Development of a machine learning model for extracting image prominent colors

Abstract

AbstractAlthough an image can consist of myriad different colors, due to the limitations of the human visual system and the mechanism of selective visual attention, only a limited number of them are prominent, that is, they stand out or are noticeable at first sight. In this article, a framework for building a model for extracting image prominent colors based on machine learning is presented. The model is learned on human‐extracted themes of prominent colors and uses numerous features, which were defined based on the properties of human visual system. For the purpose of this study, we constructed a database of images with their associated human‐extracted themes of prominent colors, which are open to the public and available to other researchers. The analysis of observers' data shows a high interobserver agreement on prominent color categories as well as high diversity of prominent colors. According to our results, the most influential factors on the perception of prominent colors are associated with color coverage (which should be adjusted with a saliency map), color properties—lightness and chroma, and diversity of colors. To the best of our knowledge, this is the first study on image prominent colors and first attempt to extract them with a model trained on the real data. The presented model has a high practical importance, since it can be used for extracting image colors in different scenarios, for example, for automatic color design, image categorization, as a descriptor in content‐based image retrieval, and image content analysis frameworks.