Visual Stylometry Using Background Selection And Wavelet-HMT-Based Fisher Information Distances For Attribution And Dating Of Impressionist Paintings

摘要

The new field of visual stylometry proposes to apply mathematical and statistical tools to high-resolution images of artworks to produce a quantitative description of each work's style, or of stylistic differences between works. Such quantitative evidence regarding artistic style can then assist in addressing open art historical questions, including those of a particular work's authorship or date of creation. In this paper, we develop a new technique for visual stylometry on impressionist and/or post-impressionist paintings. We focus on the background of each painting for our analysis, hypothesizing that only this bears the signature of the artist's hand. We then introduce a new wavelet-Hidden-Markov-Tree-based Fisher information distance as a metric of stylistic similarity between brushwork samples. Tests on two datasets consisting of over 100 impressionist and post-impressionist paintings by Van Gogh and contemporaries show that an unsupervised representation of the paintings according to this new metric tends to cluster the paintings by author and, within an author, by time period. Classifying paintings under leave-one-out cross-validation using coordinates in this unsupervised representation gave accuracies for artist classification of 87.7% and 85.0% for the two datasets and for time period classification of 81.5% and 75.4%, a substantial improvement over previously developed stylistic distance metrics for paintings.