Contextual Dominant Color Name Extraction for Web Image Search

Abstract

This paper addresses the problem of extracting perceptually dominant color names of images. Our approach is motivated by the principle that the pixels corresponding to one dominant color name identified by human are often context dependent, spatially connected and form a perceptually meaningful region. Our algorithm first learns the probabilistic mapping from a RGB color to a color name. Then, a double-threshold approach is utilized to determine the color name of a RGB pixel in a specific image by considering its neighboring pixels. This scheme effectively deals with the pixels ambiguously belonging to several dominant color names. Last, the saliency information is combined to extract perceptually dominant colors. Experiments on our labeled image data set and the Ebay image set demonstrate the effectiveness of our approach.