Swin-PAFF: A SAR Ship Detection Network with Contextual Cross-Information Fusion

作     者：Yujun Zhang Dezhi Han Peng Chen 

作者机构：School of Information EngineeringShanghai Maritime UniversityShanghai201306China 

出 版 物：《Computers, Materials & Continua》 (计算机、材料和连续体（英文）)

年 卷 期：2023年第77卷第11期

页      面：2657-2675页

核心收录：

学科分类：0710[理学-生物学] 080904[工学-电磁场与微波技术] 0810[工学-信息与通信工程] 1101[军事学-军事思想及军事历史] 0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by the National Key Research and Development Program of China under Grant 2021YFC2801001 the Natural Science Foundation of Shanghai under Grant 21ZR1426500 the 2022 Graduate Top Innovative Talents Training Program at Shanghai Maritime University under Grant 2022YBR004. 

主　　题：Transformer deep learning SAR object detection ship detection 

摘      要：Synthetic Aperture Radar(SAR)image target detection has widespread applications in both military and civil domains.However,SAR images pose challenges due to strong scattering,indistinct edge contours,multi-scale representation,sparsity,and severe background interference,which make the existing target detection methods in low accuracy.To address this issue,this paper proposes a multi-scale fusion framework(Swin-PAFF)for SAR target detection that utilizes the global context perception capability of the Transformer and the multi-layer feature fusion learning ability of the feature pyramid structure(FPN).Firstly,to tackle the issue of inadequate perceptual image context information in SAR target detection,we propose an end-to-end SAR target detection network with the Transformer structure as the backbone.Furthermore,we enhance the ability of the Swin Transformer to acquire contextual features and cross-information by incorporating a Swin-CC backbone network model that combines the Spatial Depthwise Pooling(SDP)module and the self-attentive mechanism.Finally,we design a cross-layer fusion neck module(PAFF)that better handles multi-scale variations and complex situations(such as sparsity,background interference,etc.).Our devised approach yields a noteworthy AP@0.5:0.95 performance of 91.3%when assessed on the HRSID dataset.The application of our proposed technique has resulted in a noteworthy advancement of 8%in the AP@0.5:0.95 scores on the HRSID dataset.