JOURNAL ARTICLE

Cross-domain transfer learning algorithm for few-shot ship recognition in remote-sensing images

CHEN HuajieLYU DanniZHOU XiaoLIU Jun

Year: 2024 Journal:   National Remote Sensing Bulletin Vol: 28 (3)Pages: 793-804   Publisher: Science Press

Abstract

跨域迁移学习旨在利用现有公开数据集,突破源域和目标域样本类别空间须一致的约束,提升目标域样本的识别精度。针对现有跨域迁移学习算法应用于遥感图像小样本舰船目标识别时存在的迁移类别受限和负迁移问题,本文提出一种基于源域样本相关性排序的跨域迁移学习算法:首先将目标域样本逆向加入源域分类任务中,根据加入前后各类别源域样本的识别精度变化情况,对源域样本进行相关性排序,将其划分为强/弱/负相关样本;然后采取自监督联合学习策略,在目标域分类网络中引入自监督角度预测辅助分支,筛选出的强相关源域样本仅参与该辅助分支的训练,不改变目标域主分类网络的结构。算法通过相关性排序去除了弱/负相关源域样本,有效避免了负迁移;引入自监督角度预测辅助分支,在保持主分类网络结构完整性的同时,充分利用了强相关源域样本的有效信息,学习到更具泛化能力的目标特征。实验结果显示:在遥感舰船小样本目标数据集上,提出的算法优于跨域迁移学习中广泛使用的Fine-tune(微调)算法;与仅使用主分类网络的目标域识别算法相比,遥感舰船目标识别精度提升了17.59%。

Keywords:
Shot (pellet) Transfer of learning Computer science Domain (mathematical analysis) Artificial intelligence Transfer (computing) Computer vision Pattern recognition (psychology) Algorithm Remote sensing Geology Mathematics Materials science

Metrics

1
Cited By
0.64
FWCI (Field Weighted Citation Impact)
17
Refs
0.64
Citation Normalized Percentile
Is in top 1%
Is in top 10%

Citation History

Topics

Machine Learning and ELM
Physical Sciences →  Computer Science →  Artificial Intelligence
Optical Systems and Laser Technology
Physical Sciences →  Engineering →  Electrical and Electronic Engineering
Advanced Algorithms and Applications
Physical Sciences →  Engineering →  Control and Systems Engineering

Related Documents

© 2026 ScienceGate Book Chapters — All rights reserved.