Search in art image collections based on color semantics (R-1875)

One of the most felicitous analogies for presenting the existing semantic gap in area of Content-Based Image Retrieval (CBIR) can be found in "The Hitch-Hiker┌s Guide to Galaxy" by Douglas Adams. In this story, a group of hyper-intelligent pan-dimensional beings demand to learn the "Answer to Life, the Universe, and Everything" from the supercomputer Deep Thought, specially built for this purpose. It takes Deep Thought 7½ million years to compute and check the answer, which turns out to be "42". The efforts of covering the semantic gap in CBIR are turned to avoid these misunderstanding between human perceiving and the ways of communications and computer manner of low-level representations. As it is mentioned in [Castelli et al] the user questions in image search are partitioned into three main levels: Low level ├ this level includes basic perceptual features of visual content (dominant colors, color distribution, texture pattern, etc.). Low-level queries and analysis can support the retrieval of art images in order to seek some specifics or common characteristics between artists, schools or movements. Intermediate level ├ this level forms next step of extraction from visual content, connected with emotional perceiving of the images, which usually is difficult to express in rational and textual terms. The visual art is area, where these features play significant role. Typical features in this level are color contrasts, because one of the goals of the painting is to produce specific psychological effects in the observer, which are achieved with different arrangements of colors. High level ├ this level includes queries according to rational criterions. In many cases the image itself does not contain information which would be sufficient to extract some of the characteristics. For this reason current high-level semantic systems still use huge amount of manual annotation. Different features' levels imply different ways for communication between the user and the CBIR system. When a system uses low-level properties such as color percentages, color layout, and textures, (see e.g. QBIC, developed by IBM) the queries do not need to be described in words. When working with such systems, the user can select a sample image and the system returns all images which are "similar" to it. For systems, which operate with high level features, only choosing a sample or drawing a sketch and search similar characteristics is not sufficient, even because such system has to "know" which of characteristics are targeted by the user. There are two mutually connected tasks in this area: ┘ Defining features and terms, which present certain effect or criterion and describing correlation between defined concepts; ┘ Finding appropriate algorithms for generating metadata, which alone or in combination with present terminal features and terms will allow improved image search as well as proposing adequate methods and tools for establishing belonging of a sample to same concept. In order to present properly concepts and their correlation between low and intermediate levels as well as the connections to the high level, every system usually creates its own dataset. This allows implementing the specific elements of the used methods and tools.

Keywords:DATA MINING

Disciplines:Economics and business