基于轻量化网络的水下目标检测算法  

Underwater Target Detection Algorithm Based on Lightweight Network

作　　者：许朝龙 解志斌[1] 宋科宁 XU Chaolong;XIE Zhibin;SONG Kening(Ocean College,Jiangsu University of Science and Technology,Zhenjiang 212003,China;Unit 95829,PLA,Xiaogan 432100,China)

机构地区：[1]江苏科技大学海洋学院,江苏镇江212003 [2]中国人民解放军95829部队,湖北孝感432100

出　　处：《无线电工程》2025年第2期264-270,共7页Radio Engineering

基　　金：国家自然科学基金(62276117);高端外国专家引进计划(G2023014110)。

摘　　要：基于机器视觉的无人水下航行器(Unmanned Underwater Vehicle, UUV)在工作时往往面临着嵌入式设备计算资源有限、实时检测速度慢的问题,为了解决这些问题,设计了一种轻量化网络检测算法——YOLOv8-FasterECA-Slim-neck-Focaler-EIoU(YOLOv8-FESF)。在骨干网络中,基于FasterNet Block和高效通道注意力(Efficient Channel Attention, ECA)机制构建全新的C2f_Faster_ECA模块,降低特征网络的参数量和计算量,采用Slim-neck作为颈部结构,进一步压缩模型的规模;重新设计检测头,利用参数共享的思想合并特征提取模块,从而轻量化模型,提高检测速度;使用边框回归损失函数Focaler-EIoU动态调整损失值,解决难易样本不平衡的问题,以提高检测精度。实验结果证明,所提模型在RUOD数据集上表现良好,相较于YOLOv8n基准模型,参数量和计算量分别减少40%和54%,检测速度提高10.5帧/秒,平均精度均值(mean Average Precision, mAP)仅下降0.1%,适合部署在计算设备资源受限的水下目标检测平台。Unmanned Underwater Vehicle(UUV)based on machine vision frequently encounters the problems of limited computing resources of embedded devices and slow real-time detection speed during operation.To solve these problems,a lightweight network detection algorithm,namely YOLOv8-FasterECA-Slim-neck-Focaler-EIoU(YOLOv8-FESF),is designed.In the backbone network,a novel C2f_Faster_ECA module is established based on the FasterNet Block and the Efficient Channel Attention(ECA)mechanism to reduce the number of parameters and computational load of the feature network.Moreover,the Slim-neck is employed as the neck structure to further compress the model scale.The detection head is reengineered to leverage the concept of parameter sharing to merge the feature extraction modules,thereby reducing the model s weight and improving the detection speed.The frame regression loss function Focaler-EIoU is utilized to dynamically adjust the loss value to resolve the problem of imbalanced sample difficulty and improve the detection accuracy.The experimental results show that the proposed model performs well on the RUOD dataset.Compared with the YOLOv8n baseline model,it witnesses a 40%reduction in parameters and a 54%decrease in computation,a 10.5 frame/s increase in detection speed,and only a 0.1%drop in mean Average Precision(mAP),rendering it suitable for deployment on underwater target detection platforms with constrained computing device resources.

关 键 词：Focaler-EIoU YOLOv8 水下目标检测 轻量化网络 PConv 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]