The Learning-Based Principal Component Analysis Technique in Low Resolution and High Resolution Spectral Images

Abstract

This paper presents a learning-based principal component analysis technique for accurate representation of spectral color in low and high resolution spectral images. Three learning techniques, LLE, ISOMAP, and regressive principal component analysis (PCA), are studied for this purpose. The basic concepts for the regressive PCA technique, which is computationally efficient and represents a combination of standard PCA and regression, are examined. To utilize dimensionality reduction techniques such as LLE and ISOMAP as parametric mapping procedures, the methods must be modified by combining them with a regression approach which provides data mapping from a low-dimensional space to the input space. The LLE, ISOMAP, and regressive PCA learning techniques are compared with standard PCA using low-resolution spectral images. We show that the LLE and ISOMAP approaches are computationally demanding and are not well suited to high resolution image analysis. Regressive and standard PCA are then used in a test with high resolution spectral images. The comparative study based on the S-CIELAB ΔE and RMSE employs regressive PCA measures to illustrate accurate color representation.