Um modelo bayesiano combinando análise semântica latente e atributos espaciais para recuperação de informação visual

Abstract

This thesis presents an implementation and analysis of three new methodologies for feature extraction and retrieval of visual information of digital images. The first implemented methodology is a new algorithm, called GRAS (Graph Region Arrow Shot), which uses information from the spacial relationship between the images regions. The GRAS information is added to features of classical methods such as color histograms, edge maps and texture information. The second one uses singular value decomposition to extract information from the image context, transforming the original space in a new space, which is calledlatent space. The third methodology uses a bayesian network model which combines the results from the classical methods, achieving better performance than each method alone. For performance evaluation, two image databases are used: one natural and another artificial.In the natural database, which is composed of several manually classified natural images, the results of the new techniques are better. In the experiments, when the GRAS algorithm is used, the improvements reach up to 60% in terms of average precision.