A fast dictionary-learning-based classification scheme using undercomplete dictionaries

In dictionary-learning-based classification methods, a given data point is classified based on its representation over one or possibly more learned dictionaries. The goal is to find dictionaries that minimize the classification error. Previous works aimed to train dictionaries with representation and classification powers by using overcomplete dictionaries and sparse coding. These approaches are computationally expensive and do not scale readily to problems with high dimensional data. This paper presents a dictionary-learning-based classification method with the primary goal of classification and not representation. We propose to train multiple undercomplete dictionaries (one for each class of the problem). Each dictionary approximates the given test data, and the one with the lowest reconstruction error determines the class. Singular value decomposition (SVD) is used to obtain a straightforward algorithm for the resulted optimization problem. Simulation results show that our method achieves a higher accuracy compared with a number of successful sparse representation based classification methods, while having a significantly lower computational cost.

Signal Processing