A Novel Subspace Clustering Method for Dictionary Design

Abstract

In this work, we propose a novel subspace clustering method for learning dictionaries from data. In our method, we seek to identify the subspaces where our data lives and find an orthogonal set of vectors that spans each of those subspaces. We use an Orthogonal Subspace Pursuit (OSP) decomposition method to identify the subspaces where the observation data lives, followed by a clustering operation to identify all observations that lives in the same subspace. This work is motivated by the need for faster dictionary training methods in blind source separation (BSS) of audio sources. We show that our dictionary design method offer considerable computational savings when compared to the K-SVD[1] for similar performance. Furthermore, the training method also offers better generalizability when evaluated on data beyond the training set, and consequently is well suited for continuously changing audio scenes.