基于正样本重构对比学习的小样本高光谱图像分类研究  

Hyperspectral Image Classification with Limited Samples based on the Positive Samples Reconstruction and Contrast Learning

作　　者：丰阳 罗天 戴元杰 朱甜甜 陈卓轩 陈腾飞 赵林[1,2] 唐峰 吴健辉 FENG Yang;LUO Tian;DAI Yuanjie;ZHU Tiantian;CHEN Zhuoxuan;CHEN Tengfei;ZHAO Lin;TANG Feng;WU Jianhui(Hunan Engineering Research Center of 3D Reconstruction and Intelligent Application Technology,Hunan Institute of Science and Technology,Yueyang 414006,China;College of Information and Communication Engineering,Hunan Institute of Science and Technology,Yueyang 414006,China;College of Physics and Electronic Science,Hunan Institute of Science and Technology,Yueyang 414006,China;College of Mechanical Engineering,Hunan Institute of Science and Technology,Yueyang 414006,China)

机构地区：[1]湖南理工学院三维重建与智能应用技术湖南省工程研究中心,湖南岳阳414006 [2]湖南理工学院信息科学与工程学院,湖南岳阳414006 [3]湖南理工学院物理与电子科学学院,湖南岳阳414006 [4]湖南理工学院机械工程学院,湖南岳阳414006

出　　处：《成都工业学院学报》2024年第4期34-38,共5页Journal of Chengdu Technological University

基　　金：湖南省研究生科研创新项目(CX20221237,CX20231224);湖南省自然科学基金项目(2023JJ30284);湖南省教育厅科研项目(23A0488,22C0365)。

摘　　要：对比学习作为一种有效的预训练方式,通过区分正负样本对,促使编码器学到良好的视觉表征,使用少量样本即可获得优越的模型性能。然而现有的对比学习方法,仅通过数据增强这一单一方式构造正样本,导致所构建的正样本缺乏真实性和丰富性,限制了模型的性能。为缓解这一问题,提出一种基于正样本重构的对比学习框架。通过构造样本支持集,搜索当前样本的最近邻域作为正样本,将正样本扩充到真实场景,进一步提升算法的分类性能。实验结果表明,在公开的高光谱图像数据集Pavia University和Houston上,所提出的算法仅使用每类10个标记样本即可获得优越的分类效果,在减少算法数据依赖的同时,提升了模型性能。Contrast learning,as an effective pre-training method,promotes the encoder to learn good visual representation by distinguishing positive and negative sample pairs,and superior model performance can be obtained by using a small number of samples.However,the existing contrast learning method only constructs positive samples through data enhancement,which leads to the lack of authenticity and richness of the constructed positive samples,and limits the performance of the model.To alleviate this problem,a contrastive learning framework based on positive sample reconstruction was proposed.By constructing the sample support set,the nearest neighbor of the current sample was searched as the positive sample,and the positive sample was extended to the real scene to further improve the classification performance of the algorithm.The experimental results show that in the published hyperspectral image datasets Pavia University and Houston,the proposed algorithm can obtain superior classification effect by using only 10 labeled samples for each category,which reduces the data dependence of the algorithm and improves the model performance.

关 键 词：对比学习 自监督学习 编码器 高光谱图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]