Image retrieval using spatiograms of colors quantized by Gaussian Mixture Models

In this paper we propose a novel image representation method that characterizes an image as a spatiogram--a generalized histogram--of colors quantized by Gaussian Mixture Models (GMMs). First, we quantize the color space using a GMM, which is learned by the Expectation-Maximization (EM) algorithm from the training images. The number of Gaussian components (i.e., the number of quantized color bins) is determined automatically according to the Bayesian Information Criterion (BIC). Second, we incorporate the spatiogram representation with the quantized Gaussian mixture color model. Intuitively, a spatiogram is a histogram in which the distribution of colors is spatially weighted by the locations of the pixels contributing to each color bin. We have modified the spatiogram representation to fit our framework, which employs Gaussian color components instead of discrete color bins. Finally, the comparison between two images is achieved by measuring the similarity between two spatiograms, for which purpose we propose a new measurement adopting the Jensen–Shannon Divergence (JSD). We applied the new image representation and comparison method to the image retrieval task. The experiments on several publicly available COREL image datasets demonstrate the effectiveness of our proposed image representation for image retrieval.