Abstract
Perceived similarity between images is an achievement of complex processes of vision, of cognitions and emotions. It may be influenced by the context and the content of images, semantic associations and the intention of the observer, to name a few. But simple, low-level features are used successfully to compare images. For instance, color histograms are used in content-based image retrieval (CBIR) systems and frequently result in perceptual and often semantically similar images.
Our goal was to create an image indexing system based on some of the known properties of human vision. A color space (DKL) found to underlie the second stage of human color vision (Krauskopf et al., 1982) was used to create color histograms. In the DKL space color bins, independent of luminance, were created using a logarithmic spacing which reflects the granularity of higher order color perception. The frequency and the average luminance level of the color bins were stored in a chromaticity and a luminance histogram respectively. A spatial index encodes — in analogy to transformation in the visual cortex — information about orientation and spatial frequencies of the images. The information was extracted using a 2-dimensional discrete Fourier transformation (DFT). The Fourier spectrum was divided into logarithmic-radial bins representing contrasts of distinct orientation and spatial frequency ranges.
In three experiments we quantitatively measured the relationship between the similarity order induced by the indices (and index combinations) and perceived similarity in a 2-alternative forced-choice design. For the experiments we used a large, commercially available database (COREL) of 60,000 digitized photographs. In contrast to previous evaluation approaches we measured the relationship not only for the few best matching images but for relatively distinct images, too.
Krauskopf, J., Williams, D. R., and Heeley, D. W. (1982). Cardinal directions of color space. Vision Research, 22, 1123–1131.