PConv-YOLO: A Lightweight Oriented Object Detection Network for Edge Deployment in UAV Remote Sensing
Year: 2026
Type: conference-paper
Abstract: Ship detection in remote sensing images is a critical task for maritime surveillance and port management. Due to the extreme aspect ratios and arbitrary orientations of ships, oriented bounding box (OBB) detection is more suitable than horizontal bounding box methods. However, existing high-precision rotation detection networks (e.g., YOLO-OBB series) contain over 2.9M parameters, making them difficult to deploy on edge devices such as drones. While lightweight mechanisms like GhostConv achieve compression through cheap operations, their inherent “feature hallucination” process may blur high-frequency spatial details, resulting in a severe degradation of rotation angle prediction accuracy (with \(mA{P}_{50 - 95}\) dropping to 71.4%). To overcome this bottleneck, this paper proposes PConv-YOLO, an ultra-lightweight oriented object detection network for edge deployment. We present a task-oriented architectural optimization that introduces Partial Convolution (PConv) into the Neck of a YOLO-OBB framework. Specifically, we design the C3k2_PConv module to replace the standard C3k2. By processing only 25% of the input channels through convolution while preserving the remaining 75% via identity mapping, PConv reduces Neck layer computation by 40%. Experiments on the HRSC2016 dataset demonstrate that PConv-YOLO achieves 79.0% \(mA{P}_{50 - 95}\) with only 2.42M parameters (a 9.1% reduction compared to the baseline), achieving a highly competitive trade-off between parameter efficiency and geometric accuracy. Compared to the GhostConv approach, our method improves \(mA{P}_{50 - 95}\) by 7.6 percentage points and reduces missed detections by 30%. The results validate that PConv-YOLO provides a practical, task-specific lightweight solution for real-time ship detection on UAVs.
Authors: Yongqi Xia, Qi Cao, Zenghui Li
Institutions: Xi’an Jiaotong-Liverpool University, Shanghai University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204797658