Avanços recentes em caracterização e classificação de imagens de texturas: explorando teoria da informação, aprendizado profundo e de variedades

• Ano de defesa: 2020
• Autor(a) principal: Nsimba, Cédrick Bamba
• Orientador(a): Levada, Alexandre Luis Magalhães
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Classificação de textura; Campos aleatórios gaussianos markovianos; Teoria da informação; Informação de Fisher; Transformada Wavelet; Aprendizado profundo; Aprendizado de variedades
• Palavras-chave em Inglês: Texture classification; Gaussian markov random field; Information theory; Fisher information; Wavelet Transform; Deep learning; Manifold learning
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/16199

The task of extracting features from images is a very important activity for many computer vision and image processing applications. Especially the characterization and identification of textures is a fundamental issue in this area which allows the promotion of the development of a wide range of interesting applications including medical image analysis, content image retrieval, object recognition, among others. Due to this importance, many kinds of research have been carried out in this area and, consequently, resulted in the development of a set of texture descriptors. However, in general, algorithms considered as the state of the art still present performance problems when applied in the presence of noise, since it does not maintain intrinsic properties of the image being analyzed. Assuming that random fields and wavelets are appropriate mathematical tools to assist in this process, opportunities open for the development of very interesting applications, with the potential to promote a high level of texture classification accuracy. While random fields model better the statistical properties of texture images, wavelets provide a robust tool for decomposition and multiresolution analysis. In addition, information theory-based measurements, such as Fisher information and Shannon entropy of Gaussian Markov random field models obtained from the sub-bands of a wavelet decomposition tree are able to measure texture patterns. If the dimensionality of the calculated texture descriptors is high, nonlinear dimensionality reduction methods based on manifold learning techniques can be used. Our perception is that manifold learning techniques are capable of selecting discriminative characteristics from a high dimensionality space. The aim of this research work was to propose new approaches based on the use of information theory, deep and manifold learning to characterize and classify texture images. To this end, three new approaches have been proposed: (1) descriptor design based on Fisher information matrices of Gaussian Markov random field models obtained from the sub-bands of a wavelet decomposition tree; (2) descriptor design based on the application of manifold learning methods acting as a selector of relevant texture information; and (3) descriptor design based on the combination of deep and manifold learning. Experimental results demonstrated that the proposed methods are competitive with related state of the art descriptors.