融合注意力引导的遥感影像场景分类方法研究
Research on Scene Classification Method of Remote Sensing Images with Integrated Attention Guidance
秦望博 1葛斌 1彭曦晨1
作者信息
- 1. 安徽理工大学计算机科学与工程学院,安徽淮南,232000
- 折叠
摘要
为了解决遥感影像中主体部分和背景信息容易混杂,导致场景分类任务中存在难以提取有效特征等问题,提出一种融合注意力引导的遥感影像场景分类方法.该方法将融合注意力模块和ShuffleNet单元模块交叉组合嵌入到轻量级网络ShuffleNetV2 中,使得网络能够充分提取遥感影像中的空间结构信息和通道权重信息,综合提取出图像的语义信息,聚焦影像中的核心关键部分,从而增强网络的特征识别能力,并在保证模型轻量的同时提升性能.通过在三个广泛使用的公开遥感数据集UCM、AID和NWPU上进行实验对比,结果表明,提出的方法在性能方面优于其他方法,验证了所提方法的有效性.
Abstract
In order to address the challenge of disentangling the main subject from the background information in re-mote sensing imagery,which often complicates the extraction of effective features for scene classification tasks,a no-vel method for remote sensing image scene classification that incorporates fused attention guidance is proposed.This method seamlessly integrates a fused attention module with ShuffleNet unit modules into the lightweight network ShuffleNetV2.This integration empowers the network to effectively capture both spatial structural details and channel weighting information from remote sensing imagery,thereby synthesizing meaningful semantic insights from the ima-ges and concentrating on the core and pivotal elements of the images;this approach significantly enhances the net-work's ability to recognize features while maintaining a lightweight model.Experimental comparisons,conducted on three widely utilized public remote sensing datasets—UCM,AID and NWPU—demonstrate that the proposed method surpasses other approaches in terms of performance,thus substantiating its effectiveness.
关键词
遥感影像/场景分类/注意力机制/轻量级网络/特征提取Key words
Remote sensing image/Scene classification/Attention mechanism/Lightweight network/Feature extrac-tion引用本文复制引用
基金项目
国家自然科学基金(62102003)
国家重点研发计划(2020YFB1314103)
安徽省自然科学基金(2108085QF258)
安徽省博士后基金(2022B623)
出版年
2024
国家科技期刊平台
NSTL
万方数据