Entropia Diferencial Multiescala na Descrição e Análise de Formas

Abstract

Shape analysis and recognition are fundamental in the design of systems based on computer vision. The greatest challenge is to develop robust methods capable of extracting significant features from these shapes to represent them. In this context, multiscale descriptors are an alter- native in the characterization of shapes, using a versatile and efficient tool. This work proposes a shape descriptor based on the differential entropy of the multiscale curvature (MEC), which presents better shape discrimination power when compared to the already consolidated descriptor normalized multiscale bending energy (NMBE). In a first approach, these two descriptors are compared with monoscale descriptors in shape classification experiments from three public bases: Kimia-99, MPEG7-CE and Flavia, which is a dataset of plant leaves. The results showed that the monoscale descriptors were more effective in Kimia-99 dataset, while the MEC had the best performance in the last two datasets, which are even the most challenging. In the second approach, the adjustment parameters or scales of the MEC and NMBE descriptors are optimized by a meta-heuristic algorithm called Simulated Annealing. For that, we employ a clustering validation cost function, the Silhouette, aiming to find, through optimization, the best set of scales that maximize the intraclass cohesion and the separation between classes. After optimization, we carried out classification experiments whose results showed a considerable gain in precision, recall and accuracy measures, with special emphasis on the descriptor proposed on the leaves dataset, which exceeded by 10% the gain in precision. Finally, the proposed third approach considers the concatenation of both multiscale descriptors, the latter being the one that reflected the greatest performance gain in the classification experiments.