摘要
随着深度学习技术的发展,基于卷积神经网络(CNN)和Transformer的深度学习方法在全监督遥感图像场景分类任务中得到了广泛的关注与研究.然而,如何在标注样本有限的情况下实现良好的分类性能仍然具有挑战.考虑到CNN和Transformer在深度特征提取方式上的差异,提出一种CNN和Transformer半监督交叉学习的遥感图像场景分类方法(SCL-CTNet),通过构建CNN和Transformer输出的一致性约束,更好地提取未标记数据中的信息,指导模型训练.半监督交叉学习方法将弱增强图像在一个网络上的输出作为伪标签用于监督强增强图像在另一个网络的预测结果,充分利用未标记样本的局部-全局信息,鼓励两个网络对相同输入图像预测间的一致性,提高模型泛化性.使用自适应阈值筛选伪标签,提高伪标签可靠性.在AID和NWPU-RESISC45数据集上的实验结果证明了所提出方法的有效性.
With the development of deep learning technology,deep learning methods based on Convolutional Neural Networks(CNN)and Transformers have received extensive attention and research in fully supervised remote sensing image scene classification tasks.However,achieving good classification performance with lim-ited labeled samples remains challenging.Considering the differences in deep feature extraction methods between CNN and Transformers,a semi-supervised cross-learning method for remote sensing image scene classification(SCL-CTNet)was proposed.By constructing consistency constraints on the outputs of CNN and Transformers,information from unlabeled data to guide model training would be better extracted.The semisupervised cross learning method utilizes the output of weakly augmented images in one network as pseudolabels to supervise the predictions of strongly augmented images in another network,fully leveraging the localglobal information of unlabeled samples,encouraging consistency in predictions for the same input image between the two networks,and enhancing model generalization.Adaptive thresholding is used to filter pseudolabels,improving their reliability.Experimental results on the AID and NWPU-RESISC45 datasets demonstrate the effectiveness of the proposed method.
作者
单飞龙
吕鹏远
李梦晨
Shan Feilong;LüPengyuan;Li Mengchen(School of Information Engineering,Ningxia University,Yinchuan 750021,China;Ningxia Key Laboratory of Artificial Intelligence and Information Security for Channeling Computing Resources from the East to the West,Yinchuan 750021,China)
出处
《宁夏大学学报（自然科学版）》
CAS
2024年第3期325-332,共8页
Journal of Ningxia University(Natural Science Edition)
基金
国家自然科学基金资助项目(42001307)。
