系统工程与电子技术 ›› 2023, Vol. 45 ›› Issue (8): 2415-2422.doi: 10.12305/j.issn.1001-506X.2023.08.15

• 电子技术 • 上一篇    下一篇

基于改进YOLOv5s的红外舰船检测算法

李海军1, 孔繁程1,*, 林云2   

  1. 1. 海军航空大学岸防兵学院, 山东 烟台 264001
    2. 烟台大学教务处, 山东 烟台 264005
  • 收稿日期:2022-03-31 出版日期:2023-07-25 发布日期:2023-08-03
  • 通讯作者: 孔繁程
  • 作者简介:李海军(1965—), 男, 教授, 博士, 主要研究方向为航空导弹精确制导与作战运用
    孔繁程(1998—), 男, 硕士研究生, 主要研究方向为图像目标检测与识别
    林云(1981—), 男, 博士, 主要研究方向为计算机仿真

Infrared ship detection algorithm based on improved YOLOv5s

Haijun LI1, Fancheng KONG1,*, Yun LIN2   

  1. 1. Coastal Defense College, Naval Aviation University, Yantai 264001, China
    2. Office of Academic Affairs, Yantai University, Yantai 264005, China
  • Received:2022-03-31 Online:2023-07-25 Published:2023-08-03
  • Contact: Fancheng KONG

摘要:

针对反舰导弹红外成像导引头在舰船尺度角度变化剧烈以及检测弱小舰船目标时检测能力差的问题, 提出一种基于改进YOLOv5s的目标检测算法,该方法综合考虑舰船目标特点。首先, 采用深度可分离卷积模块降低网络模型参数; 其次, 在主干网络引入坐标注意力机制提升关注目标通道信息特征的能力; 然后, 使用自适应空间特征融合策略优化空间权重分配; 最后,对损失函数进行改进, 提高目标检测框的可靠度。对比实验证明所提算法可将检测精度由原来的86.47%提升至91.64%, 全类平均精度(mean average precision, mAP)指标由原来的85.56%提升到89.35%,且在相同条件下所提算法亦优于其他目标检测网络模型。

关键词: 舰船检测, 红外图像, YOLOv5s, 坐标注意力机制, 特征融合

Abstract:

Infrared imaging seeker of anti-ship missile varies sharply in ship scale angle and its detection ability is poor when detecting small and weak ship targets. To solve this problem, a target detection algorithm based on improved YOLOv5s is proposed in this paper. Firstly, the depth separable convolution module is used to reduce the network model parameters. Secondly, the coordinate attention mechanism is introduced into the backbone network to improve the ability to pay attention to the target channel information features. Then, the adaptive spatial feature fusion strategy is used to optimize the spatial weight allocation. Finally, the loss function is improved to improve the reliability of the target detection frame. Comparative experiments verify that the proposed method improve the detection precision from the original 86.47% to 91.64%, and the mean average precision (mAP) index is improved from the original 85.56% to 89.35%. The improved algorithm also outperforms other target detection network models under the same conditions.

Key words: ship detection, infrared imagery, YOLOv5s, coordinate attention mechanism, feature fusion

中图分类号: