Dictionary learning-based classification of ink strokes in Vincent van Gogh's drawings

Abstract

Discriminative dictionary learning techniques are applied in an art historical context with the goal of automatic classification of an ink drawing based on the type of stroke. Two kinds of K-SVD-based multi-class classification are tested on van Gogh's drawings. First is a classical dictionary method where atoms are learned from pixels directly. A second feature-based dictionary method is introduced where atoms are learned based on features characterising collections of individual strokes. Results indicate that feature-based method provides better classification while incurring significantly less computational expense than the classical method. A multi-level feature-based method extends K-SVD for larger images. Such an automated classification would provide a new resource for scholars of van Gogh and students learning the art of drawing, be useful in automated photograph to computer drawing translation and point towards the more difficult problem of identifying painted brush strokes. Algorithms and data are provided to encourage modifications and extensions.