Remote sensing object detection is a rapidly growing field in artificial intelligence, playing a critical role in advancing the use of Unmanned Aerial Vehicles (UAVs) for real-world applications such as disaster response, urban planning, and environmental monitoring. Yet, designing models that balance both high accuracy and fast, lightweight performance remains a challenge. UAVs often capture images where objects appear in different sizes, angles, and lighting conditions, all while operating on devices with limited computing power. This creates the need for innovative deep learning models that can deliver robust results without relying on heavy computational resources.
To address these challenges, a research team from Osaka Metropolitan University, led by graduate student Hoang Viet Anh Le and Associate Professor Tran Thi Hong with her collaborator team, has developed a novel detection framework tailored for UAVs. At the core of this work is the Partial Reparameterization Convolution Block (PRepConvBlock), which reduces the complexity of convolution operations while maintaining strong feature extraction. This innovation makes it possible to use larger kernels, enabling longer-range feature interactions and significantly expanding receptive fields. Building on this, the researchers introduced a Shallow Bi-directional Feature Pyramid Network (SB-FPN), which fuses information between shallow and deeper feature scales to enhance visual representation.
These innovations come together in a new architecture named SORA-DET (Shallow-level Optimized Reparameterization Architecture Detector). Designed specifically for UAV remote sensing, SORA-DET employs up to four detection heads and achieves both high accuracy and efficiency. In benchmark testing, the detector reached 39.3% mAP50 on the challenging VisDrone2019 dataset and 84.0% mAP50 on the SeaDroneSeeV2 validation set—outperforming most large-scale models while being significantly smaller and faster. In fact, SORA-DET requires nearly 88.1% fewer parameters than conventional one-stage detectors, with an inference speed as fast as 5.4 milliseconds.
This combination of compact design, high detection performance, and real-time adaptability makes SORA-DET a promising solution for UAV-based remote sensing. By enabling accurate object detection on lightweight devices, this research opens the door to impactful applications in disaster management, search-and-rescue operations, and beyond.
The findings were published in Scientific Reports.
