一种小样本情境下的高光谱图像分类算法  

A hyperspectral image classification algorithm for few shot contexts

作　　者：张裕 陈立伟[1] 崔颖[1] ZHANG Yu;CHEN Liwei;CUI Ying(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《应用科技》2024年第3期135-140,共6页Applied Science and Technology

基　　金：国家自然科学基金项目(61675051).

摘　　要：Gabor滤波器是一种常见的空间特征提取技术,在针对高光谱图像分类中已标记样本稀缺的问题上,该算法通过设置不同方向的多个3D-Gabor滤波器,生成大量多视图。在多视图数据基础上生成多个图连接实现标签传播,将多个图标签传播后的分类结果融合得到预测标结果。而超像素主成分分析法算法则是一种简单但非常有效的无监督特征提取方法,将预测结果与加入了超像素主成分分析法的分类器相加权融合得到更为准确的分类结果。将算法在3个数据集上进行仿真实验,结果表明通过应用Gabor滤波器的传统高光谱图像分类算法存在运算量大且耗时长,而该算法能够在保证精度的同时有效减少计算及时间上的消耗,节约成本。Gabor filter is a common spatial feature extraction technique.In order to address the problem of sparse labelled samples in hyperspectral image classification,the algorithm in this paper generates a large number of multiple views by setting multiple 3D-Gabor filters in different directions.The algorithm generates multiple graph connections on the basis of the multi-view data to achieve label propagation,and fuses the classification results of multiple graph labels propagated to obtain the predicted label results.The superpixel principal component analysis(Super PCA)algorithm is a simple but very effective unsupervised feature extraction method,where the prediction results are weighted and fused with the classifier incorporating Super PCA to obtain more accurate classification results.Simulations of this algorithm on three datasets show that traditional hyperspectral image classification algorithms using Gabor filters are computationally intensive and time-consuming,whereas this algorithm can reduce computational and time consumption while ensuring accuracy and cost savings.

关 键 词：小样本 高光谱图像分类 3D-Gabor滤波器 多视图 标签传播 超像素分割 半监督学习 主动学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]