海洋测绘2025，Vol.45Issue(5)：37-41,5.DOI:10.3969/j.issn.1671-3044.2025.05.008
GEM-YOLO:一种轻量化前视声纳图像目标高效检测模型
GEM-YOLO:a lightweight and efficient detection model for forward-looking sonar image targets
摘要
Abstract
To address the problems of low resolution,strong noise interference,and blurred target features in forward-looking sonar images caused by complex underwater acoustic environments,as well as the issues of high computational complexity and deployment difficulties in existing YOLO models for sonar target detection,this paper proposes an improved GEM-YOLO model.Detection performance is enhanced through a three-stage optimization strategy.First,Ghost convolution is employed to construct a lightweight feature extraction network,significantly reducing model complexity.Second,a novel C3_EMSC module incorporating multi-scale dilated convolutions is designed to enhance feature representation of underwater targets.Finally,the EIOU loss function is introduced to improve the accuracy of bounding box regression.Experimental results on the URPC2022 sonar dataset show that:(1)In terms of detection performance,the improved model achieves a Precision of 0.973,Recall of 0.968,and mAP@0.5 of 0.974;(2)In terms of computational efficiency,the model parameters are reduced by 31.2％,FLOPs decrease by 35.8％,and inference speed reaches 105.778 FPS.Ablation studies further validate the effectiveness of each improvement module.This research provides a viable technical solution for real-time underwater target detection in resource-constrained environments and holds significant value for advancing intelligent applications of autonomous underwater equipment.关键词
前视声纳图像/目标检测/YOLOv5模型/轻量化模型/多尺度特征提取Key words
forward-looking sonar/object detection/YOLOv5/lightweight model/multi-scale feature extraction分类
天文与地球科学引用本文复制引用
祝捍皓,闫愣愣,文洪涛,苏峥,郝奕飞..GEM-YOLO:一种轻量化前视声纳图像目标高效检测模型[J].海洋测绘,2025,45(5):37-41,5.基金项目
福建省海洋物理与地质过程重点实验室开放基金(KLMPG2301) （KLMPG2301）
水声对抗技术国防重点实验室基础基金(JCKY2024207CH04). （JCKY2024207CH04）
