Stylometrics of artwork: uses and limitations

Abstract

A number of digital image analysis techniques have been developed in recent years to address art historical questions. These techniques allow non-destructive analyses of art images that can target outstanding problems of attribution, historical ordering, and other stylistic dimensions. However, great care must be taken in designing the comparisons to which these techniques are applied. In this paper, we review recent work by our lab and by others aimed at establishing a toolbox of stylometrics, and we discuss some of the uses and limitations of these methods. We describe a technique that provides robust classification of authentic drawings by Pieter Bruegel the Elder, and we demonstrate new techniques for art historical analysis applied to the works of other masters. Specifically, we demonstrate the use of two low-level statistics (the slope of the log amplitude spectrum and color histogram correlation) to analyze the works of Picasso and Braque. Finally, we show that face detection and recognition techniques may play a useful role in the attribution of works of art. The rationale for employing vision coding-like methods (e.g., sparse coding) in stylometry is also reviewed. We conclude that generic authentication tools are unlikely to provide reliable stylometric predictions but that with careful construction of comparison sets -which we believe must be done in close collaboration with art historians -these techniques provide important predictions that can be weighed against other art historical evidence. We argue further that concurrent predictions derived from analysis of many independent dimensions of image data (e.g., color, luminance, and spatial statistics) provide the strongest evidence for digital stylometric determinations. MOTIVATIONDespite decades, and in some cases centuries, of analysis, art historians continue to grapple with questions of authenticity and attribution of important works of art. [1][2][3][4][5] Traditionally, methods such as connoisseurship and provenance studies have provided the primary means for attributing works to particular artists. However, since at

Statistics, vision, and the analysis of artistic style

“…However, the addition of works in the Expressionist and Surrealist style causes performance to drop dramatically. Other work 9 confirms the potential utility of these low level statistics for genre classification. While Cubist, Classicist, and Impressionist works may each be relatively distinct in terms of these low level features, it is clear that this approach does not scale up to finer grained distinctions.…”

Empirical Mode Decomposition Analysis for Visual Stylometry

“…Nevertheless, both approaches produce feature vectors that work well on this data set. Interestingly, the sparse coding classification approach of [13] also failed to correctly identify drawing 011, and in general accuracy rates were somewhat lower using sparse coding features as opposed to using the EMD features presented here. This comparison establishes the utility of the EMD features in that they provide results at least as good or better than those obtained with previous methods on this data set.…”

“…Examples include box counting analysis, Fourier spectra, wavelet spectra, brushstroke analysis, etc. (see, e.g., [2], [13], [16], [20], [35], [46]). The different approaches appear to articulate different stylistic elements so that, in short, there is much less agreement on the fundamental analytic elements in stylometry.…”

Graph-based methods for the automatic annotation and retrieval of art prints

“…The main focus of the papers is on the artistic identification problem, where the goal is to classify original and fake paintings of a given artist [4,32,39] or to produce stylistic analysis of paintings [20,24,25]. Most of the methods above can be regarded as adaptations from the content-based image retrieval systems [14], where the emphasis is placed on the characterization of brush strokes using texture or color.…”