Underwater object detection plays a crucial role in safeguarding and exploiting marine resources effectively. Addressing the prevalent issues of limited storage capacity and inadequate computational power in underwater robots, this study proposes FEB-YOLOv8, a novel lightweight detection model. FEB-YOLOv8, rooted in the YOLOv8 framework, enhances the backbone network by refining the C2f module and introducing the innovative P-C2f module as a replacement. To compensate for any potential reduction in detection accuracy resulting from these modifications, the EMA module is incorporated. This module augments the network's focus on multi-scale information, thus boosting its feature extraction capabilities. Furthermore, inspired by Bi-FPN concepts, a new feature pyramid network structure is devised, achieving an optimal balance between model lightness and detection precision. The experimental results on the underwater datasets DUO and URPC2020 reveal that our FEB-YOLOv8 model enhances the mAP by 1.2% and 1.3% compared to the baseline model, respectively. Moreover, the model's GFLOPs and parameters are lowered to 6.2G and 1.64M, respectively, marking a 24.39% and 45.51% decrease from the baseline model. These experiments validate that FEB-YOLOv8, by harmonizing lightness with accuracy, presents an advantageous solution for underwater object detection tasks.
Copyright: © 2024 Zhao et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.