| 摘要: |
| 高分辨一维距离像（High Resolution Range Profile, HRRP）常应用于雷达自动目标识别领域，HRRP数据结构复杂，从中提取稳定可靠的特征是HRRP目标识别的关键，本文提出一种融合网络模型，用于舰船HRRP的目标识别。模型首先通过BERT（Bidirectional Encoder Representations from Transformers）进行初步特征提取，再通过并行网络提取深度特征，左侧分支使用多尺度卷积神经网络（Multi?scale Convolutional Neural Network, MCNN）模块提取不同尺度的局部特征信息，并通过SE（Squeeze?and?Excitation）对卷积结果进行优化，更好地关注数据中的关键信息，右侧分支使用双向门控循环网络（Bidirectional Gated Recurrent Unit, BiGRU）捕捉序列中的长期依赖关系，结合多头注意力模块可以更好地捕捉不同位置间的相关性，最后对结果进行拼接，最大程度地利用不同网络的优势，提升模型的分类性能。实验结果表明，模型能够有效学习HRRP序列中的特征，有较好的识别性能。 |
| 关键词: 高分辨距离像 BERT模块 MCNN网络 BiGRU网络 |
| DOI：DOI:10.3969/j.issn.1672-2337.2025.02.010 |
| 分类号:TN957.51 |
| 基金项目:国家自然科学基金重点项目（No.62293544） |
|
| HRRP Target Recognition Method Based on Fusion Network |
|
WU Wenjing, DAN Bo, WANG Zhongxun
|
|
1. School of Physics and Electronic Information, Yantai University, Yantai 264005, China;2. Shandong Data Open Innovation Application Laboratory of Smart Grid Advanced Technology, Yantai University, Yantai 264005, China;3. Naval Aviation University, Yantai 264001, China
|
| Abstract: |
| High resolution range profile（HRRP） is commonly used in the field of radar automatic target recognition. The data structure of HRRP is complex, and extracting stable and reliable features from it is crucial for HRRP target recognition. This paper proposes a fusion network model for the target recognition of ship HRRP. The model first performs preliminary feature extraction through bidirectional encoder representations from transformers（BERT）, followed by deep feature extraction through a parallel network. The left branch uses a multi?scale convolutional neural network （MCNN） module to extract local feature information at different scales. The convolution results are optimized by squeeze?and?excitation（SE） to better focus on key information in the data. The right branch employs a bidirectional gated recurrent unit（BiGRU） to capture long?term dependencies in the sequence. Combined with a multi?head attention module, the correlations between different positions can be better captured. Finally, the results are concatenated to maximize the advantages of different networks and improve the model’s classification performance. The experimental results show that the model can effectively learn the features from the HRRP sequences and has good recognition performance. |
| Key words: high resolution range profile BERT module MCNN network BiGRU network |
