Bases de imagens armazenadas em sistemas computacionais da área médica correspondem a uma valiosa fonte de conhecimento. Assim, a mineração de imagens pode ser aplicada para extrair conhecimento destas bases com o propósito de apoiar o diagnóstico auxiliado por computador (Computer Aided Diagnosis - CAD). Sistemas CAD apoiados por mineração de imagens tipicamente realizam a extração de características visuais relevantes das imagens. Essas características são organizadas na forma de vetores de características que representam as imagens e são utilizados como entrada para classificadores. Devido ao problema conhecido como lacuna semântica, que corresponde à diferença entre a percepção da imagem pelo especialista médico e suas características automaticamente extraídas, um aspecto desafiador do CAD é a obtenção de um conjunto de características que seja capaz de representar de maneira sucinta e eficiente o conteúdo visual de imagens médicas. Foi desenvolvido neste trabalho o extrator de características FFS (Fast Fractal Stack) que realiza a extração de características de forma, que é um atributo visual que aproxima a semântica esperada pelo ser humano. Adicionalmente, foi desenvolvido o algoritmo de classificação Concept, que emprega mineração de regras de associação para predizer a classe de uma imagem. O aspecto inovador do Concept refere-se ao algoritmo de obtenção de representações de imagens, denominado MFS-Map (Multi Feature Space Map) e também desenvolvido neste trabalho. O MFS-Map realiza agrupamento de dados em diferentes espaços de características para melhor aproveitar as características extraídas no processo de classificação. Os experimentos realizados para imagens de tomografia pulmonar e mamografias indicam que tanto o FFS como a abordagem de representação adotada pelo Concept podem contribuir para o aprimoramento de sistemas CAD / Medical image databases represent a valuable source of data from which potential knowledge can be extracted. Image mining can be applied to knowledge discover from these data in order to help CAD (Computer Aided Diagnosis) systems. The typical set-up of a CAD system consists in the extraction of relevant visual features in the form of image feature vectors that are used as input to a classifier. Due to the semantic gap problem, which corresponds to the difference between the humans image perception and the features automatically extracted from the image, a challenging aspect of CAD is to obtain a set of features that is able to succinctly and efficiently represent the visual contents of medical images. To deal with this problem it was developed in this work a new feature extraction method entitled Fast Fractal Stack (FFS). FFS extracts shape features from objects and structures, which is a visual attribute that approximates the semantics expected by humans. Additionally, it was developed the Concept classification method, which employs association rules mining to the task of image class prediction. The innovative aspect of Concept refers to its image representation algorithm termed MFS-Map (Multi Feature Space Map). MFS-Map employs clustering in different feature spaces to maximize features usefulness in the classification process. Experiments performed employing computed tomography and mammography images indicate that both FFS and Concept methods for image representation can contribute to the improvement of CAD systems

As plantações de eucaliptos representam grande potencial econômico para a indústria de papel, celulose, entre outras, além de apresentar uma série de características positivas como alta produtividade, grande potencial de adaptação e ampla diversidade de espécies. Em consequência a tais vantagens, há décadas diversas pesquisas vem sendo realizadas com o intuito de monitorar e detectar diversas doenças que aferem este tipo de cultura. O monitoramento rápido das doenças em eucaliptos torna-se um requisito para evitar grandes perdas econômicas. Neste projeto de pesquisa utilizou-se imagens aéreas obtidas por VANTs (Veículos Aéreos Não-Tripulados) para detectar um tipo específico de estresse que afeta as plantações de eucaliptos: a Murcha de Ceratocyst is. Após rotular eucaliptos doentes e saudáveis e outras estruturas em imagens aéreas, técnicas de Aprendizado de Máquina Supervisionado foram desenvolvidas para generalizar o conhecimento e possibilitar uma rápida detecção através das imagens RGB e multiespectrais. Dentre as técnicas utilizadas, destacou-se a arquitetura de Redes Neurais Convolucional chamada de Custom- CNN, inspirada no modelo da tradicional arquitetura Lenet -5 agregando-se melhorias do estado-da-arte, como a camada convolucional 1x1. Na classificação do conjunto RGB, a Custom-CNN obteve o maior F-score, de 0,81, sendo que a técnica SVM-rbf obteve 0,67. No conjunto de dados com imagens multiespectrais, a Lenet -5 e a Custom-CNN at ingiram, respectivamente, 0,63 e 0,66 de F-score, enquanto o SVM-rbf obteve 0,46. Esta dissertação apresenta a metodologia utilizada para a classificação, elencando as principais características dos algoritmos utilizados, bem como os resultados experimentais obtidos. Há ainda uma aplicação do classificador Regressão Logística para o planejamento de trajetória com VANTs. / Eucalypt us plantations represent great economic potential for t he paper, pulp, among others, in addition to presenting a number of positive characteristics such as high productivity, great potential for adaptaion and wide diversity of species. In consequence of t hese advantages, there are several decades research has been conducted in order to monitor and detect various diseases that affect s this type of culture. The rapid monitoring of diseases in eucalyptus becomes a requirement to avoid major economic losses. In t his research project we used aerial images obtained by UAVs (Unmanned Aerial Vehicles) to detect an specific type of stress t hat a effect s eucalyptus plantations: the Ceratocyst is wilt . After labeling diseased eucalyptus, healthy eucalyptus and other structures in aerial images, Supervised Machine Learning techniques were developed to generalize knowledge and enable rapid detection through RGB and multispectral images. Among the techniques used, stood out t he Convolutional Neural Network architecture called Custom-CNN, that was inspired by the model of t raditional Lenet -5 architecture and with state-of-the-art improvements, such as t he 1x1 convolution layer. In t he classification of RGB dataset , the Custom-CNN obtained the highest F-score of 0.81, and SVM-RBF technique obtained 0.67. In t he dataset with multispectral images, Lenet -5 and Custom-CNN obtained, respectively, 0.63 and 0.66 of F-score, while SVM-rbf obtained 0.46. This paper presents the methodology used for classification, listing the main features of the algorithms and the experimental results. There is also an application of Logistic Regression classifier for path planning with UAVs.

Traffic sign recognition is an important problem for autonomous cars and driver assistance systems. With recent developments in the field of machine learning, high performance can be achieved, but typically at a large computational cost. This thesis aims to investigate the relation between classification accuracy and computational complexity for the visual recognition problem of classifying traffic signs. In particular, the benefits of partitioning the classification problem into smaller sub-problems using prior knowledge in the form of shape or current region are investigated. In the experiments, the convolutional neural network (CNN) architecture MobileNetV2 is used, as it is specifically designed to be computationally efficient. To incorporate prior knowledge, separate CNNs are used for the different subsets generated when partitioning the dataset based on region or shape. The separate CNNs are trained from scratch or initialized by pre-training on the full dataset. The results support the intuitive idea that performance initially increases with network size and indicate a network size where the improvement stops. Including shape information using the two investigated methods does not result in a significant improvement. Including region information using pretrained separate classifiers results in a small improvement for small complexities, for one of the regions in the experiments. In the end, none of the investigated methods of including prior knowledge are considered to yield an improvement large enough to justify the added implementational complexity. However, some other methods are suggested, which would be interesting to study in future work.

Made available in DSpace on 2015-03-05T13:53:43Z (GMT). No. of bitstreams: 0 Previous issue date: 24 / Nenhuma / A recuperação e classificação de imagens é um tema bastante pesquisado atualmente. Além dos desafios encontrados no campo teórico e prático para permitir que máquinas possuam a capacidade de visão, sua pesquisa resulta em várias aplicações práticas para o dia-a-dia. A visão computação, grande área na qual está inserida a recuperação e classificação de imagens, possui aplicações e práticas dentre as quais podemos citar softwares capazes de recuperarem imagens em bases de dados de imagens, reconhecimento de pessoas por características de biometria(impressões digitais, reconhecimento por íris ou face), localização e quantificação de logomarcas na mídia, localização de objetos numa cena e mecanismos de visão para a robótica. A pesquisa desenvolvida nesta dissertação foca-se em obter uma generalização através do aprendizado das características de uma coleção de imagens pertencentes a uma mesma classe as quais servirão como exemplo de aprendizagem, com isto obtendo um modelo que identifique esta classe. Para tan / Image retrieval and classification are today the subject of extensive research. This topic poses both theoretical and practical challenges as researchers attempt to give machines such as computers and robots the ability to “see”. Image retrieval and classification are part of a wider field known as Computer Vision, which encompasses several practical applications such as image retrieval from databases storing only raw images, biometric recognition (from images of finger-prints, face or iris), retrieval of visual trademarks and logos from advertisements, location of objects in a scene and vision techniques in robotics. The research developed in this work is focused on obtaining a generalization of characteristics extracted from a collection of images belonging to a single class using supervised learning techniques. The result is a model that “identifies” a given class of images. To achieve this, a review of the state-of-the-art in content-based image retrieval systems and Machine Learning techniques was neede

Cette thèse s'intéresse au problème de la reconnaissance d'objets dans les images vidéo et plus particulièrement à celui de leur localisation. Elle a été conduite dans le contexte d'une collaboration scientifique entre l'INRIA Rhône-Alpes et MBDA France. De ce fait, une attention particulière a été accordée à l’applicabilité des approches proposées aux images infra-rouges. La méthode de localisation proposée repose sur l'utilisation d'une fenêtre glissante incluant une cascade à deux étages qui, malgré sa simplicité, permet d'allier rapidité et précision. Le premier étage est un étage de filtrage rejetant la plupart des faux positifs au moyen d’un classifieur SVM linéaire. Le deuxième étage élimine les fausses détections laissées par le premier étage avec un classifieur SVM non-linéaire plus lent, mais plus performant. Les fenêtres sont représentées par des descripteurs HOG et Bag-of-words. La seconde contribution de la thèse réside dans une méthode permettant de combiner localisation d'objets et catégorisation d'images. Ceci permet, d'une part, de prendre en compte le contexte de l'image lors de la localisation des objets, et d'autre part de s'appuyer sur la structure géométrique des objets lors de la catégorisation des images. Cette méthode permet d'améliorer les performances pour les deux tâches et produit des détecteurs et classifieurs dont la performance dépasse celle de l'état de l'art. Finalement, nous nous penchons sur le problème de localisation de catégories d'objets similaires et proposons de décomposer la tâche de localisation d'objets en deux étapes. Une première étape de détection permet de trouver les objets sans déterminer leurs positions tandis qu’une seconde étape d’identification permet de prédire la catégorie de l'objet. Nous montrons que cela permet de limiter les confusions entre les classes, principal problème observé pour les catégories d'objets visuellement similaires. La thèse laisse une place importante à la validation expérimentale, conduites sur la base PASCAL VOC ainsi que sur des bases d’images spécifiquement réalisées pour la thèse. / This thesis addresses the problem of object recognition in images and more precisely the problem of object localization. It have been conducted in the context of a scientific collaboration between INRIA Rhônes-Alpes and MBDA France. Therefore, a particular attention was accorded to the applicability of the proposed approaches on infrared images. The localization method proposed here relies on the sliding windows mechanism combined with a two stage cascade that, despite its simplicity, allies rapidity and precision. The first stage is a filtering stage that rejects most of the false positives using a linear classifier. The second stage prunes the detections of the first classifier using a slower yet efficient non-linear classifier. Windows are represented with HOG and Bag-of-words descriptors. The second contribution of this thesis is a method that combines object localization and image categorization. This allows, on the one hand, to take into account context information in localization, and on the other hand, to rely on geometrical structure of objects while performing image categorization. This combination leads to a significant quality improvement and obtains performance superior to the state of the art for both tasks. Finally, we consider the problem of localizing visually similar object categories and suggest to decompose the task of object localization into two steps. The first is a detection step that allows to find objects without determining their category while the second step, an identification step, predicts the objects categories. We show that this approach limits inter-class confusion, which is the main difficulty faced when localizing visually similar object classes. This thesis accords an important place to experimental validation conducted on PASCAL VOC databases as well as other databases specifically introduced for the thesis.

Ce travail de thèse a visé le développement d’un outil d’évaluation des flux métalliques annuels émis par les matériaux de couvertures à l’échelle d’un bassin versant, dans le contexte architectural et météorologique de l’Île-de-France. La méthodologie mise en place pour tendre vers ce but repose sur (1) l’évaluation des émissions annuelles de métaux par différents matériaux métalliques de couverture classiquement utilisés dans la région grâce à une approche expérimentale sur bancs d’essais, (2) l’établissement d’un cadre méthodologique pour la modélisation des flux métalliques émis à l’échelle annuelle par les toitures d’un bassin versant qui se base d’une part sur la modélisation des émissions métalliques par les matériaux à différentes échelles spatiales et temporelles (en fonction de la pluviométrie, de la géométrie du toit…) à partir des données obtenues sur les bancs d’essais, et d’autre part sur la quantification des surfaces métalliques des toitures d’un bassin versant. La première partie du travail a donc consisté à développer et à exploiter, sur deux sites différents, des bancs d’essais expérimentaux d’1/2 m², testant 12 matériaux métalliques issus de 5 familles (zinc, cuivre, plomb acier, aluminium), sous différentes mises en oeuvre (panneaux, gouttières, crochets de fixation). 13 espèces métalliques ont été quantifiées dans les eaux de ruissellement collectées ce qui a permis (1) d’acquérir une importante base de données de taux de ruissellement annuels par les différents matériaux, mettant en évidence que les taux de ruissellement annuels obtenus peuvent être assez importants, de l’ordre de plusieurs grammes par m² et par an pour les éléments constitutifs des matériaux, (2) de hiérarchiser ces matériaux en fonction de leur potentiel polluant, à travers la définition d’un indice de contamination métallique se basant sur les émissions de Cd, Cu, Ni, Pb et Zn et permettant de tenir compte des différences de toxicité des métaux. Une modélisation des émissions métalliques par les matériaux à différentes échelles de temps a été réalisée, conduisant à la conclusion que la hauteur de pluie, ainsi que la durée d’exposition sont des paramètres fondamentaux. Il est apparu que la hauteur de pluie seule est suffisante pour modéliser les émissions métalliques par les matériaux à des échelles de temps longues mais ne suffit pas à modéliser ces émissions sur quelques semaines. Un modèle plus complexe, basé sur une hypothèse d’accumulation / dissolution de produits de corrosion à la surface des matériaux donne des résultats satisfaisant à ces échelles de temps plus courtes. L’extrapolation spatiale des résultats de ruissellement obtenus sur les bancs d’essais s’est basé sur d’autres expérimentations, d’abord sur bancs d’essais conduisant à la conclusion que la longueur d’écoulement n’a pas d’influence sur la masse de métal entraînée dans le ruissellement, qui peut être calculée à partir de la hauteur de pluie, de la surface projetée et de l’inclinaison du panneau (qui s’avère négligeable quand elle est inférieure à 50°); puis à l’échelle de toits réels pour une étape de validation. Dans la seconde partie de ce travail, la quantification des surfaces de rampants à l’échelle d’un bassin versant a été effectuée grâce à un outil de classification d’image basé sur l’analyse de la radiométrie des matériaux à partir d’une photo aérienne. Les résultats obtenus sont encourageants, avec environ 75 à 80% des toits qui bien classés à l’issue de la classification. Les principales erreurs reposent sur des confusions de l’outil entre des matériaux de radiométries voisines (ardoise / zinc par exemple, qui peuvent être proches en fonction du degré d’ensoleillement)... / This thesis aimed to develop a tool for the evaluation of annual metallic flows emitted from roofing materials at the scale of a watershed in the architectural and meteorological context of Paris conurbation. The methodology used in this work is based on (1) the assessment of annual metallic emissions from different metallic materials typically used for roofing in the region considered through an experimental test bed, (2) the establishment of a methodological framework for modelling the metallic flow emitted from the roofs of catchment area, which is based both on the modelling of metal emissions from the materials at different spatial and temporal scales (depending on rainfall, geometry of the roof ...) from data obtained on the test bed, and on the quantification of metallic surface areas of roofs in the catchment area. The first part of the work has been based on the exploitation of experimental test beds of 1 / 2 m², testing 12 metallic materials from 5 families (zinc, copper, lead, steel, aluminium) in various implemented (panels, gutters, fixing brackets, exposed on two different sites. 13 metallic species were quantified in the collected runoff which allowed (1) to acquire a large database of annual runoff rates by different materials, highlighting that the annual runoff rates obtained can be fairly important, with an order of magnitude of several grams per square meter per year for the constitutive elements of materials, (2)to classify these materials according to their polluting potential, through the definition of an index of metal contamination taking into consideration the emissions of Cd, Cu, Ni, Pb and Zn and the differences in toxicity of metals. A modelling of metal emissions from the materials at different time scales has been conducted, leading to the conclusion that the rainfall quantity and the duration of exposure are fundamental parameters. It appeared that the rainfall value is sufficient to model metallic emission from materials for long time scales but not enough to model these emissions on a few weeks period. A more complex model, based on an assumption of accumulation / dissolution of corrosion products on the surface of the material gives satisfactory results for these time-scales periods. The spatial extrapolation of results obtained on the test bed scale was based on other experiments, first on test beds, leading to the conclusion that the length of flow has no influence on the mass of metal entrained in the runoff, which can be calculated from the rainfall quantity, the projected area and inclination of the panel (which is negligible when it is below 50 °), and then at the real roof scale for a validation step. In the second part of this work, quantification of surface areas of roofs at the scale of the catchment was conducted using a classification tool image analysis based on the radiometry of materials. The results are encouraging, with about 75 to 80% of roofs ranked on the basis of classification. The main errors are due to confusions between materials presenting nearby radiometry (slate / zinc, for example, which can be close depending on the amount of sunshine). Exploratory work was conducted for the consideration of singular elements - usually realized in metal -, from the use of unified technical documents. The evaluation of metal surfaces concerned has proved difficult to implement in an automatic way because of the small size of these elements, not visible on an aerial photo

A good image representation is vital for image comparision and classification; it may affect the classification accuracy and efficiency. The purpose of this thesis was to explore novel and appropriate image representations. Another aim was to investigate these representations for image classification. Finally, novel features were examined for improving image classification accuracy. Images of interest to this thesis were textile design images. The motivation of analysing textile design images is to help designers browse images, fuel their creativity, and improve their design efficiency. In recent years, bag-of-words model has been shown to be a good base for image representation, and there have been many attempts to go beyond this representation. Bag-of-words models have been used frequently in the classification of image data, due to good performance and simplicity. “Words” in images can have different definitions and are obtained through steps of feature detection, feature description, and codeword calculation. The model represents an image as an orderless collection of local features. However, discarding the spatial relationships of local features limits the power of this model. This thesis exploited novel image representations, bag of shapes and region label graphs models, which were based on bag-of-words model. In both models, an image was represented by a collection of segmented regions, and each region was described by shape descriptors. In the latter model, graphs were constructed to capture the spatial information between groups of segmented regions and graph features were calculated based on some graph theory. Novel elements include use of MRFs to extract printed designs and woven patterns from textile images, utilisation of the extractions to form bag of shapes models, and construction of region label graphs to capture the spatial information. The extraction of textile designs was formulated as a pixel labelling problem. Algorithms for MRF optimisation and re-estimation were described and evaluated. A method for quantitative evaluation was presented and used to compare the performance of MRFs optimised using alpha-expansion and iterated conditional modes (ICM), both with and without parameter re-estimation. The results were used in the formation of the bag of shapes and region label graphs models. Bag of shapes model was a collection of MRFs' segmented regions, and the shape of each region was described with generic Fourier descriptors. Each image was represented as a bag of shapes. A simple yet competitive classification scheme based on nearest neighbour class-based matching was used. Classification performance was compared to that obtained when using bags of SIFT features. To capture the spatial information, region label graphs were constructed to obtain graph features. Regions with the same label were treated as a group and each group was associated uniquely with a vertex in an undirected, weighted graph. Each region group was represented as a bag of shape descriptors. Edges in the graph denoted either the extent to which the groups' regions were spatially adjacent or the dissimilarity of their respective bags of shapes. Series of unweighted graphs were obtained by removing edges in order of weight. Finally, an image was represented using its shape descriptors along with features derived from the chromatic numbers or domination numbers of the unweighted graphs and their complements. Linear SVM classifiers were used for classification. Experiments were implemented on data from Liberty Art Fabrics, which consisted of more than 10,000 complicated images mainly of printed textile designs and woven patterns. Experimental data was classified into seven classes manually by assigning each image a text descriptor based on content or design type. The seven classes were floral, paisley, stripe, leaf, geometric, spot, and check. The result showed that reasonable and interesting regions were obtained from MRF segmentation in which alpha-expansion with parameter re-estimation performs better than alpha-expansion without parameter re-estimation or ICM. This result was not only promising for textile CAD (Computer-Aided Design) to redesign the textile image, but also for image representation. It was also found that bag of shapes model based on MRF segmentation can obtain comparable classification accuracy with bag of SIFT features in the framework of nearest neighbour class-based matching. Finally, the result indicated that incorporation of graph features extracted by constructing region label graphs can improve the classification accuracy compared to both bag of shapes model and bag of SIFT models.

The Deepwater Horizon oil spill that started on April 20, 2010, in the Gulf of Mexico was the largest marine oil spill in the history of the petroleum industry. There was an unexpected and prolonged sedimentation event of oil-associated marine snow to the seafloor due to the oil spill. The sedimentation event occurred because of the coagulation process among oil associated marine particles. Marine scientists are developing models for the coagulation process of marine particles and oil, in order to estimate the amount of oil that may reach the seafloor along with marine particles. These models, used certain assumptions regarding the shape and the texture parameters of marine particles. Such assumptions may not be based on accurate information or may vary during and after the oil spill. The work performed here provided a quantitative analysis of the assumptions used in modeling the coagulation process of marine particles. It also investigated the changes in model parameters (shape and texture) during and after the Deepwater Horizon oil spill in different seasons (spring and summer). An Interactive Visualization Application was developed for data exploration and visual analysis of the trends in these parameters. An Interactive Statistical Analysis Application was developed to create a statistical summary of these parameter values.

Résumé en français non disponible

This thesis develops and evaluates three approaches for quantitative molecularly-targeted optical imaging of neoplasia. The first approach focuses on widefield imaging of biomarkers near the tissue surface for early detection applications; this approach is demonstrated in freshly resected oral tissue. Most oral cancers are not detected until the disease has spread, but topical application of targeted imaging agents allows rapid visualization of biomarker expression, giving real-time, objective information. Epidermal growth factor receptor (EGFR) expression was quantified in patient samples using fluorescent epidermal growth factor. Dysplasia (n=4) and cancer (n=13) had an average 2.3-fold and 3.8-fold increase in signal compared to normal tissue. EGFR expression was assessed along with metabolic activity using a fluorescent glucose analog, 2-NBDG, in 9 patient samples. A classification algorithm using quantitative image features resulted in an area under the curve (AUC) of 0.83, though the main advantage of this technique may be to understand spatial heterogeneity of biomarker expression and how this correlates with disease. The next approach focuses on high-resolution optical imaging through a needle to detect metastases in lymphoid tissue for clinical staging applications; this approach is demonstrated in resected lymph nodes from breast cancer patients. These patients often require removal of nodes, but an optical imaging strategy using topical application of imaging agents in vivo may classify nodes as normal or metastatic, thus reducing unnecessary removal of normal nodes and improving metastasis detection. Proflavine, a nuclear dye, was topically applied to 43 nodes. A classification algorithm developed from quantitative image features distinguished normal lymphoid tissue from metastases with an AUC of 0.84. Because optical imaging is depth limited, the final approach combines high-resolution optical imaging with magnetic resonance imaging (MRI) for multimodal evaluation of deep tissue. An imaging agent functional in both optical and MRI was developed by co-loading fluorescent EGFR antibodies and gadolinium-based contrast agents in silicon discs. These discs accumulate in tumors, resulting in localized delivery of imaging agents. The research presented here can be applied to understanding tumor biology and biomarker heterogeneity, with the future clinical goal of improving identification of disease and determination of appropriate therapy for cancer patients.