Towards aesthetics of image: A Bayesian framework for color harmony modeling

Abstract

Color harmony is one of the most important features that determine the aesthetics quality of images. The existing color harmony models can be roughly divided into two groups: empirical based models defined by artists and designers and learning based models established by discovering underlying patterns from the collected samples. However, these two types of methods treat the model of color harmony from two distinct aspects and no consolidated framework exists to ensure that the benefits can be easily reaped from each other. To overcome this problem, we proposed a Bayesian framework for constructing the color harmony model, in which the empirical rules defined by the artists or designers serve as a prior and the patterns discovered by machine learning methods from the training images are modeled as the likelihood. Particularly, under this framework, we integrate two empirical (Matsuda and Moon–Spencer) color harmony models into a latent Dirichlet allocation (LDA) based learning procedure to train the color harmony model. The experimental results on a public dataset show that the proposed Bayesian based color harmony model is superior to the conventional color harmony models in respect of the image aesthetics assessment.