Мethod for determining the level of similarity of images by color component based on correlation analysis and color space quantization

Abstract

The color component of an image is the basic information about the image. Color determines the mood of an image, its concept, and can also serve as a semantic accent. Obtaining information about the color component of an image is necessary when solving various kinds of tasks. For example, the task of finding similar images according to the content, as information about the color component can be combined with other metadata of the image and more fully describe its concept. In this paper authors describe the physical representation of color, the most popular digital representations of color, ways of extracting color components of an image, such as color histograms, quantization, coherent vectors and K -means, and ways of comparing color components of images, such as Euclidean distance, cosine distance, Manhattan distance, Chebyshev distance, Jaccard index, correlation analysis, Pearson's agreement test and Bhattacharya distance. It is suggested that the use of histogram comparison and vector comparison methods is acceptable for comparing the color components of images, since color histograms can be written in vector form. The authors conduct an experiment in which a human expert reviews combinations of color component acquisition and comparison methods and completes a test protocol based on the criteria highlighted by the authors. The criteria were based on the human interpretability of the results of the methods of determining and comparing the color component. On the basis of the experiment it has been determined that the best of the methods of comparison of color components of the definition images considered in the article is the correlation analysis of color histograms of images, which have been extracted in a quantized color space, because this approach meets all the criteria for testing and keeps the greatest amount of information about the color component.