基于局部线性重构的全景高光谱影像低维可视化

论文提出一种基于局部线性重构的半监督学习全景高光谱影像扩散坐标系延拓方法.基于半监督学习思想,利用ε-net近似逼近全景高光谱影像的特征空间,并计算出对应的扩散几何坐标,进而基于局部线性性质在ε-net所逼近的特征空间上局部线性重构出外采样点的扩散几何坐标.相对于基于坐标系配准达到大尺度高光谱影像可视化的方法,半监督学习范式能够给出全景高光谱影像更为一致的坐标表示.

Low-dimensional Visualization of Full Scene Hyperspectral Imagery Based on Locally Linear Reconstruction

In this paper, the diffusion coordinate system extension approach on full scene hyperspectral imagery is pro- posed, which is based on local linear assumption and semi-supervised learning paradigm. Abiding by semi-supervised learning paradigm, the feature space of full scene hyperspectral imagery is approximated by e-net sampling technique and the diffusion geometric coordinates are computed accordingly. As for the large-scale out-of-samples, these diffusion geometric coordinates are reconstructed based on the local linear reconstruction algorithm on the sampled subspace. Comparing with the diffusion coordinate system transformation approach studied in previous work, this semi-supervised learning approach could provide more consistent coordinates representation.