Novos métodos de análise de texturas baseados em modelos gravitacionais simplificados e caminhos mais curtos em grafos

Abstract

Image analysis is an important field of computer vision whose role is to extract significant information from images. Among several relevant attributes, texture is one of the most important because it is a rich source of information. This thesis aims to develop novel texture analysis methods (for grayscale and color images) based on "simplified gravitational systems" and "shortest paths in graphs" which provide feature vectors more discriminative than the methods already established in literature. The first approach converts an image into a simplified gravitational system whose collapse process is explored by using fractal dimension and lacunarity descriptors. The second approach converte the pixels of an image into vertices of a non-oriented weighted graph and explores the shortest paths between pairs of vertices in different scales and orientations. Additionally, this thesis proposes to apply these approaches to plant leaf identification (a relevant problem for botanists), and medical image identification/classification, increasing the confidence of medical diagnosis. The experiments are perfomed on the followiing image databases: Brodatz, UIUC, VisTex, USPTex, Outex, leaf textures, palisade parenchyma, pap-smear and breast tissues. the most significant comparison results are obtained from UIUC, USP-Tex and palisade parenchyma, with success rates of 55,00%, 96,57% and 91,56% (lower success rates) obtained by the proposed methods, respectively. These success rates are almost always superior to the results obtained by the methods used for comparison. This demonstrates that the proposed methods open promising sources of research in grayscale and color texture analysis.