Relevance feedback approach for image retrieval combining support vector machines and adapted Gaussian mixture models

A new relevance feedback (RF) approach for content-based image retrieval (CBIR) is presented, which uses Gaussian mixture (GM) models as image representations. The GM of each image is obtained as an adaptation of a universal GM which models the probability distribution of the features of the image database. In each RF round, the positive and negative examples provided by the user until the current round are used to train a support vector machine (SVM) to distinguish between the relevant and irrelevant images according to the preferences of the user. In order to quantify the similarity between two images represented as GMs, Kullback-Leibler (KL) approximations are employed, the computation of which can be further accelerated taking advantage from the fact that the GMs of the images are all refined from a common model. An appropriate kernel function, based on this distance between GMs, is used to make possible the incorporation of GMs in the SVM framework. Finally, comparative numerical experiments that demonstrate the merits of the proposed RF methodology and the advantages of using GMs for image modelling are provided.