An improved EM algorithm for content based image retrieval

Content Based Image Retrieval (CBIR) mainly contains two phases: first, to represent an image; second, to measure the dissimilarity between two images. Expectation-Maximization (EM) is a popular algorithm for clustering Gauss mixtures for the image representation, but the greedy nature of EM make it hard to get an optimal model for CBIR. In this paper , we introduce an improved EM algorithm for clustering Gaussian Mixtures (GMs) to represent an image, instead of regular EM algorithm. We also use different dissimilarity measures for different queries according to their statistics features. Experiments show this approach can greatly improve the performance of CBIR.