Multivariate dictionary learning and shift & 2D rotation invariant sparse coding

Quentin Barthélemy (1, 2) , Anthony Larue (2) , Aurélien Mayoue (2) , David Mercier (3) , Jerome I. Mars (1)

1 GIPSA-SIGMAPHY - GIPSA - Signal Images Physique
2 LOAD (CEA, LIST) - Laboratoire Outils d'Analyse des Données (CEA, LIST)
3 LIMA (CEA, LIST) - Laboratoire Information, Modèles, Apprentissage (CEA, LIST)

In this article, we present a new tool for sparse coding : Multivariate DLA which empirically learns the characteristic patterns associated to a multivariate signals set. Once learned, Multivariate OMP approximates sparsely any signal of this considered set. These methods are specified to the 2D rotation-invariant case. Shift and rotation invariant cases induce a compact learned dictionary. Our methods are applied to 2D handwritten data in order to extract the elementary features of this signals set.

Mots clés: Rotation-invariant; Multivariate DLA; Levenberg-Marquardt; machine learning; Handwritten characters; artificial intelligence; dictionary learning; statistical analysis; Multivariate OMP; Shift-invariant; Handwritten characters.; Sparse coding; Online learning; signal processing

Domaines: Traitement du signal et de l'image [eess.SP]; Recherche d'information [cs.IR]; Apprentissage [cs.LG]

Quentin Barthélemy, Anthony Larue, Aurélien Mayoue, David Mercier, Jerome I. Mars. Multivariate dictionary learning and shift & 2D rotation invariant sparse coding. SSP 2011 - 2011 IEEE Workshop on Statistical Signal Processing, Jun 2011, Nice, France. pp.649-652, ⟨10.1109/SSP.2011.5967783⟩. ⟨hal-00625352⟩