Improved separable dictionary learning

Abstract

Due to the extensive applications of sparse representation, the algorithms of dictionary learning have received widespread attentions. In this paper, we proposed a new separable dictionary learning algorithm: improved separable dictionary learning (ISeDiL). Different with the traditional separable learning methods, we divided the learning procedure into two steps: sparse coding and dictionary optimization. In the first step, we used the separable fast iterative shrinkage-thresholding algorithm (SFISTA) to obtain the sparse coefficient matrices. In the second step, we projected the dictionaries to oblique manifold, and used the conjugate gradient method to optimize the dictionaries. By projecting the dictionaries to oblique manifold, the dictionaries can be optimized as a whole part. At last, we showed the experimental procedure, and the de-noising results of the proposed algorithm.