基于回归模型与注意力的轻量化SAR舰船检测模型
作者:
作者单位:

1.中国科学院智能红外感知重点实验室 中国科学院上海技术物理研究所,上海 200083;2.国科大杭州高等研究院,浙江 杭州 310024;3.中国科学院大学,北京 100049

作者简介:

通讯作者:

中图分类号:

基金项目:

国家自然科学基金项目(61975222)和中国科学院地球微卫星热红外光谱仪项目(XDA19010102)


The research on lightweight SAR ship detection method based on regression model and attention
Author:
Affiliation:

1.Key Laboratory of Intelligent Infrared Perception, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China;2.Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Hangzhou 310024, China;3.University of Chinese Academy of Sciences, Beijing 100049, China

Fund Project:

Supported by National Natural Science Foundation of China(61975222)and CASEarth Minisatellite Thermal Infrared Spectrometer Project(XDA19010102)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    合成孔径雷达(SAR)具有不受云层干扰、可全天时、全天候对地观测的特点,基于SAR图像的舰船检测已广泛用于海洋搜救、港口侦察、领海防御等民用或军用领域。然而,与大型舰船相比,像素点少、对比度低的小型舰船存在漏检率高的问题,并且速度和精度之间的平衡成为舰船检测算法天基应用的难点。针对以上问题,本文提出了一种基于YOLOv5s模型改进的舰船检测轻量化模型(ImShips)。首先,针对船体大小差异导致的漏检问题,采取在网络底部使用感受野较小的标准卷积,提升了模型对小规模舰船空间信息的获取能力。同时在网络顶部设计了放大感受野的扩张卷积,保留了更多的语义特征,有利于大目标的特征提取。接着,提出将轻量级的通道注意力机制应用于YOLOv5的骨干网络和特征融合网络,通过对提取到的特征按重要性分配权重,实现纹理信息的筛选。最后,在下采样时采取深度可分离卷积代替标准卷积,减少了模型参数的数量,进一步提高了模型的推理速度。实验结果表明,在SAR舰船检测SSDD和ISSID数据集中,改进后的ImShips模型在保证精度的同时,所需的浮点计算数比YOLOv5s模型减少了45.61%,检测速度提高了8.31%。ImShips模型网络规模小,检测速度快,在实时天基舰船检测中具有较大的应用潜力。

    Abstract:

    Synthetic aperture radar (SAR) has the advantages of all-sky and all-weather earth observation without cloud interference. Ship detection based on SAR images has been widely used in civil and military fields, including maritime search and rescue, port reconnaissance, territorial sea defense. However, different from large ships, the misdetection rate of small ships with fewer pixels and lower contrast is high. And it is difficult to balance speed and accuracy during on-orbit ship detection. To solve the above problems, an improved lightweight ships detection method (ImShips) based on YOLOv5s is proposed. Firstly, the standard convolution with small receptive field is adopted at the bottom of the baseline to obtain spatial information of small ships. And the dilated convolution with enlarged receptive field is added at the top of the baseline to preserve more semantic features, which is conducive to extract large targets feature. Then, a lightweight channel attention mechanism is applied to the backbone and neck of YOLOv5. And the weight is allocated to filter more important texture information. Finally, the depth-wise separable convolution is adopted to replace the standard convolution during down-sampling to reduce the number of parameters and improve the inference speed. Compared with YOLOv5s model, the experimental results show that ImShips achieves an increase in AP, while the FLOPs are reduced by 45.61%, and the speed is increased by 8.31% in SSDD and ISSID datasets. The speed and accuracy of ImShips model are improved effectively for sea surface object detection. The proposed method has great application potential in on-orbit ship detection.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-09-26
  • 最后修改日期:2022-01-17
  • 录用日期:2021-10-20
  • 在线发布日期: 2022-01-10
  • 出版日期: