Color Harmony Based Computational Aesthetics Of Image And Its Application

Color harmony is an important factor for image aesthetics assessment. Although plenty of color harmony theories are proposed by artists and scientists, there is little firm consensus and ambiguous definition among them, or even contradictory between them. Moreover, these theories or metrics were always fixed, no matter what kind of images they are. In this paper, we attempt to develop a model that can adjust parameters according to samples.Firstly, we proposed a color harmony model based on statistical features of color hue, whose parameters can be adjust according to datasets. We extract features on dominant color ranges clustered via Mean-Shift [1]algorithm. Generative models can be trained on these features and used to make the aesthetics quality assessment. Experimental results on our datasets show that our method outperforms the conventional rule based image aesthetics assessment approach.Secondly, considering the limitation of the only features about color hue, we propose a hierarchical unsupervised learning approach to learn the compatible color combinations from large dataset. By using this generative color harmony model, we attempt to uncover the underlying principles that generate pleasing color combinations based on natural images. We learn co-current color groups from natural images with LDA(latent Dirichilet allocation[3]) and train the GMM(Gaussian Mixture Model)[2] to represent relationships between color groups. With this hierarchical color harmony model, we make the aesthetic quality assessment and experimental results show this method not only outperforms the method mentioned above but also reveal more information about color harmony.Finally, we implement a prototype of image retrieval system,"Beauty Hunter", for "beautiful" images based on models mentioned above.