This thesis introduces a framework for analyzing the underlying visual similarities among art images. Three types of art related images are investigated. They are: (1) painterly rendered images which are ``painting-like" images generated by painterly rendering algorithms; (2) painting images which are digitized copies of paintings; (3) graphic design images which are graphic art used in various logos, trademarks and symbols. However, the main focus of this thesis is given to the later two types of ``real" art images: \ie painting and graphic design images.
In the proposed framework, image features used for analyzing art images are defined to ``translate" qualitative art concepts or principles into quantitative numerical form. And then a series of Self-Organizing Map (SOM) based methods are introduced to discover, analyze and visualize the underlying visual similarities of art images. Furthermore, two Self-Organizing Map Best Matching Unit Entropy (SOM-BMU Entropy) based approaches are proposed to compare and visualize the data similarity of multiple sets of images. Based on the proposed framework, applications such as reverse painterly rendering, painting artistic influence analysis and logo similarity retrieval are also presented.
Different from previous art imaging systems, the proposed framework aims to bridge the gaps between artistic concepts, image features and numeric solutions, which allows the numeric analysis results to be explained by art concepts and to be visualize easily. Therefore, using the proposed framework, art image data can be better organized and explored; visual similarities of art images can be better understood and explained.