基于可变形的多尺度自注意力特征融合SAR影像舰船识别  

Deformable-based multi-scale self-attentive feature fusion for SAR image ship recognition

作　　者：周慧[1] 朱虹 陈澎[2] ZHOU Hui;ZHU Hong;CHEN Peng(School of Computer and Software,Dalian Neusoft Information University,Dalian 116023,China;Navigation College,Dalian Maritime University,Dalian 116026,China)

机构地区：[1]大连东软信息学院计算机与软件学院,辽宁大连116023 [2]大连海事大学航海学院,辽宁大连116026

出　　处：《大连海事大学学报》2024年第4期110-118,共9页Journal of Dalian Maritime University

基　　金：辽宁省教育厅基本科研项目面上项目(LJKMZ20222006)。

摘　　要：SAR图像中不同类别船舶的目标特征区分不明显,当船舶类别较多时会出现识别准确率下降的问题。为更好地提取类别特征,本文提出一种识别模型DCN-MSFF-TR,借鉴Transformer encoder-decoder思想,在主干网络中加入可变形卷积模块(DCN),同时,将经过Transformer多尺度自注意力处理的特征层按照特征金字塔的方式在模型合适的位置进行特征融合,每一层不仅能够利用自身的信息,还能够综合利用其他层的特征。在公开数据集Open SARShip-3-Complex三分类数据集和Open SARShip-6-Complex六分类数据集的验证结果表明,平均识别精确率分别达到78.1%和66.7%,说明本文方法相对其他识别模型可更有效识别出SAR图像中的舰船类别。The feature differentiation of different categories of ship targets in SAR images is not clear,and the recognition accuracy may decrease when there are many ship categories.To better extract category features,this paper proposed a recognition model DCN-MSF-TR,which drawed on the idea of Transformer encoder-decoder and added a deformable convolutional module(DCN)to the backbone network.At the same time,the feature layers processed by Transformer multi-scale self attention were fused at appropriate positions in the model in a feature pyramid manner,and each layer can not only utilize its own information,but also comprehensively utilize the features of other layers.The validation results on the Open SARShip-3 Complex three class dataset and Open SARShip-6 Complex six class dataset show that the average recognition accuracy reaches 78.1%and 66.7%,respectively,which show that the proposed method can more effectively identify ship categories in SAR images compared to other recognition models.

关 键 词：SAR图像 船舶识别 Transformer架构 可变卷积 特征融合 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]