A Color and Composition-based Image Retrieval System by Image-Query Drawing Method with Color-Impression Database

In this paper, we propose a color and composition-based image retrieval system with an image-query drawing method for color-impression database. The aim of this system to provide a method of image-query expression in image retrieval for representing his/her imaginary image. This system retrieves image data similar to an image-query drawn as a grid image corresponding to the overall color combination and composition features on the searcher’s imagination. The most significant features of this system are that it sets a color-impression database storing expert’s knowledge about impression words and their corresponding to color combinations, and an image-query drawing function on the query part of image database. Thus, the searcher chooses three color combinations as color feature and paints a N*N grid image by the color combination to reflect composition feature in the image-query. All images stored into the image database also have metadata containing keywords that describe image’s contents to categorize the images. A semantic space of the images is selected by the keywords to retrieve similar images in the specific domain. In this time, we adopted color combinations with impression words based on psychology-color research knowledge into the color-impression database. The data stored in the color-impression database can be replaced with other color-impression knowledge. So that, a searcher can paint each grid with various color usage of favorite painters and artists. For image retrieval, this system also compares color histograms, that consist of representative colors, on each grid between image-query and all images that are divided into N*N grids to calculate local similarity between grids. The sum of the local similarity on the color histograms corresponds with global similarity between the image-query and images stored into the image database. Godlove’s color difference formula in HSV color space is applied to the similarity calculation. We show the details of this method and its effectiveness and feasibility in this paper.