摘要
针对深度学习方法在图像分类任务中易产生过拟合,影响分类精度的现象,提出了一种冻结与微调相结合的迁移学习方法,构建一种能够提高遥感图像分类精度的模型。为体现新模型的性能,以AID数据集为基础,保证数据集其他因素不变,划分1/5的AID数据集构造小样本数据集;通过在公开数据集AID和构造小样本数据集上,与基础网络ResNet50及传统迁移学习方法构建的分类模型进行对比实验。结果表明,新提出的迁移学习方法构建的分类模型在数据集AID上准确率为96.63%、召回率为96.35%,在数据集S-AID上,模型准确率为90.54%、召回率为96.35%,这表明新提出的迁移学习方法构建的分类模型在遥感图像分类任务中具有更好的分类性能。
To address the issue of overfitting commonly observed in deep learning-based image classification tasks,which negatively affects classification accuracy,this study proposes a transfer learning method that combines parameter freezing with fine-tuning to enhance remote sensing image classification performance.To evaluate the proposed model,the AID dataset was used as the benchmark,and a small-sample subset(one-fifth of the AID dataset)was constructed while keeping other data characteristics constant.Comparative experiments were conducted on both the full AID dataset and the constructed small-sample dataset,using the baseline ResNet50 network and traditional transfer learning methods for reference.The experimental results show that the classification model developed with the proposed transfer learning strategy achieved an accuracy of 96.63%and a recall rate of 96.35%on the AID dataset.On the small-sample S-AID dataset,the model achieved an accuracy of 90.54%and a recall rate of 96.35%.These findings demonstrate that the proposed transfer learning method significantly improves classification performance for remote sensing image recognition tasks.
作者
夏文生
XIA Wensheng(College of Applied Engineering,Gandong University,344000,Fuzhou,Jiangxi,PRC)
出处
《江西科学》
2025年第6期1154-1161,共8页
Jiangxi Science
基金
江西省教育厅科学技术研究项目(GJJ2403803)
赣东学院院长基金项目(12224000613)。
