| 摘要: |
| 雷达高分辨距离像（High Resolution Range Profiles, HRRP）能够反映被测物体的结构信息，并且由于其易于获取的特性，近年来成为雷达自动目标识别（Radar Automatic Target Recognition, RATR）领域重要的研究方向。本文中针对雷达HRRP序列数据提出了一种通过格拉姆角场（Gramian Augular Field, GAF）方法将其转换成二维图像处理的方法，此方法能够在保留原有信息的基础上增加一个维度的特征，便于神经网络进行特征提取；并且针对小样本数据集问题，提出了一种基于迁移学习的带有高效局部注意力（Efficient Local Attention, ELA）机制的卷积神经网络架构，此架构冻结神经网络前5层卷积层的参数，对注意力模块以及全连接层进行单独训练，使模型具有更快地收敛速度和更高的识别精度。通过在5类舰船数据集上的测试发现，迁移学习方法比从头训练模型的训练时间减少了6%，准确率提升了11.62%。 |
| 关键词: 高分辨距离像 雷达自动目标识别 小样本学习 格拉姆角场 高效局部注意力 |
| DOI： |
| 分类号:TN957.52 |
| 基金项目:国家自然科学基金项目（62388102） |
|
| Few-Shot HRRP Target Recognition Network Based on Efficient Local Attention |
|
|
|
|
| Abstract: |
| Radar high resolution range profiles (HRRP) can reflect the structural information of the object under test, and due to its easy accessibility, it has become a significant research direction in the field of radar automatic target recognition (RATR) in recent years. In this paper, a method is proposed for radar HRRP sequence data to be converted into two-dimensional image processing by the Gramian angular field (GAF) method, which preserves the original information while adding an additional dimension of features, facilitating feature extraction by neural networks. To address the issue of small-sample datasets, a convolutional neural network architecture based on the transfer learning with an efficient local attention (ELA) mechanism is proposed. This architecture freezes the parameters of the first 5 convolutional layers of the neural network, and trains the attention module as well as the fully connected layer separately, so that the model has a faster convergence speed and higher recognition accuracy. Tests on five types of ship datasets demonstrates that the transfer learning approach reduced the training time by 6% and improved the accuracy by 11.62% compared to models trained from scratch. |
| Key words: high resolution range profiles radar automatic target recognition few-shot learning gramian angular field efficient local attention |