Optimized YOLOv8 for Lightweight Small Floating Object Detection on Unmanned Surface Vehicles

Abstract

In this paper, we propose PSP-YOLOv8s, an optimized version of YOLOv8 tailored for lightweight and efficient detection of small floating objects on unmanned surface vehicles (USVs). We introduce a Polarized Self-Attention module integrated with the C2f module to enhance feature extraction capabilities. Additionally, we incorporate a dedicated small object detection layer and utilize Partial Convolution to reduce computational overhead. Furthermore, Wise-IoUv3 is employed as the loss function to improve detection accuracy. Experimental results on the self-constructed USV-RSFO dataset and the Orca dataset demonstrate that PSP-YOLOv8s achieves state-of-the-art detection accuracy of 95.5% and 88.8%, respectively, while maintaining low model complexity. This work expands the application of vision technology for USVs, providing crucial support for water resource and ecosystem protection. Code is available at https://github.com/hongh07/PSP-YOLOv8s.