existing models like RT-DETR exhibit stage-specific degradation in aerial imagery analysis.

Purpose of the Study:

• To develop an enhanced real-time detection model, MicroSight-DETR, for improved small target detection in UAV scenarios.
• To address identified degradation mechanisms within the RT-DETR pipeline.

Main Methods:

• Introduced Global Efficient Modeling (GEM) backbone for global feature modeling with linear complexity.
• Developed Multi-domain Adaptive Fusion Dynamics (MAFD) encoder for adaptive spatial-frequency dual-domain fusion.
• Implemented Spatial Preserving Aggregation with Multi-scale (SPAM) neck to preserve shallow detail features.

Main Results:

• MicroSight-DETR achieved 51.3% mAP@0.5 (9.9% relative improvement) and 31.8% mAP@0.5:0.95 (12% increase) on the VisDrone2019 dataset.
• Demonstrated additive improvements from the three proposed modules through ablation studies.
• Achieved real-time inference at 78 FPS with 16.1M parameters and 64.5 GFLOPs.

Conclusions:

• MicroSight-DETR offers a robust solution for small target detection in UAV imagery.
• The complementary design of GEM, MAFD, and SPAM modules effectively mitigates degradation issues.
• The model provides efficient and accurate real-time processing for aerial surveillance applications.