Fast Selection of Linear Features in Image Data

Abstract

Identifying a small number of useful features for learning from images requires processing large amounts of data and may be very time consuming. The standard approach is to compute many features from training data and then select a subset of the features that is useful for the task at hand. We propose an algorithm for selecting these features efficiently for the case in which the features are linear. Our approach is based on Principal Components dimensionality reduction applied to the training data. It is shown that the computation of linear features as well as the task of pruning redundant features can all be performed in a compact, reduced representation. Specifically, redundant features are detected and removed using the classic Factorization algorithm, applied in the reduced space.