基于多尺度胶囊Swin Transformer的SAR图像目标识别方法  

Multi-scale capsule Swin Transformer-based method for SAR image target recognition

作　　者：侯宇超 王洁[1] 李洪涛 郝岩 段晓旗 黄凯文 田有亮[2] HOU Yuchao;WANG Jie;LI Hongtao;HAO Yan;DUAN Xiaoqi;HUANG Kaiwen;TIAN Youliang(Shanxi Key Laboratory of Cryptography and Data Security,Shanxi Normal University,Taiyuan 030031,China;State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;School of Mathematics and Statistics,Taiyuan Normal University,Taiyuan 030002,China)

机构地区：[1]山西师范大学密码学与数据安全山西省重点实验室,山西太原030031 [2]贵州大学公共大数据国家重点实验室,贵州贵阳550025 [3]太原师范学院数学与统计学院,山西太原030002

出　　处：《通信学报》2025年第3期274-290,共17页Journal on Communications

基　　金：国家自然科学基金资助项目(No.42461057,No.62272123,No.42371470);山西省基础研究计划基金资助项目(No.202303021212164,No.202303021212255);山西省高等学校科技创新基金资助项目(No.2022L405);山西省研究生科研创新基金资助项目(No.2024KY474);贵州省基础研究基金资助项目(No.2024129)。

摘　　要：通过协同胶囊单元的语义特征编码和Swin Transformer的上下文特征图建模优势相结合,提出了一种多尺度胶囊Swin Transformer网络(MSCSTN),将胶囊编码和Swin Transformer联合应用于SAR图像目标识别。该网络集成3个并行的胶囊Swin Transformer编码结构,融合后对输入图像进行分类。每个结构通过基于膨胀卷积切片划分的胶囊令牌编码器和三维胶囊Swin Transformer模块构建,能捕获更深层次、更广泛的语义特征。在运动和静止目标的获取与识别(MSTAR)数据集及FUSAR-Ship数据集上的实验结果表明,MSCSTN在各种测试条件下均优于其他方法。结果表明,MSCSTN展现了良好的识别性能、泛化能力和应用潜力。A multi-scale capsule Swin Transformer network(MSCSTN)was proposed by synergizing the semantic feature encoding of capsule units with the context feature mapping of Swin Transformer.Capsule encoding and the Swin Transformer were jointly applied to SAR image target recognition.The network was integrated with three parallel capsule Swin Transformer encoding structures,which were fused to classify the input image.Each structure was constructed through a capsule token encoder based on expanded convolutional slice partition and a 3D capsule Swin Transformer module,which designed to capture of more profound and extensive semantic features.The experimental results on the moving and stationary target acquisition and recognition(MSTAR)dataset and FUSAR-Ship dataset were shown to demonstrate that MSCSTN outperformed other methods under various test conditions.The results demonstrate that MSCSTN exhibits excellent recognition performance,generalization ability,and potential for application.

关 键 词：膨胀卷积切片分区 胶囊令牌编码器 三维胶囊Swin Transformer模块 多尺度胶囊Swin Transformer网络 SAR图像目标识别 

分 类 号：TN957.52[电子电信—信号与信息处理]