现代防御技术 ›› 2024, Vol. 52 ›› Issue (6): 88-96.DOI: 10.3969/j.issn.1009-086x.2024.06.012

• 目标特性与探测跟踪技术 • 上一篇    下一篇

复杂海况下的海上船舶目标检测算法

邢汇源1,2, 崔亚奇2,3,4, 王子玲2, 熊伟2, 蒋丙栋5   

  1. 1.中国人民解放军72506部队,山东 济南 250000
    2.海军航空大学 信息融合研究所,山东 烟台 264001
    3.中国航空工业沈阳飞机 设计研究所,辽宁 沈阳 210100
    4.西北工业大学 自动化学院,江苏 西安 610100
    5.江苏自动化研究所,江苏 连云港 222006
  • 收稿日期:2023-07-18 修回日期:2023-10-07 出版日期:2024-12-28 发布日期:2024-12-26
  • 作者简介:邢汇源(1996-),男,山东青州人。硕士生,研究方向为目标检测、多目标跟踪。
  • 基金资助:
    国家自然科学基金(62001499)

Target Detection Algorithm for Ships at Sea Under Complex Sea Conditions

Huiyuan XING1,2, Yaqi CUI2,3,4, Ziling WANG2, Wei XIONG2, Bingdong JIANG5   

  1. 1.PLA 72506 Troops,Jinan 250000,China
    2.Research Institute of Information Fusion,Naval Aviation University,Yantai 264001,China
    3.Shenyang Aircraft Design and Research Institute,Aviation Industry of China,Shenyang 210100,China
    4.School of Automation,Northwestern Polytechnical University,Xi'an 610100,China
    5.Jiangsu Institute of Automation,Lianyungang 222006,China
  • Received:2023-07-18 Revised:2023-10-07 Online:2024-12-28 Published:2024-12-26

摘要:

针对海洋环境下无人艇载光学图像背景信息复杂、小目标,当前目标检测算法特征提取能力不足、定位能力弱、检测精度差等问题,提出了一种基于YOLOv7-Tiny的改进的海上目标检测算法。利用RepVGG在推理时的“无损耗”特性设计了特征提取模块RepELAN,在不影响推理速度的同时提升网络特征提取能力;改进特征共享融合网络,融合高分辨率特征图提升对小目标特征提取能力,裁剪低分辨率特征图减少网络推理计算量;针对网络在复杂环境下定位检测能力弱的问题,设计检测头模块,区分定位与分类2种解耦头,提升网络定位检测能力。在建立的船舶目标检测数据集进行实验,检测精度提升了6.2%,设计了模块消融实验与对比实验,论证了所提算法的有效性。

关键词: 目标检测, YOLOv7-Tiny, 解耦头, 特征融合, 小目标, 海上船舶

Abstract:

Aiming at the problems of complex background information and small target of unmanned vehicle-borne optical images in marine environment, insufficient feature extraction ability, weak positioning ability and poor detection accuracy of the current target detection algorithm, an improved maritime target detection algorithm based on YOLOv7-Tiny is proposed. The feature extraction module RepELAN is designed by using the "lossless" feature of RepVGG during inference, which improves the feature extraction capability of the network without affecting the inference speed. The feature sharing and fusion network is improved, which fuses high-resolution feature maps to improve the ability to extract features of small targets, and crops low-resolution feature maps to reduce the amount of network inference calculation. Aiming at the problem that the network has weak positioning and detection capabilities in complex environments, the detection head module is designed to distinguish between two decoupling heads, positioning and classification, and improve the network positioning detection capability. In the established ship target detection dataset, the detection accuracy is improved by 6.2%, and the module ablation experiment and comparative experiment are designed, which demonstrates the effectiveness of the proposed algorithm.

Key words: object detection, YOLOv7-Tiny, decouple the head, feature fusion, small goals, ships at sea

中图分类号: