I-Generalized and Kullback - Leibler Divergences for Content-Based Image Retrieval

Technological advancement has increased the amount of information is generated every day. The social and economic relevance of image recovery systems has created the need for improvement. Computing similarity between two images is a fundamental step in a CBIR system to retrieve images of interest. A proper similarity measure choice is the essence for efficient and effective image retrieval. Therefore, recent studies have used Bregman divergences in researches because of its flexibility in similarity analysis. Thus, this work aims to propose an efficient and robust method using the divergences of Bregman I-Generalized and Kullback Leibler to be used as measures of similarity in CBIR systems. The model bag of the visual word approach based on image subregions was used to characterize the images. Experiments used public databases: Holiday and Caltech101. A result shows our proposal achieved satisfactory gains when compared to Euclidean and Cosine distances in different classifiers. Moreover, our approach shows promising and competitive results when compared to methods presented in the literature. Therefore, using divergences with appropriate treatment can be used as similarity functions in parametric classifiers to minimize semantic gap em CBIR.