Abstract
This paper presents a novel content-based image retrieval technique based on Gaussian mixture probability model. The proposed technique provides the solution toward matching arbitrary images based on color, shape and texture. Glyph structure of the image, which inclines on the shape and texture attributes, is modeled and used for content matching. Gaussian mixture model is applied to quantify the glyph structure in terms of its parameters. The formed probability density functions based on the glyph structure are refined using expectation maximization. Finally, the parameters yielded by the Gaussian mixture model allow us to perform comparison between arbitrary images based on their semantic details. It is concluded from the experimental results that relatively similar images have comparable parameters while the parameters of discordant images deviate with each other. In this way, for a certain arbitrary image, the set of resembling images is obtained from a large image base. In addition, the results show that this set is narrowed or broadened on the basis of a divergence ratio which marks the functional difference between the parameters of the images being compared.