Super-Resolution Enhanced Real-Time Detector for Small Objects in Aerial Imagery

Abstract of PraxisSuper-Resolution Enhanced Real-Time Detector For Small Objects in Aerial Imagery Small object detection in aerial imagery remains difficult because targets often occupy only a few pixels, particularly after the tiling, resizing, and early backbone downsampling of detection transformers. Real-time oriented detection transformers such as RT-DETR can deliver strong NMS-free performance, but their fixed input resolutions and multi-stage downsampling can weaken the spatial cues needed to detect and precisely localize small, densely-packed objects. This praxis evaluates whether inserting a lightweight super-resolution (SR) front end before feature extraction can mitigate that resolution bottleneck. An EDSR-based super-resolution module is integrated into two detector families, producing SRT-DETR and SRF-DETR variants at multiple scaling factors. The models are evaluated on a tiled DOTA validation set, with inference-time resizing matched to each detector family’s native input constraints/recommendations. Performance is reported using COCO-defined metrics, overall and stratified by object size, alongside real-time deployment-oriented efficiency measurements. Across both detector families, integrating super-resolution improves small object detection relative to the corresponding baselines, some by greater than 20%, with latency remaining under 50ms in the best model, and within 10% of the baseline, indicating that while accuracy gains are accompanied by increased compute overhead, they remain deployable in some real-time environments.

Author

• Robinson, Shane

Language

• en

Date created

• 2026

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Computer Science

Degree

• D.Engr.

Advisor

• Yu, Jeffrey C

Committee Member(s)

• Ng, Jonathan
• Fossaceca, John
• Yu, Jeffrey C

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/k0698875j