Publication
da Rocha, W. F., Azzag, H., Lebbah, M., Mokraoui, A. Benchmarking SRGAN-Upscaled YOLO for Enhanced Object Detection in Aerial Imagery. ICECCME 2025, pp. 1–8.
All projects
Conference paper · ICECCME 2025
SRGAN-upscaled YOLO
Super-resolution as a training strategy for aerial object detection
The problem
Object detection in aerial imagery is hard for two compounding reasons: the objects in a single scene span a very wide range of scales, and the images themselves vary a lot in quality.
A detector trained on that mixture spends capacity compensating for low-quality inputs instead of learning scale-invariant features.
The approach
-
SRGAN as a dataset-level intervention
Rather than changing the detector, a Super-Resolution Generative Adversarial Network generates higher-quality versions of the weakest images in the dataset, which then replace the originals during training and evaluation.
-
Ten YOLO architectures, three scenarios
The benchmark spans ten YOLO architectures across three evaluation scenarios, so the claim is about the strategy rather than about one lucky model.
-
Two input resolutions
Everything is measured at 416×416 and 640×640 pixels, since input resolution interacts directly with the small-object problem the method is trying to solve.
-
DOTA v1.5
A standard aerial detection benchmark with the scale heterogeneity the method targets.
Results
DOTA v1.5.
YOLOv5s-transformer
Best configuration
SRGAN ×2
Upscaling factor
10
Architectures compared
416² and 640²
Resolutions
The SRGAN-upscaled training strategy improves detection across the model family, not only for the best configuration.
Publication & code
Code
Public artefacts related to this work:
Related work
Other parts of the same research line.