Scholarly record

EXTRACTION OF MAIN COLORS FROM A COLOR DIGITAL IMAGE

A. Ciobanu, M. Costin, T. Barbu

Abstract

In order to prepare minimal feature vectors to be used in a Content Based Image Recognition (CBIR) system, we are proposing methods that can accurately identify the main colors contained in a digital image. The methods use the “LAB” color space properties that allow for all the colors of the same intensity or luminance “L” to be placed uniformly in the same “ab” plane. Basically, RGB images are first preprocessed: trimmed, scaled and transformed in the LAB color space. Then (a,b) pairs, corresponding to large pixel counts in the image, are detected using two or more planar histograms. One of our approaches determines individual (L,a,b) triplets (each defining precisely a color) belonging to dark and light colors, and treats separately black, white and gray tones. It makes three lists of colors and then reduces the number of colors in each list by merging close shades of colors until only several main colors are listed globally for the processed image. Another advanced approach finds (a,b) pairs in five preset bands of luminance “L”, resulting in five lists of colors that are also reduced in number by merging shades of colors in each band. The top colors in each band (or, if necessary, top 3 colors) are then extracted as color features for an image. In both approaches, the initial image is reconstructed with the identified main colors for visual inspection and, as a byproduct, for extracting edges with a Canny detector. Results show potential for further use in image classification or retrieval, as well as for segmentation of images in regions.

Publication details

Back to publication list