Técnicas biométricas procuraram identificar usuários pela textura da íris, impressão digital, traços faciais, entre outros. A íris humana apresenta características de textura que a classificam como uma peculiaridade biométrica de grande poder de discriminação no reconhecimento de pessoas. O objetivo deste trabalho é avaliar a eficiência de uma nova metodologia de análise de texturas em desenvolvimento no LAVI (Laboratório de Visão Computacional da EESC-USP) na identificação de indivíduos por meio da textura de sua íris. A metodologia denomina-se Local Fuzzy Pattern e tem sido utilizada com excelente desempenho com texturas gerais, naturais e artificiais. Este documento detalha as técnicas utilizadas para extração e normalização da textura da íris, a utilização e os resultados obtidos com o método Local Fuzzy Pattern aplicado à classificação biométrica da íris humana. Os resultados obtidos apresentam sensibilidade de até 99,7516% com a aplicação da metodologia proposta em bancos de imagens de íris humana disponíveis na internet demonstram a viabilidade da técnica proposta. / Biometric techniques sought to identify users by the texture of the iris, fingerprint, facial features, among others. The human iris have texture characteristics that rank it as a powerful biometric peculiarity on human recognition. The objective of this masters proposal is to investigate the efficiency of a new methodology of iris texture analysis currently in development in LAVI (Laboratório de Visão Computacional da EESC-USP). The methodology is called LFP (Local Fuzzy Pattern) and has been used with excellent overall performance on artificial and natural textures. This document details the techniques used for the extraction and normalization of the iris texture, the use and results of the local fuzzy pattern method applied to biometric classification of the human eye. The results show a sensibility of value up to 99.7516%, obtained by applying the proposed methodology on human iris photos from image database available on the internet does showing the viability of the technique.

De nouvelles options thérapeutiques sont en cours d'introduction pour les maladies musculaires génétiques telles que les dystrophies musculaires et les myopathies congénitales, maladies jusque là sans traitement causal. Ces développements récents ont suscité un intérêt renouvelé et croissant pour les méthodes atraumatiques en vue de caractériser et de suivre les muscles atteints, en particulier pendant et après une intervention thérapeutique. Dans ce contexte, les modèles animaux sont essentiels pour mieux comprendre les mécanismes des maladies et pour tester des nouvelles thérapies. Récemment, il y a eu des avancées significatives dans l'évaluation atraumatique de modèles murins de maladies musculaires génétiques. Néanmoins, nombre de lignées de souris n'ont pas encore été caractérisées de façon atraumatique et il reste à mettre au point des méthodes plus sensibles pour identifier précocement des altérations subtiles dans le muscle des souris malades. L'objectif de cette thèse est d'appliquer des techniques atraumatiques innovantes à l'étude du muscle de modèles murins de maladies musculaires génétiques avec des phénotypes variés. Trois lignées de souris modèles de dystrophies musculaires (mdx, Large_myd et mdx/Large_myd) et une lignée de souris modèle de la myopathie congénitale (KI-Dnm2_R465W) ont été étudiées par des méthodes de Résonance Magnétique Nucléaire (RMN). Deux lignées dystrophiques (Large_myd et mdx/Large_myd) plus des souris normales après une blessure ont été étudiées par micro-tomographie (micro-CT). En RMN, toutes les souches de souris affectées ont présenté un T2 musculaire augmenté, en relation avec une gamme d'anomalies histologiques, y comprises nécrose et inflammation, mais aussi des groupes de fibres en régénération ou des fibres avec altérations de l'architecture. Avec la combinaison de la RMN et de l'analyse de la texture, il a été possible d'identifier sans ambiguïté toutes les lignées dystrophiques, alors que la seule mesure du T2 ne permettait pas de les différencier. Les souris mdx ont présenté des altérations fonctionnelles et morphologiques du réseau vasculaire musculaire. Pour les souris KI-Dnm2_R465W, des études préliminaires ont révélé une tendance à développer des altérations fonctionnelles musculaires. Finalement, les images de micro-CT n'ont pas pu détecter des différences du contenu musculaire dans les souris dystrophiques. L'ensemble des résultats non seulement enrichit le panel de modèles murins de maladies génétiques musculaires caractérisés de manière atraumatique, il révèle également un certain degré de spécificité des anomalies dans l'imagerie, comme l'a montré l'analyse de texture. Les résultats démontrent aussi que des méthodes de RMN non-invasives peuvent être assez sensibles pour identifier des altérations subtiles dans le phénotype musculaire murin, même à des stades précoces. Cette thèse a été développée dans le cadre d'une co-tutelle internationale entre la France et le Brésil, et elle a comporté un important transfert de compétence, qui a permis de réaliser les premières explorations atraumatiques du muscle murin effectuées au Brésil. / Novel therapeutic approaches are being introduced for genetic muscle diseases such as muscle dystrophies and congenital myopathies, all of them having remained without cure so far. These recent developments have motivated a renewed and augmented interest in non-invasive methods for muscle characterization and monitoring, particularly during and after therapeutic intervention. In this context, animal models are essential to better understand the disease mechanisms and to test new therapies. Recently, significant advances in the non-invasive evaluation of mouse models for genetic muscle diseases have been achieved. Nevertheless, there were still several mouse strains not characterized non-invasively, and it was necessary to develop sensitive methods to identify subtle alterations in the murine affected muscle. The purpose of this thesis was to apply non-invasive techniques in the study of murine models for genetic muscle diseases with variable phenotypes. Three mouse models for muscle dystrophy (mdx, Large_myd, mdx/Large_myd) and one mouse model for congenital myopathy (KI-Dnm2_R465W) were studied with Nuclear Magnetic Resonance (NMR) methods. Two dystrophic strains (Large_myd, mdx/Large_myd) and normal mice after injury were studied through micro-Computed Tomography (micro-CT). On NMR, all affected mouse strains presented increased muscle T2, which could be related to variable features in the histological evaluation, including necrosis and inflammation, but also to clusters of fibers under regeneration or with altered cytoarchitecture. The combination of NMR and texture analyses allowed the unambiguous differential identification of all the dystrophic strains, although it was not feasible when comparing the muscle T2 measurements only. Mdx mice showed functional and morphological alterations of vascular network. In the KI-Dnm2_R465W mice, a pilot study revealed tendencies of functional impairment. Finally, micro-CT images were unable to detect differences in muscle´s content in dystrophic mice. Altogether, these results not only increased the number of murine models for genetic muscle diseases non-invasively characterized, it also demonstrated some degree of specificity of the imaging anomalies, as revealed by texture analysis. It also showed that non-invasive NMR methods can be sensitive enough to identify subtle alterations in murine muscle phenotype, even in early stages. This thesis was developed under an international joint supervision between France and Brazil, and comprised an important transfer of technology, with the first non-invasive studies of murine muscles performed in Brazil.

Caractériser précisément l’hétérogénéité tumorale constitue un enjeu majeur en cancérologie. Le calcul de biomarqueurs de cette hétérogénéité directement à partir des données d’imagerie présente de nombreux avantages : il est non-invasif, répétable plusieurs fois au cours du traitement, ne nécessite pas d’examen supplémentaire et permet de caractériser la tumeur toute entière et ses éventuelles métastases. Mon projet de recherche visait à développer et évaluer des méthodes pour une caractérisation plus complète de l’activité métabolique des tumeurs. L’analyse de texture des images TEP nécessite un protocole de calcul des index plus complexe que celui des paramètres conventionnels utilisés en clinique. Afin de déterminer l’influence des étapes préliminaires au calcul de ces index, une étude méthodologique a tout d’abord été menée. Cette analyse a montré que certains index de texture étaient redondants et qu’il existait une forte corrélation entre certains d’entre eux et le volume métabolique. Elle a également mis en évidence l’impact de la formule et du taux de discrétisation sur les valeurs des paramètres de texture et permis de clarifier l’interprétation des indices. Après avoir établi un protocole de calcul strict, une seconde partie de ce travail a consisté à évaluer la capacité de ces index pour la caractérisation des tumeurs. L’analyse de texture a ainsi permis de différencier les tissus sains des tissus tumoraux et de distinguer les types histologiques pour les tumeurs mammaires, les lésions pulmonaires ou encore les gliomes.Afin de comprendre le lien entre l’hétérogénéité tumorale quantifiée sur les images TEP et l’hétérogénéité biologique des lésions, nous avons comparé l’analyse de texture réalisée à différentes échelles sur un modèle animal. Cette étude a révélé que la texture mesurée in vivo sur les images TEP reflétait la texture mesurée ex vivo sur les images autoradiographiques. / The precise characterization of the biological heterogeneity of a tumor is a major issue in oncology. The calculation of biomarkers reflecting this heterogeneity directly from imaging data offers a number of advantages: it is non-invasive, can be repeated during the therapy, does not require supplementary examinations and the whole tumor and possible metastases can be investigated from the images. My research project was to develop and assess methods to characterize the metabolic activity distribution in tumors.Texture analysis based on PET images requires a protocol to compute index that is somehow more sophisticate than when simply measuring the conventional index used in clinical practice. To determine the role of the different steps that are involved in the computation of texture index, a methodological study was conducted. This study demonstrated that some texture parameters were redundant and that there existed a strong correlation between some of them and the metabolic volume. We have also shown that the formula and the rate of discretization impact the texture analysis and clarified the interpretation of these metrics. After the protocol of texture index computation has been established, the second part of this work was to assess the interest of these indices for the tumor characterization. We showed that some texture indices were different in tumor and in healthy tissue and could identify histological types such as the triple-negative breast tumors, the squamous cell carcinoma from adenocarcinoma in lung tumors, as well as the grade of gliomas.To understand the links between the tumor heterogeneity as measured from PET images and the biological heterogeneity of lesions, we compared the texture analysis based on different scales in a mouse model. This study revealed that the texture measured in vivo based on PET images reflects the texture measured ex vivo from autoradiographic images.

Osteoarthritis (OA) of the knee is a disease that deteriorates the bones and surrounding soft tissue of the affected joint. Categorisation of the disease into grades of severity is subject to errors of measurement and poor observer agreement. There is an urgent need for automated methods to measure radiographic features and remove, as far as possible, the element of subjectivity in assessment. This project creates a fully automated system to analyse all aspects of the knee in radiographs. The methods evaluate explicit and implicit features of: overall shape, trabecular structure, osteophytes, tibial spines and intercondylar notch, and joint space shape. The project develops the first fully automated osteophyte detection algorithms, improved trabeculae features using raw pixel intensities, and a better analysis of joint space using shape models. This project is the first to combine explicit and implicit features across the whole of the knee, and applies these features to classify radiographs using four main outcomes: current OA, current pain, later onset OA, and later onset pain. The results find a strong current OA classification rate, with an Area Under the ROC Curve (AUC) of 0.904 and weighted kappa of 0.49 (0.48-0.51). The remaining later onset and pain experiments report weaker results; these results suggest that radiographic features in Posterior-Anterior (PA) view radiographs have a weak association with clinical and later onset OA.

Este projeto descreve o desenvolvimento, estudo e aplicação de descritores fractais em análise de texturas. Nos últimos anos, a literatura vem apresentando a geometria fractal como uma ferramenta poderosa para a análise de imagens, com aplicações em variados campos da ciência. A maior parte destes trabalhos faz uso direto da dimensão fractal como um descritor do objeto representado na imagem. Entretanto, em função da complexidade de muitos problemas nesta área, algumas soluções foram propostas para melhorar essa análise, usando não apenas o valor da dimensão fractal, mas um conjunto de medidas que pudessem ser extraídas pela geometria fractal e que descrevessem as texturas com maior riqueza e precisão. Entre essas técnicas, destacam-se a metodologia de multifractais, de dimensão fractal multiescala e, mais recentemente, os descritores fractais. Esta última técnica tem se mostrado eficiente na solução de problemas relacionados à discriminação de imagens de texturas e formas, uma vez que os descritores gerados fornecem uma representação direta do padrão de complexidade (distribuição dos detalhes ao longo das escalas de observação) da imagem. Assim, essa solução permite que se tenha uma descrição rica da imagem estudada pela análise da distribuição espacial e/ou espectral dos pixels e intensidade de cores/tons de cinza, com uma modelagem que pode se aproximar da percepção visual humana para a geração de um método automático e preciso. Ocorre, entretanto, que os trabalhos apresentados até o momento sobre descritores fractais focam em métodos de estimativa de dimensão fractal mais conhecidos como Bouligand-Minkowski e Box-counting. Este projeto visa estudar mais a fundo o conceito, generalizando para outras abordagens de dimensão fractal, bem como explorando diferentes formas de se extraírem os descritores a partir da curva logarítmica associada à dimensão. Os métodos desenvolvidos são aplicados à análise de texturas, em problemas de classificação de bases públicas, cujos resultados podem ser comparados com métodos da literatura, bem como a segmentação de imagens de satélite e à identificação automática de amostras obtidas em estudos de nanotecnologia. Os resultados alcançados demonstram o potencial da metodologia desenvolvida para a solução destes problemas, mostrando tratar-se de uma nova fronteira a ser usada e explorada em análise de imagens e visão computacional como um todo. / This project describes the development, study and application of fractal descriptors to texture analysis. Recently, the literature has shown fractal geometry as a powerful tool for image analysis, with applications to several areas of science. Most of these works use fractal dimension as a descriptor of the object depicted in the image. However, due to the complexity of many problems in this context, some solutions have been proposed to improve this analysis. These proposed methods use not only the value of fractal dimension, but a set of measures which could be extracted by fractal geometry to describe the textures with greater richness and accuracy. Among such techniques, we emphasize the multifractal methodology, multiscale fractal dimension and, more recently, fractal descriptors. This latter technique has demonstrated to be efficient in solving problems related to the discrimination of texture and shape images. This is possible as the extracted descriptors provide a direct representation of the complexity (the details distribution along the scales of observation) in the image. Thus, this solution allows for a rich description of the image studied by analyzing the spatial/spectral distribution of pixels and intensity of colors/gray-levels, with a model which can approximate the human visual perception, generating an automatic and precise method. However, the works about fractal descriptors presented in the literature focus on classical methods to estimate fractal dimension, such as Bouligand-Minkowski and Box-counting. This project aims at studying more deeply the concept, generalizing to other approaches in fractal dimension, as well as exploring different ways of extracting the key features from the logarithmic curve associated with the dimension. The developed methods are applied to texture analysis, in classification problems over public databases, whose results can be compared with literature methods, as well as to the segmentation of satellite images and automatically identifying samples obtained from studies on nanotechnology. The results demonstrate the potential of the methodology developed to solve such problems, showing that this is a new frontier to be explored and used in image analysis and computer vision at all.

As imagens digitais são frequentemente corrompidas por ruídos ou distorcidas pelo processo de aquisição. A teoria dos conjuntos fuzzy e a lógica fuzzy constituem uma alternativa mais adequada para lidar com tais incertezas, em comparação com os sistemas convencionais, baseados na lógica tradicional (crisp). Este trabalho propõe uma nova metodologia para análise de micropadrões de imagens digitais baseada em números fuzzy. Um micropadrão é uma estrutura de níveis de cinza dos pixels de uma vizinhança e pode descrever o contexto espacial da imagem, como borda, textura, linha, canto e padrões mais complexos. Na literatura de visão computacional, algumas abordagens foram desenvolvidas para extrair estas características, tais como Texture Unit (TU), Local Binary Pattern (LBP) e Fuzzy Number Edge Detector (FUNED). O trabalho apresenta um novo método que modela a distribuição dos níveis de cinza de um micropadrão como um conjunto fuzzy, e com base nas funções de pertinência usadas gera códigos-fuzzy que representam o grau de pertinência de cada pixel vizinho com nível de cinza próximo do pixel central. A metodologia proposta é chamada de Local Fuzzy Pattern (LFP) e é aplicada na análise de textura usando a função sigmoide (LFP-s), a função triangular e simétrica (LFP-t) e a função gaussiana (LFP-g) para calcular o grau de pertinência do pixel central em relação à sua vizinhança. Para avaliar o desempenho da técnica proposta foram usados bases de texturas, cujas imagens foram amostradas aleatoriamente. Após processá-las pelas abordagens LFP-s, LFP-t, LFP-g e LBP, foram comparadas as taxas de acertos alcançadas usando a distância Chi-quadrado. Nos experimentos realizados também é avaliado o esforço computacional do LFP, comparando-o com o descritor LBP. Os resultados mostram que o LFP é eficaz na descrição de textura e que supera o LBP nos diferentes testes realizados. Neste trabalho também é demonstrado que a formulação do LFP é uma generalização de técnicas previamente publicadas, como Texture Unit, Local Binary Pattern e FUNED. / Digital images are often corrupted by noise and distorted by the acquisition process. The fuzzy set theory and fuzzy logic are an alternative more appropriate to deal with these uncertainties, in comparison with conventional treatment based on traditional logic (crisp). This work proposes a new methodology for the analysis of micro-patterns of digital images based on fuzzy numbers. A micro-pattern is the structure of the gray-level pixels within a neighborhood and can describe the spatial context of the image, such as edge, texture, line, corner and more complex patterns. In the literature of computer vision, some approaches have been developed to extract these features, such as Texture Unit (TU), Local Binary Pattern (LBP) and Fuzzy Number Edge Detector (FUNED). This work presents a new method that models the distribution of the gray levels of a micro-pattern as a fuzzy set, and based on the membership functions used generates fuzzy-codes that represent the membership degree of each neighbor pixel neighbor with gray-levels near of the central pixel. The proposed methodology is called Local Fuzzy Pattern (LFP) and is applied in the texture analysis by using a sigmoid (LFP-s), a symmetrical triangular (LFP-t) function and Gaussian function (LFP-g) for calculating the membership degree of a central pixel of a neighborhood. To evaluate the performance of the proposed technique were used two database, whose images were randomly sampled. After processing these images by the LFP-s, LFP-t, LFP-g and LBP approaches, it was compared the hit-rate reached by using the Chi-square distance. In the experiments also evaluated the computational effort of the LFP and surpasses the LBP that the different tests. The results show that the LFP-s is efficient to describe texture and that it surpasses the LBP in different tests. This work also demonstrates that the proposed formulation for the LFP is a generalization of previously published techniques such as Texture Unit, LBP and FUNED.

O greening ou Huanglongbing (HLB) é uma das mais graves doenças dos citros presentes nos pomares do Brasil. Causada pela bactéria Candidatus Liberibacter spp, é transmitida pelo inseto psilídeo Diaphorina citri, que ao se alimentar de uma planta doente transmite a doença às demais plantas. O greening apresenta como sintoma, manchas amareladas nas folhas, muitas vezes confundidas com deficiências nutricionais. A erradicação da planta e o controle do inseto transmissor são as únicas formas de prevenção para evitar a sua propagação. Este trabalho teve por objetivo avaliar uma metodologia baseada em segmentação por cor e outra baseada em análise de textura para avaliação de folhas de citros sintomáticas, identificando se estão contaminadas com o greening ou outras doenças e deficiências nutricionais. Foram fornecidas pelo grupo FISHER, 324 amostras de folhas cítricas, contendo folhas com doenças (greening, CVC e rubelose) e deficiências nutricionais (manganês, magnésio e zinco). As folhas foram digitalizadas por um scanner de mesa, com duas resoluções, utilizando somente a parte frontal da folha. Foram montados três bancos de imagens. Os resultados gerados com a metodologia baseada em segmentação por cor utilizando RNA PMC, mostraram que essa metodologia não é eficiente. Na metodologia baseada na análise por textura foram avaliados os descritores LBP, LFP e os de Haralick. Para estes descritores foram extraídas amostras por folha e por quadrantes das folhas nos canais de cores vermelho e verde e amostras em níveis de cinza. Os resultados gerados pelos descritores foram classificados pela distância &#9672 e pelos algoritmos IBK e RNA PMC do toolbox Weka. Os melhores resultados foram para os descritores LBP e LFP-s para distância &#9672, com valores de sensibilidade acima de 97% e 93%, respectivamente, e para o LBP com o algoritmo IBK, com valores de sensibilidade acima de 98,5%. Os resultados obtidos evidenciam que o descritor LBP é o mais eficiente seguido pelo LFP-s na diferenciação do greening das outras pragas. / The greening or Huanglongbing (HLB) is one of the most serious diseases of citrus orchards present in Brazil. HLB is caused by the bacterium Candidatus Liberibacter spp, it is transmitted by the psyllid insect (Diaphorina citri) that, when feeding on a diseased plant, it transmits the disease to other plants. One of the symptoms of the greening are yellowish spots on the leaves, often confused with nutritional deficiencies. The eradication of plants and control of insect are the only forms of prevention. This work aims to evaluate two methodologies: one based on color segmentation and the other based on texture analysis for assessment of symptomatic citrus leaves, identifying whether they are infected with greening and other diseases and nutritional deficiencies. A number of 324 samples of citrus leaves were provided by FISHER group, infected with diseases (greening, CVC, rubelose) and nutritional deficiencies ( manganese, magnesium, zinc) . The leaves were acquired by a flatbed scanner with two different resolutions, using only the front side of the leaf. Three datasets of images were constructed. The results generated using the methodology based on color segmentation with ANN MLP, showed that this methodology is not efficient. In the methodology based on texture analysis it was evaluated the LBP, LFP and the Haralick descriptors. For these descriptors it was extracted samples from the leaves and quadrants of leaves, in red and green color channels and grayscale. The results generated by the descriptors were classified by &#9672 distance and the algorithms IBK and ANN MLP from the toolbox Weka. The best results were for LBP descriptor and LFP-s for &#9672 distance with values of sensitivity above 97% and 93%, respectively, and the LBP with IBK algorithm, with values of sensitivity above 98.5%. The results showed that the LBP descriptor is the most efficient followed by LFP-s in the differentiation of the greening from other pests.

As imagens digitais são frequentemente corrompidas por ruídos ou distorcidas pelo processo de aquisição. A teoria dos conjuntos fuzzy e a lógica fuzzy constituem uma alternativa mais adequada para lidar com tais incertezas, em comparação com os sistemas convencionais, baseados na lógica tradicional (crisp). Este trabalho propõe uma nova metodologia para análise de micropadrões de imagens digitais baseada em números fuzzy. Um micropadrão é uma estrutura de níveis de cinza dos pixels de uma vizinhança e pode descrever o contexto espacial da imagem, como borda, textura, linha, canto e padrões mais complexos. Na literatura de visão computacional, algumas abordagens foram desenvolvidas para extrair estas características, tais como Texture Unit (TU), Local Binary Pattern (LBP) e Fuzzy Number Edge Detector (FUNED). O trabalho apresenta um novo método que modela a distribuição dos níveis de cinza de um micropadrão como um conjunto fuzzy, e com base nas funções de pertinência usadas gera códigos-fuzzy que representam o grau de pertinência de cada pixel vizinho com nível de cinza próximo do pixel central. A metodologia proposta é chamada de Local Fuzzy Pattern (LFP) e é aplicada na análise de textura usando a função sigmoide (LFP-s), a função triangular e simétrica (LFP-t) e a função gaussiana (LFP-g) para calcular o grau de pertinência do pixel central em relação à sua vizinhança. Para avaliar o desempenho da técnica proposta foram usados bases de texturas, cujas imagens foram amostradas aleatoriamente. Após processá-las pelas abordagens LFP-s, LFP-t, LFP-g e LBP, foram comparadas as taxas de acertos alcançadas usando a distância Chi-quadrado. Nos experimentos realizados também é avaliado o esforço computacional do LFP, comparando-o com o descritor LBP. Os resultados mostram que o LFP é eficaz na descrição de textura e que supera o LBP nos diferentes testes realizados. Neste trabalho também é demonstrado que a formulação do LFP é uma generalização de técnicas previamente publicadas, como Texture Unit, Local Binary Pattern e FUNED. / Digital images are often corrupted by noise and distorted by the acquisition process. The fuzzy set theory and fuzzy logic are an alternative more appropriate to deal with these uncertainties, in comparison with conventional treatment based on traditional logic (crisp). This work proposes a new methodology for the analysis of micro-patterns of digital images based on fuzzy numbers. A micro-pattern is the structure of the gray-level pixels within a neighborhood and can describe the spatial context of the image, such as edge, texture, line, corner and more complex patterns. In the literature of computer vision, some approaches have been developed to extract these features, such as Texture Unit (TU), Local Binary Pattern (LBP) and Fuzzy Number Edge Detector (FUNED). This work presents a new method that models the distribution of the gray levels of a micro-pattern as a fuzzy set, and based on the membership functions used generates fuzzy-codes that represent the membership degree of each neighbor pixel neighbor with gray-levels near of the central pixel. The proposed methodology is called Local Fuzzy Pattern (LFP) and is applied in the texture analysis by using a sigmoid (LFP-s), a symmetrical triangular (LFP-t) function and Gaussian function (LFP-g) for calculating the membership degree of a central pixel of a neighborhood. To evaluate the performance of the proposed technique were used two database, whose images were randomly sampled. After processing these images by the LFP-s, LFP-t, LFP-g and LBP approaches, it was compared the hit-rate reached by using the Chi-square distance. In the experiments also evaluated the computational effort of the LFP and surpasses the LBP that the different tests. The results show that the LFP-s is efficient to describe texture and that it surpasses the LBP in different tests. This work also demonstrates that the proposed formulation for the LFP is a generalization of previously published techniques such as Texture Unit, LBP and FUNED.

Texture analysis has been used extensively in the computer-assisted interpretation of SAR sea ice imagery. Provision of maps which distinguish relevant ice types is significant for monitoring global warming and ship navigation. Due to the abundance of SAR imagery available, there exists a need to develop an automated approach for SAR sea ice interpretation. Grey level co-occurrence probability (<i>GLCP</i>) texture features are very popular for SAR sea ice classification. Although these features are used extensively in the literature, they have a tendency to erode and misclassify texture boundaries. Proposed is an advancement to the <i>GLCP</i> method which will preserve texture boundaries during image segmentation. This method exploits the relationship a pixel has with its closest neighbors and weights the texture measurement accordingly. These texture features are referred to as <i>WGLCP</i> (weighted <i>GLCP</i>) texture features. In this research, the <i>WGLCP</i> and <i>GLCP</i> feature sets are compared in terms of boundary preservation, unsupervised segmentation ability, robustness to increasing boundary density and computation time. The <i>WGLCP</i> method outperforms the <i>GLCP</i> method in all aspects except for computation time, where it suffers. From the comparative analysis, an inconsistency with the <i>GLCP</i> correlation statistic was observed, which motivated an investigative study into using this statistic for image segmentation. As the overall goal of the thesis is to improve SAR sea ice segmentation accuracy, the concepts developed from the study are applied to the image segmentation problem. The results indicate that for images with high contrast boundaries, the <i>GLCP</i> correlation statistical feature decreases segmentation accuracy. When comparing <i>WGLCP</i> and <i>GLCP</i> features for segmentation, the <i>WGLCP</i> features provide higher segmentation accuracy.

Texture analysis has been used extensively in the computer-assisted interpretation of SAR sea ice imagery. Provision of maps which distinguish relevant ice types is significant for monitoring global warming and ship navigation. Due to the abundance of SAR imagery available, there exists a need to develop an automated approach for SAR sea ice interpretation. Grey level co-occurrence probability (<i>GLCP</i>) texture features are very popular for SAR sea ice classification. Although these features are used extensively in the literature, they have a tendency to erode and misclassify texture boundaries. Proposed is an advancement to the <i>GLCP</i> method which will preserve texture boundaries during image segmentation. This method exploits the relationship a pixel has with its closest neighbors and weights the texture measurement accordingly. These texture features are referred to as <i>WGLCP</i> (weighted <i>GLCP</i>) texture features. In this research, the <i>WGLCP</i> and <i>GLCP</i> feature sets are compared in terms of boundary preservation, unsupervised segmentation ability, robustness to increasing boundary density and computation time. The <i>WGLCP</i> method outperforms the <i>GLCP</i> method in all aspects except for computation time, where it suffers. From the comparative analysis, an inconsistency with the <i>GLCP</i> correlation statistic was observed, which motivated an investigative study into using this statistic for image segmentation. As the overall goal of the thesis is to improve SAR sea ice segmentation accuracy, the concepts developed from the study are applied to the image segmentation problem. The results indicate that for images with high contrast boundaries, the <i>GLCP</i> correlation statistical feature decreases segmentation accuracy. When comparing <i>WGLCP</i> and <i>GLCP</i> features for segmentation, the <i>WGLCP</i> features provide higher segmentation accuracy.