An Analysis of Face Portraiture: Feature Selection and Identification

Renaissance portraits were depictions of some important people of those times. These encompass a wide range of art works such as sculptures, death masks, etc. Apart from being used for a variety of dynastic and commemorative purposes, they were used to depict individuals often to convey an aura of power, beauty or other abstract qualities. A large number of these portraits, however, have lost the identities of their subjects through the fortunes of time. Analysis of faces in these portraits can provide valuable dynastical information in addition to enriching personal details of the depicted sitter. Traditionally, identification of many of these portraits has been limited to often quite variable personal opinion.  In this work, we evaluate the application of face recognition technology to portrait art and in turn aid art historians by providing a quantitative source of evidence to help answer questions regarding subject identity and artists' styles. Due to subjective interpretations of artists, portraits of the same sitter can vary from artist to artist. This results in considerable variability in the renditions, which has to be accounted for by the face recognition algorithms. Further, we have only a few images owing to the lack of authenticity and cost involved in their procurement. Given these challenges, we are required to choose a set of features that possess high discriminating power across artists/sitters.  By means of random subspace ensemble learning and statistical permutation tests, we describe an algorithm that is capable of automatically learning a set of weighted features characteristic of an artist's style, wherein the weights denote the relative importance of the features. The learned features is used to obtain a measure of similarity between image pairs under consideration, to yield match scores (depicting same person) and non-match scores (depicting different people) as appropriate. We also learn the distribution of match and non-match scores, referred to as the ``Portrait Feature Space" across various artists in order to serve  identification tasks. The sparse similarity scores computed between image pairs by an artist are cross-validated using a non-parametric statistical test called the Siegel-Tukey test that is known to work well for small samples. We also cross-validate the relatively larger set of similarity scores across artists using standard cross-validation methods.  Through statistical hypothesis tests, we analyze identification paradigms with respect to the learned Portrait Feature Space to arrive at appropriate conclusions.