The Classification of Style in Fine-Art Painting

The Classification of Style in Fine-art Painting by Thomas Edward Lombardi Submitted in partial fulfillment of the requirements for the degree of Doctor of Professional Studies in Computing August 2005 The computer science approaches to the classification of painting concentrate on problems of attribution. While this goal is certainly worthy of pursuit, there are other valid tasks related to the classification of painting including the identification of period styles, the description of styles, and the analysis of the relationship between different painting styles. This dissertation proposed and developed a general approach to the classification of style and achieved this goal using a semantically-relevant feature set. The resulting automated painting analysis system supports the following tasks: recognize painting styles, identify key relationships between styles, outline the basis for style proximity, and evaluate and visualize classification results. The study initially conducted a review of the features currently applied to this domain and implemented these features, supplementing them with commonly used features in image retrieval applications. The study evaluated these features for classification accuracy, speed, storage space, and semantic relevance, and mapped the features considered to formal elements discussed in the domain, including light, line, texture, and color. In particular, the study successfully employed several color features not previously applied to painting classification, such as color autocorrelograms and dynamic spatial chromatic histograms. The dissertation proposed and developed a palette description feature for describing the color content of paintings. In tests, the palette description feature classified style as well as comparable color features. The study evaluated the features against two databases of paintings using a variety of supervised and unsupervised classification techniques including k-nearest neighbor, hierarchical clustering, self-organizing maps, and multidimensional scaling. In summary, the dissertation proposed and developed a theoretical style center and variance as both an analytical tool and an evaluation technique for classification accuracy. A style description ratio based on the theoretical style center and variance served as a reliable basis for the evaluation of classification results.