A Shift Tolerant Dictionary Training Method

Traditional dictionary learning method work by vectorizing long signals, and training on the frames of the data, thereby restricting the learning to time-localized atoms. We study a shift-tolerant approach to learning dictionaries, whereby the features are learned by training on shifted versions of the signal of interest. We propose an optimized Subspace Clustering learning method to accommodate the larger training set for shift-tolerant training. We illustrate up to 50% improvement in sparsity on training data for the Subspace Clustering method, and the KSVD method [1] with only a few integer shifts. We demonstrate improvement in sparsity for data outside the training set, and show that the improved sparsity translates into improved source separation of instantaneous audio mixtures.