Content based image retrieval using multi-quanitzation in color spaces

Abstract

Content based image retrievals are more and more important in processing image data in the multimedia age. Among these methods, color based image retrieval among them quantizes color distribution of images respectively and measure similarities among them[1,2,3]. Conventional algorithms use 3-dimensional histograms and turn out to be remarkably robust in variations such as a change in orientation, a shift in viewing position, a change in the scene background, partial occlusion, or even a radical change in shape[1]. However, the method neglects affinity with the human perception of color similarity (i.e., a red image and a blue image have the same dissimilarity from an orange one). In order to correct this discrepancy, Harfer uses a weighing factor, but this method has a higher computational complexity than the histogram intersection[4]. Another problem is that the number of bins in a histogram is quite large[4,5,6]. Particularly, it is hard to quantize color space finely with a small number of bin numbers because of quantization structure limits. In order to correct the demerit, 1-dimentional histogram, which independently quantizes three coordinates, is used. This method sometimes produces a color classification mistake which retrieves the worst image[7]. Multi-quantization which consists of three finely quantized 1-dimensional histograms and one coarsely quantized 3-dimensional histogram is proposed as a solution[8]. This method considers an adjacent color without additional special calculation complexity and has a good result in similar image retrievals and sub-image retrievals. Particularly in the cases that color distribution peaks in some regions such as opponent color space, HSV color space, YCbCr color space and so on. Using finely divided color spaced quantization, better results are obtained[9]. Therefore, multi-quatization, which quantizes color space finely with a small number of bins without color classification mistakes, has better results than...