Color clustering techniques for color-content-based image retrieval from image databases

Abstract

Image retrieval based on color content is an auxiliary function for traditional text-annotated image databases. Most color-based image retrieval systems adopt color histograms as the feature of color content. One of the most important steps in these systems is to reduce histogram dimensions with the least loss in color content. A good clustering technique is vital for this purpose. This paper examines the color conservation property by applying different clustering techniques in perceptually uniform color spaces and different images. For studying color spaces, the perceptual uniform spaces, such as Mathematical Transformation to Munsell system (MTM) and C.I.E. L*a*b*, are investigated. For evaluating clustering techniques, the equalized quantization approach, the hierarchical clustering approach, and the Color-Naming-System (CNS) supervised clustering approach are studied. For analyzing color loss, the error bound, the quantized error in color space conversion, and the average quantized error of 400 color images are explored. An image retrieval application based on color content is shown to demonstrate the difference in applying these clustering techniques. These simulation results suggest that good clustering techniques usually lead to more effective retrieval.