Using independent subspace analysis to build independent spectral representations of images

In this work we propose the use of independent subspace analysis (ISA) for selecting filters used to build image representations. ISA is an extension of independent component analysis (ICA), a technique employed to decompose an image into independent features. In ISA, complete independence of features is not required; features that possess some mutual dependence are associated in feature subspaces. A characteristic of the ISA model is that these subspaces enclose features of similar frequency and orientation. This, in turn, helps in determining a reduced set of filters to be employed in image classification. Here we address the task of classifying patches of textured images. Preliminary results here presented show that our proposed ISA criterion can attain performance comparable to other filter based classification schemes while resulting in a considerably smaller filter bank.