Rotation-invariant texture classification based on graylevel co-occurrence matrices.

• Ano de defesa: 2013
• Autor(a) principal: Barrera Acuña, Mauricio Andrés
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Classificação de texturas; Lógica nebulosa; Matriz de co-ocorrência; Ortoimagem
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/3/3142/tde-09072014-110549/

Texture is one of the most primitive characteristics of objects. Digital images represent this property as local intensity variations in the image. Consequently, texture is an attribute that is innately present in virtually every digital image; mathematically describing this information leads to a myriad of different applications, from food qualities processing up to medical image analysis. It comes as no surprise that texture has been one of the most researched topics in the computer vision community, and it continues to receive a great deal of attention. One of the most classic approaches to model texture is the statistically-based co-occurrence matrix method. The present dissertation work revolves around a clever variation of the co- occurrence matrix that incorporates rotation-invariance, a very desirable property for texture classification. This variation is taken from previous work in the literature and is used to propose a robust fuzzy orthoimage classifier. Moreover, the original rotation- invariant approach is modified though a generalization and benchmarked with one of the most widely-used texture description methods in the recent literature: the Local Binary Patterns approach. The results indicate that Fuzzy Logic is a powerful tool to build texture-based classifiers that have to deal with diversely-sourced image samples; the results also indicate that the generalization proposal, which originates the here named CCM and RCM texture description methods, offers a significant performance boost that is comparable, and even better than Local Binary Patterns approach when comparing accuracy scores.