基于CBAM-Swin-Transformer迁移学习的海上微动目标分类方法

打开文本图片集

关键词：雷达目标分类；海上微动目标；迁移学习；Swin-Transformer 网络；注意力机制；时频分析中图分类号： TN958.2 文献标志码：A DOI：10.12305/j.issn.1001-506X.2025.04.1

Abstract：As an important means for maritime target detection and identification，radar requires fine grained description and classification of the motion characteristics of maritime targets，which is a key technology.Deep learning-based convolutional network clasification method，although model-independent，stillstruggle to adapt to the complex and diverse maritime environment and the variety of maritime targets，with limited generalization ability.Convolutional block attention module（CBAM） is integrated into the Swin-Transformer network，and based on transfer learning（TL） strategy，a maritime micromotion target classification method（TL-CBAMSwin-Transformer）is proposed to consider both ship targets and low-altitude rotorcraft flight targets，thereby enhancing the model classification adaptability under various observation conditions.Firstly，a maritime micromotion target model is established，and based on three radar measurement data sets，micromotion timefrequency datasets of sea surface non-uniform translational motion，threeaxis rotation，helicopters，and fixed rotor drones are constructed. Then，the TL-CBAM-Swin-Transformer network is designed，where CBAM extracts features from both the channel and spatial dimensions，enhancing its ability to extract multi-head attention informationat small scales.Experimental data verification result shows that compared to SwinTransformer，the classification accuracy of the proposed algorithm is improved by 3.43% ．By using TL method，the proposed network is pre-trainedon the ImageNet dataset and is transferred to Council for Scientific and Industrial Research（CSIR） micromotion targets with intelligent pixel processing（IPIX） radar micromotion targets as the source domain for pre-training，achieving a classification probability of 97.9% .With helicopter rotors as the source domain for pre-training，the proposed algorithm transfer to fixed rotorcraft drones，achieving a classification probability of 98.8% ，which vadidates the strong generalization ability of the proposed algorithm.

Keywords：radar target clasification；maritime micromotion target；transfer learning（TL）；Swin-Transformer network；attention mechanism；time-frequency analysis

0引言

海上目标监视在海洋权益保护和国家安全领域中具有广泛应用，雷达是海上目标探测的主要手段1，适用于复杂海洋环境和多样海上目标的自适应检测和分类方法是雷达的关键技术之一。（剩余17463字）