Machine learning for painting classification: State of the art and future prospects

Abstract

Being able to classify paintings by painter, style or genre is something most art historians can do fairly accurately. Recently attempts have been made to do this with machine learning classifiers. This paper provides a look at the state of the art of classifying paintings using machine learning classifiers and feature sets. State of the art classifiers and feature sets will be discussed for various classification approaches. Per classification approach I discuss applications, benefits, drawbacks, practical usefulness and performance. Analysis of performance is based on results from literature. From all of this I conclude that feature sets specifically made for painting classification should be created for classifying by painter. Based on current literature it seems that a feature set with features created by a convolutional neural network (CNN) could work well for classifying by style. For classifiers, it seems that researchers often use the support vector machine classifier without testing performance of other classifiers. More research should be done on the performance of classifiers other than support vector machines on the subject of painting classification. Other approaches and their problems are also discussed briefly.