Incremental Compressive Pattern Classification

Compressive sensing (CS) is a technique based on a proven observation that if a signal is compressible, i.e. has a sparse representation in some basis, it can be reconstructed from only a very limited number of measurements. This idea, directly applicable to data compression, has been recently exploited in pattern classification. In this setting an input pattern is projected onto a sparsity basis spanned by a number of other, pre-labeled patterns (a training set), such that expectedly only those representing the same category as the tested pattern are assigned non-zero coefficients. Label of a novel pattern is then determined by generating K signals with attenuated coefficients of the basis patterns not sharing each of K possible labels, and picking the one minimizing the reconstruction error. Our contribution consists in adding recursion to the above scheme. As a result, a newly labeled pattern, if considered "good enough", is added to the training set. The following test patterns are classified in a sparse representation re-calculated under the compressive assumptions from this extended training set. The proposed approach, called "Incremental Compressive Pattern Classification" has been tested on several public image datasets showing superiority to the original algorithm and comparable or better performance than the alternative methods in terms of the classification accuracy.