YOLOv7-F:一种轻量级船舶实时检测算法  

YOLOv7-F:A lightweight real-time ship detection algorithm

作　　者：王微 郁强 王五桂 邢博闻 WANG Wei;YU Qiang;WANG Wugui;XING Bowen(Shanghai Zhong Chen Xin Wei Aerospace Technology Co.,Ltd.,Shanghai 201306,China;College of Engineering Science and Technology,Shanghai Ocean University,Shanghai 201306,China;Shanghai Zhongchuan NERC-SDT Co.,Ltd.,Shanghai 201114,China;China Ship Research and Design Center,Wuhan 430064,China)

机构地区：[1]上海中辰新威航天科技有限公司,上海201306 [2]上海海洋大学工程学院,上海201306 [3]上海中船船舶设计技术国家工程研究中心有限公司,上海201114 [4]中国船舰研究设计中心,武汉430064

出　　处：《兵器装备工程学报》2024年第11期11-18,共8页Journal of Ordnance Equipment Engineering

摘　　要：船舶检测在内河航运管理中至关重要,在复杂的水面条件下,船舶检测很难兼顾准确性和实时性。针对这个问题,提出了一种改进YOLOv7的船舶实时检测方法YOLOv7-F。YOLOv7-F将GhostNet引入骨干网络进行特征提取,再将分布移位卷积引入特征融合网络,实现了模型轻量化。在特征融合网络中引入注意力机制,弥补模型轻量化带来的精度损失。损失函数也进行相应改进,使检测模型更适用于船舶数据集。HPRship数据集的实验结果表明,与传统YOLOv7检测模型相比,计算量减少了38.8×10^(9),模型参数量减少了5.7×10^(6),检测精度mAP0.5提升了0.7%,为98.80%。YOLOv7-F在轻量化和检测精度之间取得了良好的平衡,能够准确实时地完成船舶检测任务,适合部署到存储和计算有限的小型设备上。Ship detection is crucial in inland waterway transportation management,and it is challenging to balance accuracy and real-time performance in complex water surface conditions.To address these issues in ship real-time detection,the YOLOv7-F model based on an improved YOLOv7 model has been proposed.GhostNet is introduced into the backbone network for feature extraction,and the distributed shift convolution is introduced into the feature fusion network to achieve model lightweight.An attention mechanism is introduced into the feature fusion network to compensate for the accuracy loss caused by model lightweight.The loss function is improved to make the detection model more suitable for ship datasets.Experimental results on the HPRship dataset show that compared with the traditional YOLOv7 detection model,the computational cost is reduced by 38.8×10^(9),the model parameter quantity is reduced by 5.7×10^(6),and the detection accuracy mAP0.5 is increased by 0.7%to 98.80%.YOLOv7-F achieves a good balance between lightweight and detection accuracy,allowing for accurate real-time ship detection tasks and is suitable for deployment on small devices with limited storage and computing resources.

关 键 词：船舶检测 深度学习 模型轻量化 注意力机制 YOLOv7 

分 类 号：U664.82[交通运输工程—船舶及航道工程] TP18[交通运输工程—船舶与海洋工程]