Classification of color combinations based on distance between color distributions

Color features of a given image can be represented by its color distribution, distribution of constituent colors in a color space. If distance between color distributions is defined, color features can be classified objectively according to their similarity. Though a digital image consists of a cast number of representative colors is required. The obtained color distribution, therefore is a discrete distribution of finite number of representative colors. Interpreting the relation between discrete distributions as a network flow problem, which is a kind of linear programming problem, the distance between discrete distributions is defined. Using the distance, hierarchical clustering analysis with furthest neighbor method and performed to classify several paintings and designs. The results well agreed with our observation on color features of the images. The notion of this distance will be applied to many fields, such as image retrieval, pattern recognition, including other various possibilities.