Image Statistics for Clustering Paintings According to their Visual Appearance

Untrained observers readily cluster paintings from different art periods into distinct groups according to their overall visual appearance or “look” [WCF08]. These clusters are typically influenced by both the content of the paintings (e.g. portrait, landscape, still-life, etc.), and stylistic considerations (e.g. the “flat” appearance of Gothic paintings, or the distinctive use of colour in Fauve works). Here we aim to identify a set of image measurements that can capture this “naïve visual impression of art”, and use these features to automatically cluster a database of images of paintings into appearance-based groups, much like an untrained observer. We combine a wide range of features from simple colour statistics, through mid-level spatial features to high-level properties, such as the output of face-detection algorithms, which are intended to correlate with semantic content. Together these features yield clusters of images that look similar to one another despite differences in historical period and content. In addition, we tested the performance of the feature library in several classification tasks yielding good results. Our work could be applied as a curatorial or research aid, and also provides insight into the image attributes that untrained subjects may attend to when judging works of art.