基于ViT-CNN混合网络的合成孔径雷达图像船舶分类

Synthetic aperture radar image ship classification based on ViT-CNN hybrid network

为了解决视觉转换器模型缺乏多尺度与局部特征捕获能力,难以适应合成孔径雷达图像船舶分类任务的问题,本文提出一种混合网络模型用于合成孔径雷达图像船舶分类。利用分阶段下采样网络结构,解决了ViT无法捕获多尺度特征的问题。通过在ViT模型的3个核心模块中融入卷积结构,设计了卷积标记嵌入、卷积参数共享注意力和局部前馈网络3个模块,使得网络能够同时捕获船舶图像的全局和局部特征,进一步增强了网络归纳偏置和特征提取能力。研究表明:本文所提模型在OpenSARShip和FUSAR-Ship2个通用合成孔径雷达船舶图像数据集上,分类准确率较最优方法分别提高了2.96%和4.18%,有效地提升了合成孔径雷达图像船舶分类性能。

In recent years,vision transformer(ViT)has made significant breakthroughs in the field of image classification.However,it is difficult to adapt to the task of synthetic aperture radar image ship classification due to its lack of multiscale and local feature capture capability.For this reason,this paper proposes a hybrid network model for synthetic aperture radar image ship classification.A staged downsampling network structure is designed to solve the problem that ViT is unable to capture multi-scale features.By incorporating the convolutional structure into three core modules of the ViT model,three modules,namely,convolutional token embedding,convolutional parameters sharing attention,and local feed-forward network,are designed,which enable the network to capture both global and local features of the ship images,and further enhance the network's inductive biasing and feature extraction ability.Experimental results show that the proposed model in this paper improves the classification accuracy by 2.96%and 4.18%compared with the existing optimal method on two generalized SAR ship image datasets,OpenSARShip and FUSARShip,respectively,which effectively improves the performance of SAR image ship classification.