Selection and combination of unsupervised learning methods for image retrieval

Abstract

The evolution of technologies to store and share images has made imperative the need for methods to index and retrieve multimedia information based on visual content. The CBIR (Content-Based Image Retrieval) systems are the main solution in this scenario. Originally, these systems were solely based on the use of low-level visual features, but evolved through the years in order to incorporate various supervised learning techniques. More recently, unsupervised learning methods have been showing promising results for improving the effectiveness of retrieval results. However, given the development of different methods, a challenging task consists in to exploit the advantages of diverse approaches. As different methods present distinct results even for the same dataset and set of features, a promising approach is to combine these methods. In this work, a framework is proposed aiming at selecting the best combination of methods in a given scenario, using different strategies based on effectiveness and correlation measures. Regarding the experimental evaluation, six distinct unsupervised learning methods and two different datasets were used. The results as a whole are promising and also reveal good perspectives for future works.