Descritores de imagens baseados em assinatura textual

Abstract

The technique of representing images by a set of visual words, known as bag of visual words, has been successfully applied to content-based image retrieval. In this work, it is proposed the paradigm Signature-based Bag of Visual Words (S-BoVW), a formal defini-tion for methods that describe images by visual words, without the previous construction of a visual vocabulary. Methods based on this paradigm not require the use of cluste-ring algorithms, which allows to reduce the cost associated with the images description step. The coding and combination of features such as color and texture were explored in this work in order to define new descriptors of images based on the S-BoVW paradigm. Experiments were carried out in order to propose effective and efficient ways to apply the concept proposed by S-BoVW paradigm in the definition of new methods of content-based image retrieval. The results obtained demonstrate that proper choice of query pro-cessing technique and the similarity function ensures obtaining optimized performance by S-BoVW methods and also ensures their competitiveness compared to baselines in many scenarios.