基于小样本分类的光谱图像关联特征降维挖掘  

Dimension Reduction Mining of Spectral Image Association Features Based on Small Sample Classification

作　　者：于春霞 车银超[2] YU Chun-xia;CHE Yin-chao(School of Information Engineering,Huanghe Science and Technology University,Zhengzhou Henan 450046,China;College of Information and Management Science,Henan Agricultural University,Zhengzhou Henan 450046,China)

机构地区：[1]黄河科技学院信息工程学院,河南郑州450046 [2]河南农业大学信息与管理科学学院,河南郑州450046

出　　处：《计算机仿真》2023年第6期240-244,共5页Computer Simulation

基　　金：河南省教育厅2019年品牌专业建设项目(ZLG201903)。

摘　　要：光谱图像包含较多的空间样本信息,混合像元问题严重,同物异谱间影响明显,相邻波段的强相关性也导致图像高维特征易出现冗余,直接影响图像关联特征挖掘效果。为此,提出基于小样本分类的光谱图像关联特征挖掘方法。采用蒙特卡罗算法对光谱图像降维,并结合稀疏与低秩矩阵分解方法完成光谱图像去噪;基于此,建立光谱图像灰度共生矩阵,提取光谱图像角二阶矩、相关性、对比度、熵、相异性以及逆差矩光谱关联特征;通过支持向量机对获取的关联特征完成光谱图像的小样本分类,完成光谱图像的关联特征挖掘。实验结果表明,研究方法的光谱图像特征挖掘错误率低于2%,同类特征可显著聚类成团,且图像去噪效果明显。The spectral image contains more spatial sample information,the mixed pixel problem is serious,the in-fluence between the same object and different spectra is obvious,and the strong correlation of adjacent bands also leads to the redundancy of high-dimensional features of the image,which directly affects the effect of image correlation feature mining.In this paper,a new method of spectral image association feature mining based on small sample classification is proposed.The Monte Carlo algorithm is used to reduce the dimension of spectral image,and the sparse and low rank matrix decomposition method is used to denoise the spectral image;Based on this,the gray level co-occurrence matrix of spectral image is established,and the spectral correlation features of angular second mo-ment,correlation,contrast,entropy,dissimilarity and inverse moment of spectral image are extracted;Support vector machine(SVM)is used to classify the small samples of spectral images and to mine the correlation features of spec-tral images.The experimental results show that the error rate of spectral image feature mining is less than 2%,the same features can be clustered significantly,and the image denoising effect is obvious.

关 键 词：小样本分类 光谱图像 关联特征挖掘 图像去噪 蒙特卡罗算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]