Learning unions of orthonormal bases with thresholded singular value decomposition

摘要: We propose a new method to learn overcomplete dictionaries for sparse coding structured as unions of orthonormal bases. The interest such structure is manifold. Indeed, it seems that many signals or images can be modeled the superimposition several layers with decompositions in Moreover, dictionaries, efficient block coordinate relaxation (BCR) algorithm used compute decompositions. show possible design an iterative learning produces dictionary required structure. Each step based on coefficients estimation, using variant BCR, followed by update one chosen basis, singular value decomposition. assess experimentally how well recovers may not have structure, and what extent noise level disturbing factor.