基于元学习的高光谱鲁棒性学习  

Meta-learning-based Robustness Learning for Hyperspectral Image

作　　者：廖启明 张桂烽 张国云 石乘仲 周晨明 李新平 赵林 LIAO Qiming;ZHANG Guifeng;ZHANG Guoyun;SHI Chengzhong;ZHOU Chenming;LI Xinping;ZHAO Lin(School of Information and Communication Engineering,Hunan Institute of Science and Technology,Yueyang 414000,China)

机构地区：[1]湖南理工学院信息科学与工程学院,湖南岳阳414000

出　　处：《成都工业学院学报》2023年第5期34-38,共5页Journal of Chengdu Technological University

基　　金：湖南省研究生创新项目(CX20211187);湖南理工学院研究生创新项目(YCX2022A29);湖南省自然科学基金项目(2020JJ4343);湖南省教育厅科研项目(22C0365)。

摘　　要：在有监督的高光谱图像分类任务中,标签噪声会严重影响分类器的性能,导致所训练模型容易产生错误的分类结果。因此,建立一个稳健的深度学习分类框架来处理带有标签噪声的高光谱图像数据集是一个重要且具有挑战性的问题。为缓解标签噪声对高光谱图像分类任务的影响,提出一种基于元学习的样本加权和分类框架。通过计算样本属于标签噪声的可能性,赋予相应的权重以降低标签噪声的影响并提高基础分类模型的鲁棒性。实验结果表明,在高光谱图像数据集Salinas和Botswana上,该框架能有效抑制所加标签噪声,提升基础分类网络的鲁棒性,获得更高的分类精度。Label noise refers to the erroneous labeling of data caused by human factors or other reasons.In the supervised Hyperspectral Image classification task,Label noise can significantly affect the performance of classifiers,resulting in incorrectly classified results from trained models.Therefore,building a robust deep learning classification model to handle Hyperspectral Image(HSI)datasets with label noise is a challenging and important problem.In the training process of a deep learning classification network,a sample weighting framework based on meta-learning was proposed.To reduce the impact of label noise and improve the robustness of the underlying classification model,weights were assigned to samples according to their likelihood of belonging to label noise.Experimental results show that in two common Hyperspectral Image datasets Salinas and Botswana,the meta-learning-based weighting framework can effectively suppress label noise,enhance the robustness of the base classification network and improve the accuracy of its classification results.

关 键 词：高光谱图像 元学习 样本加权 标签噪声 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]