Sparse Coded Spatial Features from Spectral Images
Arto Kaarna · 2006
In this study we apply the non-negative matrix factorization (NMF) to extract features from a spectral image. NMF features are sparse, they carry more localized information than the principal component analysis (PCA) features which are holistic in nature. The sparseness of the NMF feature vectors were controlled with a sparseness constraint. This approach allowed even more sparse feature vectors. The experiments indicate that NMF is a suitable tool also with spectral images in extracting sparse feature vectors even though spectral images have their own special properties compared e.g. to facial image databases.