Learning overcomplete dictionaries based on parallel atom-updating

Mostafa Sadeghi (1) , Massoud Babaie-Zadeh (1) , Christian Jutten (2)

1 EE department of SUT [Tehran] - Department of Electrical Engineering [Tehran]
2 GIPSA-VIBS - GIPSA - Vision and Brain Signal Processing

In this paper we propose a fast and efficient algorithm for learning overcomplete dictionaries. The proposed algorithm is indeed an alternative to the well-known K-Singular Value Decomposition (K-SVD) algorithm. The main drawback of K-SVD is its high computational load especially in high- dimensional problems. This is due to the fact that in the dictionary update stage of this algorithm an SVD is performed to update each column of the dictionary. Our proposed algorithm avoids performing SVD and instead uses a special form of alternating minimization. In this way, as our simulations on both synthetic and real data show, our algorithm outperforms K-SVD in both computational load and the quality of the results.

Mots clés: Sparse approximation; compressive sensing; dictionary learning; alternative minimization

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Mostafa Sadeghi, Massoud Babaie-Zadeh, Christian Jutten. Learning overcomplete dictionaries based on parallel atom-updating. MLSP 2013 - IEEE 23rd International Workshop on Machine Learning for Signal Processing, Sep 2013, Southampton, United Kingdom. 5 p. ⟨hal-00867170⟩