@Article{cmc.2026.083186,
AUTHOR = {Abdulhamid Victor Ibrahim, Haoyuan Li, Bingyang Guo, Ruiyun Yu},
TITLE = {Optimizing Small Object Detection in Low-Resolution Imagery: A Unified Super-Resolution and Detection Framework Using YOLO-Flex},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/27727},
ISSN = {1546-2226},
ABSTRACT = {The detection of small objects in low-resolution aerial imagery presents a persistent challenge in computer vision, where hardware constraints, imaging altitude, and scene complexity collectively degrade spatial detail to the point where standard detection frameworks fail. Existing super-resolution methods offer partial remedies but are limited by substantial computational costs and by feature discrepancies between Generative Adversarial Network-enhanced and real high-resolution images that degrade downstream detection accuracy. This paper presents YOLO-Flex, a unified framework that addresses these challenges through the co-design of a super-resolution module and a task-adapted object detection network, jointly optimized through a shared loss function. The YOLO-Flex approach employs a Generative Adversarial Network (GAN)-based super-resolution module in which the Residual Feature Cluster Network (RFCN) serves as the generator, trained adversarially against the Average Relativistic Discriminator to produce high-quality super-resolved images from low-resolution inputs, preserving block-level residual features that standard super-resolution-based methods progressively integrate in deep networks. The detection network is restructured from the YOLOv5 baseline through principled backbone reduction, addition of a stride-4 detection head, task-aware SE attention placement, and a trimmed bidirectional feature pyramid prioritizing high-resolution fusion pathways. The complete YOLO-Flex framework achieves competitive super-resolution quality with Peak Signal-to-Noise Ratio (PSNR) of 31.852 dB on DOTA, surpassing notable super-resolution techniques, while the detector subnetwork contains 7.05 million parameters and requires 8.0 GFLOPs, representing a 6.0 times reduction in computational cost compared to the YOLOv5m baseline. Experimental validation is conducted on three benchmark datasets: VisDrone2019-DET, DOTA, and COWC.},
DOI = {10.32604/cmc.2026.083186}
}