A new method for visual stylometry on impressionist paintings

A new emerging field, that of visual stylometry of art, proposes to apply image analysis and machine learning tools to high-resolution digital images of artwork in order to assist art connoisseurs in determining the painting´s likely creator. The premise is that each artist´s brushwork is likely to contain features that are characteristic of the artist´s unique habitual physical movements; these features could be identified and characterized through machine learning. In this paper, we describe a new technique for this problem. We extract, as features for our classifier, parameters of both Hidden Markov Tree models and linear predictor models of the painting´s wavelet coefficients. We then use the FINE dimensionality reduction technique [1] to produce an unsupervised low-dimensional embedding of the data. Tests on two dataset consisting of over 100 high-resolution digital images of impressionist paintings by Van Gogh and contemporaries shows good separation between paintings of Van Gogh and others is achieved via this unsupervised process. We further show (through comparison with the alternative) that our method benefits greatly from (1) using only background sections of each painting in our analysis, (2) the FINE technique, and (3) the use of both HMT and linear predictor features together in the same analysis. All three of these technique choices are new in this paper. We hope that our method can be a tool, used carefully in conjunction with the connoisseur´s expertise and other examinations, to determine a painting´s true authorship.