文献类型：期刊文献

中文题名：结合注意力机制的残差网络卫星图像切片分类方法

英文题名：Satellite Image Slice Classification Method Combining Attention Mechanism with Residual Network

作者：袁涌博[1];冯楠[2];刘泽宇[1];王梓鉴[1];曹林林[1]

第一作者：袁涌博

机构：[1]北京联合大学城市轨道交通与物流学院,北京100101;[2]北京化工大学化学工程学院,北京100029

第一机构：北京联合大学城市轨道交通与物流学院

年份：2025

卷号：40

期号：4

起止页码：146-155

中文期刊名：遥感信息

外文期刊名：Remote Sensing Information

收录：;北大核心:【北大核心2023】；

基金：北京市教委科研计划(KM201911417007)。

语种：中文

中文关键词：深度学习;影像分类;注意力机制;ResNet50;AMP

外文关键词：deep learning;image classification;attention mechanism;ResNet50;AMP

摘要：为了解决卫星图像分类任务中存在的复杂场景识别困难、泛化能力有待提升和训练效率的问题,提出了一种CMP-ResNet50的图像切片分类方法。该方法在传统的ResNet50模型中引入卷积块注意力模块(convolutional block attention module,CBAM),增强了模型对图像语义细节的挖掘和敏感性,在保留ResNet50模型优势的同时,有效提升了特征表达能力。同时,采用自动混合精度训练(automatic mixed precision training,AMP)策略,使每轮训练速度提升了23 s。实验在4个公开卫星图像数据集上进行。结果表明,该方法在卫星图像分类任务上的准确率均优于ResNeXt50、EfficientNet-b0、DenseNet121、MobileNet2、GoogLeNet等网络,特别在NWPU数据集上准确率提高了25.5个百分点,验证了所提出方法的有效性。消融实验进一步验证了CBAM模块和AMP训练策略对模型性能的积极影响。
To address the difficulties in recognizing complex scenes,the need to improve generalization capabilities,and training efficiency issues in satellite image classification tasks,a CMP-ResNet50 image slicing classification method is proposed.Specifically,this method introduces the convolutional block attention module(CBAM)into the traditional ResNet50 model,enhancing the model’s ability to mine and be sensitive to semantic details in images.While retaining the advantages of the ResNet50 model,it effectively improves the feature expression capability.Additionally,the automatic mixed precision training(AMP)strategy is adopted,which speeds up each round of training by 23 seconds.Experiments on four public satellite image datasets show that the proposed method outperforms networks such as ResNeXt50,EfficientNet-b0,DenseNet121,MobileNet2,and GoogLeNet in satellite image classification tasks,with an accuracy improvement of 25.5%on the NWPU dataset in particular,verifying the effectiveness of the proposed method.Ablation experiments further confirm the positive impact of the CBAM module and the AMP training strategy on model performance.