Tesi sul tema "Unsupervised image segmentation"

Image segmentation forms the first stage in many image analysis procedures including image sequence re-timing and the emerging field of content based retrieval. By dividing the image into a set of disjoint connected regions, each of which is homogeneous with respect to some measure of the image content, the scene can be analysed and metadata extracted more efficiently, and in many cases more effectively, than on a pixel by pixel basis. Though a great number of segmentation techniques exist (and continue to be developed,) many of them fall short of the requirements of these applications. This thesis first defines these requirements and reviews established segmentation methods describing their qualities and shortfalls. Selecting the watershed transform and connected operators from those techniques reviewed a number of novel adaptations are introduced, developed and shown to produce pleasing results both in terms of a new evaluation metric and subjective appraisal. Finally, the use of the image segmentation is shown to improve established methods of image noise removal using the discrete wavelet transform.

The development of a fully unsupervised algorithm to achieve image segmentation is the central theme of this dissertation. Existing literature falls short of such a goal providing many algorithms capable of solving a subset of this highly challenging problem. Unsupervised segmentation is the process of identifying and locating the constituent regions of an observed image, while having no prior knowledge of the number of regions. The problem can be formulated in a Bayesian framework and through the use of an assumed model unsupervised segmentation can be posed as a problem of optimisation. This is the approach pursued throughout this dissertation. Throughout the literature, the commonly adopted model is an hierarchical image model whose underlying components are various forms of Markov Random Fields. Gaussian Markov Random Field models are used to model the textural content of the observed image's regions, while a Potts model provides a regularisation function for the segmentation. The optimisation of such highly complicated models is a topic that has challenged researchers for several decades. The contribution of this thesis is the introduction of new techniques allowing unsupervised segmentation to be carried using a single optimisation process. It is hoped that these algorithms will facilitate the future study of hierarchical image models and in particular the discovery of further models capable of more closely fitting real world data. The extensive literature surrounding Markov Random Field models and their optimisation is reviewed early in this dissertation, as is the literature concerning the selection of features to identify the textural content of an observed image. In the light of these reviews new algorithms are proposed that achieve a fusion between concepts originating in both these areas. Algorithms previously applied in statistical mechanics form an important part of this work. The use of various Markov Chain Monte Carlo algorithms is prevalent and in particular, the reversible jump sampling algorithm is of great significance. It is the combination of several of these algorithms to form a single optimisation framework that lies at the heart of the most successful algorithms presented here.

We propose a novel approach to investigate and implement unsupervised segmentation of color images particularly natural color images. The aim is to devise a robust unsu- pervised segmentation approach that can segment a color textured image accurately. Here, the color and texture information of each individual pixel along with the pixel's spatial relationship within its neighborhood have been considered for producing precise segmentation of color images. Precise segmentation of images has tremendous potential in various application domains like bioinformatics, forensics, security and surveillance, the mining and material industry and medical imaging where subtle information related to color and texture is required to analyze an image accurately. We intend to implement a robust unsupervised segmentation approach for color im- ages using a newly developed multidimensional spatially variant ¯nite mixture model (MSVFMM) using a Markov Random Fields (MRF) model for improving the over- all accuracy in segmentation and Haar wavelet transform for increasing the texture sensitivity of the proposed approach. [...]
Master of Computing

Biomedical image analysis plays a crucial role in the development of healthcare, with a wide scope of applications including the disease diagnosis, clinical treatment, and future prognosis. Among various biomedical image analysis techniques, segmentation is an essential step, which aims at assigning each pixel with labels of interest on the category and instance. At the early stage, the segmentation results were obtained via manual annotation, which is time-consuming and error-prone. Over the past few decades, hand-craft feature based methods have been proposed to segment the biomedical images automatically. However, these methods heavily rely on prior knowledge, which limits their generalization ability on various biomedical images. With the recent advance of the deep learning technique, convolutional neural network (CNN) based methods have achieved state-of-the-art performance on various nature and biomedical image segmentation tasks. The great success of the CNN based segmentation methods results from the ability to learn contextual and local information from the high dimensional feature space. However, the biomedical image segmentation tasks are particularly challenging, due to the complicated background components, the high variability of object appearances, numerous overlapping objects, and ambiguous object boundaries. To this end, it is necessary to establish automated deep learning-based segmentation paradigms, which are capable of processing the complicated semantic and morphological relationships in various biomedical images. In this thesis, we propose novel deep learning-based methods for fully supervised and unsupervised biomedical image segmentation tasks. For the first part of the thesis, we introduce fully supervised deep learning-based segmentation methods on various biomedical image analysis scenarios. First, we design a panoptic structure paradigm for nuclei instance segmentation in the histopathology images, and cell instance segmentation in the fluorescence microscopy images. Traditional proposal-based and proposal-free instance segmentation methods are only capable to leverage either global contextual or local instance information. However, our panoptic paradigm integrates both of them and therefore achieves better performance. Second, we propose a multi-level feature fusion architecture for semantic neuron membrane segmentation in the electron microscopy (EM) images. Third, we propose a 3D anisotropic paradigm for brain tumor segmentation in magnetic resonance images, which enlarges the model receptive field while maintaining the memory efficiency. Although our fully supervised methods achieve competitive performance on several biomedical image segmentation tasks, they heavily rely on the annotations of the training images. However, labeling pixel-level segmentation ground truth for biomedical images is expensive and labor-intensive. Subsequently, exploring unsupervised segmentation methods without accessing annotations is an important topic for biomedical image analysis. In the second part of the thesis, we focus on the unsupervised biomedical image segmentation methods. First, we proposed a panoptic feature alignment paradigm for unsupervised nuclei instance segmentation in the histopathology images, and mitochondria instance segmentation in EM images. To the best of our knowledge, we are for the first time to design an unsupervised deep learning-based method for various biomedical image instance segmentation tasks. Second, we design a feature disentanglement architecture for unsupervised object recognition. In addition to the unsupervised instance segmentation for the biomedical images, our method also achieves state-of-the-art performance on the unsupervised object detection for natural images, which further demonstrates its effectiveness and high generalization ability.

Advances in medical imaging have greatly aided in providing accurate and fast medical diagnosis, followed by recent deep learning developments enabling the efficient and cost-effective analysis of medical images. Among different image processing tasks, medical segmentation is one of the most crucial aspects because it provides the class, location, size, and shape of the subject of interest, which is invaluable and essential for diagnostics. Nevertheless, acquiring annotations for training data usually requires expensive manpower and specialised expertise, making supervised training difficult. To overcome these problems, unsupervised domain adaptation (UDA) has been adopted to bridge knowledge between different domains. Despite the appearance dissimilarities of different modalities such as MRI and CT, researchers have concluded that structural features of the same anatomy are universal across modalities, which unfolded the study of multi-modality image segmentation with UDA methods. The traditional UDA research tackled the domain shift problem by minimising the distance of the source and target distributions in latent spaces with the help of advanced mathematics. However, with the recent development of the generative adversarial network (GAN), the adversarial UDA methods have shown outstanding performance by producing synthetic images to mitigate the domain gap in training a segmentation network for the target domain. Most existing studies focus on modifying the network architecture, but few investigate the generative adversarial training strategy. Inspired by the recent success of state-of-the-art data augmentation techniques in classification tasks, we designed a novel mix-up strategy to assist GAN training for the better synthesis of structural details, consequently leading to better segmentation results. In this thesis, we propose SynthMix, an add-on module with a natural yet effective training policy that can promote synthetic quality without altering the network architecture. SynthMix is a mix-up synthesis scheme designed for integration with the adversarial logic of GAN networks. Traditional GAN approaches judge an image as a whole which could be easily dominated by discriminative features, resulting in little improvement of delicate structures in synthesis. In contrast, SynthMix uses the data augmentation technique to reinforce detail transformation at local regions. Specifically, it coherently mixes up aligned images of real and synthetic samples at local regions to stimulate the generation of fine-grained features examined by an associated inspector for domain-specific details. We evaluated our method on two segmentation benchmarks among three publicly available datasets. Our method showed a significant performance gain compared with existing state-of-the-art approaches.

We propose a novel approach to investigate and implement unsupervised segmentation of color images particularly natural color images. The aim is to devise a robust unsu- pervised segmentation approach that can segment a color textured image accurately. Here, the color and texture information of each individual pixel along with the pixel's spatial relationship within its neighborhood have been considered for producing precise segmentation of color images. Precise segmentation of images has tremendous potential in various application domains like bioinformatics, forensics, security and surveillance, the mining and material industry and medical imaging where subtle information related to color and texture is required to analyze an image accurately. We intend to implement a robust unsupervised segmentation approach for color im- ages using a newly developed multidimensional spatially variant ¯nite mixture model (MSVFMM) using a Markov Random Fields (MRF) model for improving the over- all accuracy in segmentation and Haar wavelet transform for increasing the texture sensitivity of the proposed approach. [...]
Master of Computing

We propose a novel approach to investigate and implement unsupervised segmentation of color images particularly natural color images. The aim is to devise a robust unsu- pervised segmentation approach that can segment a color textured image accurately. Here, the color and texture information of each individual pixel along with the pixel's spatial relationship within its neighborhood have been considered for producing precise segmentation of color images. Precise segmentation of images has tremendous potential in various application domains like bioinformatics, forensics, security and surveillance, the mining and material industry and medical imaging where subtle information related to color and texture is required to analyze an image accurately. We intend to implement a robust unsupervised segmentation approach for color im- ages using a newly developed multidimensional spatially variant ¯nite mixture model (MSVFMM) using a Markov Random Fields (MRF) model for improving the over- all accuracy in segmentation and Haar wavelet transform for increasing the texture sensitivity of the proposed approach. [...]
Master of Computing

L'accès aux séquences d'images 3D s'est aujourd'hui démocratisé, grâce aux récentes avancées dans le développement des capteurs de profondeur ainsi que des méthodes permettant de manipuler des informations 3D à partir d'images 2D. De ce fait, il y a une attente importante de la part de la communauté scientifique de la vision par ordinateur dans l'intégration de l'information 3D. En effet, des travaux de recherche ont montré que les performances de certaines applications pouvaient être améliorées en intégrant l'information 3D. Cependant, il reste des problèmes à résoudre pour l'analyse et la segmentation de scènes intérieures comme (a) comment l'information 3D peut-elle être exploitée au mieux ? et (b) quelle est la meilleure manière de prendre en compte de manière conjointe les informations couleur et 3D ? Nous abordons ces deux questions dans cette thèse et nous proposons de nouvelles méthodes non supervisées pour la classification d'images 3D et la segmentation prenant en compte de manière conjointe les informations de couleur et de profondeur. A cet effet, nous formulons l'hypothèse que les normales aux surfaces dans les images 3D sont des éléments à prendre en compte pour leur analyse, et leurs distributions sont modélisables à l'aide de lois de mélange. Nous utilisons la méthode dite « Bregman Soft Clustering » afin d'être efficace d'un point de vue calculatoire. De plus, nous étudions plusieurs lois de probabilités permettant de modéliser les distributions de directions : la loi de von Mises-Fisher et la loi de Watson. Les méthodes de classification « basées modèles » proposées sont ensuite validées en utilisant des données de synthèse puis nous montrons leur intérêt pour l'analyse des images 3D (ou de profondeur). Une nouvelle méthode de segmentation d'images couleur et profondeur, appelées aussi images RGB-D, exploitant conjointement la couleur, la position 3D, et la normale locale est alors développée par extension des précédentes méthodes et en introduisant une méthode statistique de fusion de régions « planes » à l'aide d'un graphe. Les résultats montrent que la méthode proposée donne des résultats au moins comparables aux méthodes de l'état de l'art tout en demandant moins de temps de calcul. De plus, elle ouvre des perspectives nouvelles pour la fusion non supervisée des informations de couleur et de géométrie. Nous sommes convaincus que les méthodes proposées dans cette thèse pourront être utilisées pour la classification d'autres types de données comme la parole, les données d'expression en génétique, etc. Elles devraient aussi permettre la réalisation de tâches complexes comme l'analyse conjointe de données contenant des images et de la parole
Access to the 3D images at a reasonable frame rate is widespread now, thanks to the recent advances in low cost depth sensors as well as the efficient methods to compute 3D from 2D images. As a consequence, it is highly demanding to enhance the capability of existing computer vision applications by incorporating 3D information. Indeed, it has been demonstrated in numerous researches that the accuracy of different tasks increases by including 3D information as an additional feature. However, for the task of indoor scene analysis and segmentation, it remains several important issues, such as: (a) how the 3D information itself can be exploited? and (b) what is the best way to fuse color and 3D in an unsupervised manner? In this thesis, we address these issues and propose novel unsupervised methods for 3D image clustering and joint color and depth image segmentation. To this aim, we consider image normals as the prominent feature from 3D image and cluster them with methods based on finite statistical mixture models. We consider Bregman Soft Clustering method to ensure computationally efficient clustering. Moreover, we exploit several probability distributions from directional statistics, such as the von Mises-Fisher distribution and the Watson distribution. By combining these, we propose novel Model Based Clustering methods. We empirically validate these methods using synthetic data and then demonstrate their application for 3D/depth image analysis. Afterward, we extend these methods to segment synchronized 3D and color image, also called RGB-D image. To this aim, first we propose a statistical image generation model for RGB-D image. Then, we propose novel RGB-D segmentation method using a joint color-spatial-axial clustering and a statistical planar region merging method. Results show that, the proposed method is comparable with the state of the art methods and requires less computation time. Moreover, it opens interesting perspectives to fuse color and geometry in an unsupervised manner. We believe that the methods proposed in this thesis are equally applicable and extendable for clustering different types of data, such as speech, gene expressions, etc. Moreover, they can be used for complex tasks, such as joint image-speech data analysis

Cette thèse présente plusieurs algorithmes développés dans le cadre du projet ANR COCLICO et contient deux axes principaux :Le premier axe concerne l'introduction d'un algorithme applicable aux images satellite à très haute résolution, qui est basé sur les champs aléatoires de Markov et qui apporte des notions sémantiques sur les clusters découverts. Cet algorithme est inspiré de l'algorithme Iterated conditional modes (ICM) et permet de faire un clustering sur des segments d'images pré-traitées. La méthode que nous proposons permet de gérer des voisinages irréguliers entre segments et d'obtenir des informations sémantiques de bas niveau sur les clusters de l'image traitée.Le second axe porte sur le développement de méthodes de clustering collaboratif applicables à autant d'algorithmes que possible, ce qui inclut les algorithmes du premier axe. La caractéristique principale des méthodes proposées dans cette thèse est leur applicabilité aux deux cas suivants : 1) plusieurs algorithmes travaillant sur les mêmes objets dans des espaces de représentation différents, 2) plusieurs algorithmes travaillant sur des données différentes ayant des distributions similaires. Les méthodes que nous proposons peuvent s'appliquer à de nombreux algorithmes comme l'ICM, les K-Moyennes, l'algorithme EM, ou les cartes topographiques (SOM et GTM). Contrairement aux méthodes précédemment proposées, notre modèle permet à des algorithmes très différents de collaborer ensemble, n'impose pas de contrainte sur le nombre de clusters recherchés et a une base mathématique solide
This thesis presents several algorithms developed in the context of the ANR COCLICO project and contains two main axis: The first axis is concerned with introducing Markov Random Fields (MRF) based models to provide a semantic rich and suited algorithm applicable to images that are already segmented. This method is based on the Iterated Conditional Modes Algorithm (ICM algorithm) and can be applied to the segments of very high resolution (VHR) satellite pictures. Our proposed method can cope with highly irregular neighborhood dependencies and provides some low level semantic information on the clusters and their relationship within the image. The second axis deals with collaborative clustering methods developed with the goal of being applicable to as many clustering algorithms as possible, including the algorithms used in the first axis of this work. A key feature of the methods proposed in this thesis is that they can deal with either of the following two cases: 1) several clustering algorithms working together on the same data represented in different feature spaces, 2) several clustering algorithms looking for similar clusters in different data sets having similar distributions. Clustering algorithms to which these methods are applicable include the ICM algorithm, the K-Means algorithm, density based algorithms such as DB-scan, all Expectation-Maximization (EM) based algorithms such as the Self-Organizing Maps (SOM) and the Generative Topographic Mapping (GTM) algorithms. Unlike previously introduced methods, our models have no restrictions in term of types of algorithms that can collaborate together, do not require that all methods be looking for the same number of clusters, and are provided with solid mathematical foundations

La fusion d’informations a été largement étudiée dans le domaine de l’intelligence artificielle. Une information est en général considérée comme imparfaite. Par conséquent, la combinaison de plusieurs sources d’informations (éventuellement hétérogènes) peut conduire à une information plus globale et complète. Dans le domaine de la fusion on distingue généralement les approches probabilistes et non probabilistes dont fait partie la théorie de l’évidence, développée dans les années 70. Cette méthode permet de représenter à la fois, l’incertitude et l’imprécision de l’information, par l’attribution de fonctions de masses qui s’appliquent non pas à une seule hypothèse (ce qui est le cas le plus courant pour les méthodes probabilistes) mais à un ensemble d’hypothèses. Les travaux présentés dans cette thèse concernent la fusion d’informations pour la segmentation d’images.Pour développer cette méthode nous sommes partis de l’algorithme de la « Ligne de Partage des Eaux » (LPE) qui est un des plus utilisés en détection de contours. Intuitivement le principe de la LPE est de considérer l’image comme un relief topographique où la hauteur d’un point correspond à son niveau de gris. On suppose alors que ce relief se remplit d’eau par des sources placées au niveau des minima locaux de l’image, formant ainsi des bassins versants. Les LPE sont alors les barrages construits pour empêcher les eaux provenant de différents bassins de se mélanger. Un problème de cette méthode de détection de contours est que la LPE directement appliquée sur l’image engendre une sur-segmentation, car chaque minimum local engendre une région. Meyer et Beucher ont proposé de résoudre cette question en spécifiant un ensemble de marqueurs qui seront les seules sources d’inondation du relief. L'extraction automatique des marqueurs à partir des images ne conduit pas toujours à un résultat satisfaisant, en particulier dans le cas d'images complexes. Plusieurs méthodes ont été proposées pour déterminer automatiquement ces marqueurs.Nous nous sommes en particulier intéressés à l’approche stochastique d’Angulo et Jeulin qui estiment une fonction de densité de probabilité (fdp) d'un contour (LPE) après M simulations de la segmentation LPE classique. N marqueurs sont choisis aléatoirement pour chaque réalisation. Par conséquent, une valeur de fdp élevée est attribuée aux points de contours correspondant aux fortes réalisations. Mais la décision d’appartenance d’un point à la « classe contour » reste dépendante d’une valeur de seuil. Un résultat unique ne peut donc être obtenu.Pour augmenter la robustesse de cette méthode et l’unicité de sa réponse, nous proposons de combiner des informations grâce à la théorie de l’évidence.La LPE se calcule généralement à partir de l’image gradient, dérivée du premier ordre, qui donne une information globale sur les contours dans l’image. Alors que la matrice Hessienne, matrice des dérivées d’ordre secondaire, donne une information plus locale sur les contours. Notre objectif est donc de combiner ces deux informations de nature complémentaire en utilisant la théorie de l’évidence. Les différentes versions de la fusion sont testées sur des images réelles de la base de données Berkeley. Les résultats sont comparés avec cinq segmentations manuelles fournies, en tant que vérités terrain, avec cette base de données. La qualité des segmentations obtenues par nos méthodes sont fondées sur différentes mesures: l’uniformité, la précision, l’exactitude, la spécificité, la sensibilité ainsi que la distance métrique de Hausdorff
Information fusion has been widely studied in the field of artificial intelligence. Information is generally considered imperfect. Therefore, the combination of several sources of information (possibly heterogeneous) can lead to a more comprehensive and complete information. In the field of fusion are generally distinguished probabilistic approaches and non-probabilistic ones which include the theory of evidence, developed in the 70s. This method represents both the uncertainty and imprecision of the information, by assigning masses not only to a hypothesis (which is the most common case for probabilistic methods) but to a set of hypothesis. The work presented in this thesis concerns the fusion of information for image segmentation.To develop this method we start with the algorithm of Watershed which is one of the most used methods for edge detection. Intuitively the principle of the Watershed is to consider the image as a landscape relief where heights of the different points are associated with grey levels. Assuming that the local minima are pierced with holes and the landscape is immersed in a lake, the water filled up from these minima generate the catchment basins, whereas watershed lines are the dams built to prevent mixing waters coming from different basins.The watershed is practically applied to the gradient magnitude, and a region is associated with each minimum. Therefore the fluctuations in the gradient image and the great number of local minima generate a large set of small regions yielding an over segmented result which can hardly be useful. Meyer and Beucher proposed seeded watershed or marked-controlled watershed to surmount this oversegmentation problem. The essential idea of the method is to specify a set of markers (or seeds) to be considered as the only minima to be flooded by water. The number of detected objects is therefore equal to the number of seeds and the result is then markers dependent. The automatic extraction of markers from the images does not lead to a satisfying result especially in the case of complex images. Several methods have been proposed for automatically determining these markers.We are particularly interested in the stochastic approach of Angulo and Jeulin who calculate a probability density function (pdf) of contours after M simulations of segmentation using conventional watershed with N markers randomly selected for each simulation. Therefore, a high pdf value is assigned to strong contour points that are more detected through the process. But the decision that a point belong to the "contour class" remains dependent on a threshold value. A single result cannot be obtained.To increase the robustness of this method and the uniqueness of its response, we propose to combine information with the theory of evidence.The watershed is generally calculated on the gradient image, first order derivative, which gives comprehensive information on the contours in the image.While the Hessian matrix, matrix of second order derivatives, gives more local information on the contours. Our goal is to combine these two complementary information using the theory of evidence. The method is tested on real images from the Berkeley database. The results are compared with five manual segmentation provided as ground truth, with this database. The quality of the segmentation obtained by our methods is tested with different measures: uniformity, precision, recall, specificity, sensitivity and the Hausdorff metric distance

L'iris est considérée comme une des modalités les plus robustes et les plus performantes en biométrie à cause de ses faibles taux d'erreurs. Ces performances ont été observées dans des situations contrôlées, qui imposent des contraintes lors de l'acquisition pour l'obtention d'images de bonne qualité. Relâcher ces contraintes, au moins partiellement, implique des dégradations de la qualité des images acquises et par conséquent une réduction des performances de ces systèmes. Une des principales solutions proposées dans la littérature pour remédier à ces limites est d'améliorer l'étape de segmentation de l'iris. L'objectif principal de ce travail de thèse a été de proposer des méthodes originales pour la segmentation des images dégradées de l'iris. Les chaînes de Markov ont été déjà proposées dans la littérature pour résoudre des problèmes de segmentation d'images. Dans ce cadre, une étude de faisabilité d'une segmentation non supervisée des images dégradées d'iris en régions par les chaînes de Markov a été réalisée, en vue d'une future application en temps réel. Différentes transformations de l'image et différentes méthodes de segmentation grossière pour l'initialisation des paramètres ont été étudiées et comparées. Les modélisations optimales ont été introduites dans un système de reconnaissance de l'iris (avec des images en niveaux de gris) afin de produire une comparaison avec les méthodes existantes. Finalement une extension de la modélisation basée sur les chaînes de Markov cachées, pour une segmentation non supervisée des images d'iris acquises en visible, a été mise en place
Iris is considered as one of the most robust and efficient modalities in biometrics because of its low error rates. These performances were observed in controlled situations, which impose constraints during the acquisition in order to have good quality images. The renouncement of these constraints, at least partially, implies degradations in the quality of the acquired images and it is therefore a degradation of these systems’ performances. One of the main proposed solutions in the literature to take into account these limits is to propose a robust approach for iris segmentation. The main objective of this thesis is to propose original methods for the segmentation of degraded images of the iris. Markov chains have been well solicited to solve image segmentation problems. In this context, a feasibility study of unsupervised segmentation into regions of degraded iris images by Markov chains was performed. Different image transformations and different segmentation methods for parameters initialization have been studied and compared. Optimal modeling has been inserted in iris recognition system (with grayscale images) to produce a comparison with the existing methods. Finally, an extension of the modeling based on the hidden Markov chains has been developed in order to realize an unsupervised segmentation of the iris images acquired in visible light

Les chaînes de Markov cachées (HMC) sont très utilisées pour la segmentation bayésienne non supervisée de données discrètes. Elles sont particulièrement robustes et, malgré leur simplicité, elles sont suffisamment efficaces dans de nombreuses situations. En particulier pour la segmentation d'image, malgré leur nature unidimensionnelle, elles sont capables, grâce à une transformation des images bidimensionnelles en séquences monodimensionnelles avec le balayage de Peano (PS), de produire des résultats satisfaisants. Cependant, dans certains cas, on peut préférer des modèles plus complexes tels que les champs de Markov cachées (HMF) malgré leur plus grande complexité en temps, pour leurs meilleurs résultats. De plus, les modèles de Markov cachés (les chaînes aussi bien que les champs) ont été étendus aux modèles de Markov couples et triplets, qui peuvent être intéressant dans des cas plus complexes. Par exemple, lorsque le temps de séjour n'est pas géométrique, les chaînes de semi-Markov cachées (HSMC) ont tendance à être plus performantes que les HMC, and on peut dire de même pour les chaînes de Markov évidentielles cachées (HEMC) dans le cas de données non-stationnaires. Dans cette thèse, nous proposons dans un premier lieu une nouvelle chaîne de Markov triplet (TMC), qui étend simultanément les HSMC et les HEMC. Basée sur les chaînes de Markov triplets cachées (HTMC), la nouvelle chaîne de semi-Markov évidentielle cachée (HESMC) peut être utilisée de manière non supervisée, les paramètres étant estimés avec l'algorithme Expectation-Maximization (EM). Nous validons l'intérêt d'un tel modèle grâce à des expériences sur des données synthétiques. Nous nous intéressons ensuite au problème de l'unidimensionnalité des HMC avec PS dans le cadre de la segmentation d'image, en construisant le balayage de Peano contextuel (CPS). Il consiste à associer à chaque indexe dans le HMC obtenu à partir du PS, deux observations sur les pixels qui sont voisins du pixel en question dans l'image considérée, mais qui ne sont pas voisins dans la HMC. On obtient donc trois observations pour chaque point du balayage de Peano, ce qui induit une nouvelle chaîne de Markov conditionnelle (CMC) avec une structure plus complexe, mais dont la loi a posteriori est toujours markovienne. Ainsi, nous pouvons appliquer la méthode classique d'estimation des paramètres : l'algorithme Stochastic Expectation-Maximization (SEM), ainsi qu'étudier la segmentation non supervisée obtenue avec l'estimateur du mode des marginales a posteriori (MPM). Les segmentations supervisées et non supervisées par MPM, basées sur la CMC avec CPS, sont comparés aux HMC avec PS et aux HMF à travers des expériences sur des images synthétiques. Elles améliorent de manière significative les premières, et peuvent même être compétitives avec ces derniers. Finalement, nous étendons les CMC-CPS aux chaînes de Markov couples conditionnelles (CPMC) et à deux chaînes de Markov triplets particulières : les chaînes de Markov évidentielles conditionnelles (CEMC) et les chaînes de semi-Markov conditionnelles (CSMC). Pour chacune de ces extensions, nous montrons qu'elles peuvent améliorer de manière notable leur contrepartie non conditionnelle, ainsi que les CMC-CPS, et peuvent même être compétitives avec les HMF. Par ailleurs, elles permettent de mieux utiliser la généralité du triplet markovien dans le cadre de la segmentation d'image, en contournant les problèmes de temps de calcul considérables qui apparaissent lorsque l'on passe des champs de Markov cachés aux triplets
Hidden Markov chains (HMC) are widely used in unsupervised Bayesian hidden discrete data restoration. They are very robust and, in spite of their simplicity, they are sufficiently efficient in many situations. In particular for image segmentation, despite their mono-dimensional nature, they are able, through a transformation of the bi-dimensional images into mono-dimensional sequences with Peano scan (PS), to give satisfying results. However, sometimes, more complex models such as hidden Markov fields (HMF) may be preferred in spite of their increased time complexity, for their better results. Moreover, hidden Markov models (the chains as well as the fields) have been extended to pairwise and triplet Markov models, which can be of interest in more complex situations. For example, when sojourn time in hidden states is not geometrical, hidden semi-Markov (HSMC) chains tend to perform better than HMC, and such is also the case for hidden evidential Markov chains (HEMC) when data are non-stationary. In this thesis, we first propose a new triplet Markov chain (TMC), which simultaneously extends HSMC and HEMC. Based on hidden triplet Markov chains (HTMC), the new hidden evidential semi-Markov chain (HESMC) model can be used in unsupervised framework, parameters being estimated with Expectation-Maximization (EM) algorithm. We validate its interest through some experiments on synthetic data. Then we address the problem of mono-dimensionality of the HMC with PS model in image segmentation by introducing the “contextual” Peano scan (CPS). It consists in associating to each index in the HMC obtained from PS, two observations on pixels which are neighbors of the pixel considered in the image, but are not its neighbors in the HMC. This gives three observations on each point of the Peano scan, which leads to a new conditional Markov chain (CMC) with a more complex structure, but whose posterior law is still Markovian. Therefore, we can apply the usual parameter estimation method: Stochastic Expectation-Maximization (SEM), as well as study unsupervised segmentation Marginal Posterior Mode (MPM) so obtained. The CMC with CPS based supervised and unsupervised MPM are compared to the classic scan based HMC-PS and the HMF through experiments on artificial images. They improve notably the former, and can even compete with the latter. Finally, we extend the CMC-CPS to Pairwise Conditional Markov (CPMC) chains and two particular triplet conditional Markov chain: evidential conditional Markov chains (CEMC) and conditional semi-Markov chains (CSMC). For each of these extensions, we show through experiments on artificial images that these models can improve notably their non conditional counterpart, as well as the CMC with CPS, and can even compete with the HMF. Beside they allow the generality of markovian triplets to better play its part in image segmentation, while avoiding the substantial time complexity of triplet Markov fields

In this research work, we develop an obstacle detection and emergency exit sign recognition system on a mobile phone by extending the feature from accelerated segment test detector with Harris corner filter. The first step often required for many vision based applications is the detection of objects of interest in an image. Hence, in this research work, we introduce emergency exit sign detection method using colour histogram. The hue and saturation component of an HSV colour model are processed into features to build a 2D colour histogram. We backproject a 2D colour histogram to detect emergency exit sign from a captured image as the first task required before performing emergency exit sign recognition. The result of classification shows that the 2D histogram is fast and can discriminate between objects and background with accuracy. One of the challenges confronting object recognition methods is the type of image feature to compute. In this work therefore, we present two feature detectors and descriptor methods based on the feature from accelerated segment test detector with Harris corner filter. The first method is called Upright FAST-Harris and binary detector (U-FaHB), while the second method Scale Interpolated FAST-Harris and Binary (SIFaHB). In both methods, feature points are extracted using the accelerated segment test detectors and Harris filter to return the strongest corner points as features. However, in the case of SIFaHB, the extraction of feature points is done across the image plane and along the scale-space. The modular design of these detectors allows for the integration of descriptors of any kind. Therefore, we combine these detectors with binary test descriptor like BRIEF to compute feature regions. These detectors and the combined descriptor are evaluated using different images observed under various geometric and photometric transformations and the performance is compared with other detectors and descriptors. The results obtained show that our proposed feature detector and descriptor method is fast and performs better compared with other methods like SIFT, SURF, ORB, BRISK, CenSurE. Based on the potential of U-FaHB detector and descriptor, we extended it for use in optical flow computation, which we termed the Nearest-flow method. This method has the potential of computing flow vectors for use in obstacle detection. Just like any other new methods, we evaluated the Nearest flow method using real and synthetic image sequences. We compare the performance of the Nearest-flow with other methods like the Lucas and Kanade, Farneback and SIFT-flow. The results obtained show that our Nearest-flow method is faster to compute and performs better on real scene images compared with the other methods. In the final part of this research, we demonstrate the application potential of our proposed methods by developing an obstacle detection and exit sign recognition system on a camera phone and the result obtained shows that the methods have the potential to solve this vision based object detection and recognition problem.

L'iris est considérée comme une des modalités les plus robustes et les plus performantes en biométrie à cause de ses faibles taux d'erreurs. Ces performances ont été observées dans des situations contrôlées, qui imposent des contraintes lors de l'acquisition pour l'obtention d'images de bonne qualité. Relâcher ces contraintes, au moins partiellement, implique des dégradations de la qualité des images acquises et par conséquent une réduction des performances de ces systèmes. Une des principales solutions proposées dans la littérature pour remédier à ces limites est d'améliorer l'étape de segmentation de l'iris. L'objectif principal de ce travail de thèse a été de proposer des méthodes originales pour la segmentation des images dégradées de l'iris. Les chaînes de Markov ont été déjà proposées dans la littérature pour résoudre des problèmes de segmentation d'images. Dans ce cadre, une étude de faisabilité d'une segmentation non supervisée des images dégradées d'iris en régions par les chaînes de Markov a été réalisée, en vue d'une future application en temps réel. Différentes transformations de l'image et différentes méthodes de segmentation grossière pour l'initialisation des paramètres ont été étudiées et comparées. Les modélisations optimales ont été introduites dans un système de reconnaissance de l'iris (avec des images en niveaux de gris) afin de produire une comparaison avec les méthodes existantes. Finalement une extension de la modélisation basée sur les chaînes de Markov cachées, pour une segmentation non supervisée des images d'iris acquises en visible, a été mise en place
Iris is considered as one of the most robust and efficient modalities in biometrics because of its low error rates. These performances were observed in controlled situations, which impose constraints during the acquisition in order to have good quality images. The renouncement of these constraints, at least partially, implies degradations in the quality of the acquired images and it is therefore a degradation of these systems’ performances. One of the main proposed solutions in the literature to take into account these limits is to propose a robust approach for iris segmentation. The main objective of this thesis is to propose original methods for the segmentation of degraded images of the iris. Markov chains have been well solicited to solve image segmentation problems. In this context, a feasibility study of unsupervised segmentation into regions of degraded iris images by Markov chains was performed. Different image transformations and different segmentation methods for parameters initialization have been studied and compared. Optimal modeling has been inserted in iris recognition system (with grayscale images) to produce a comparison with the existing methods. Finally, an extension of the modeling based on the hidden Markov chains has been developed in order to realize an unsupervised segmentation of the iris images acquired in visible light

The title of the thesis focuses the attention on hyperspectral image segmentation, that is, we want to detect salient regions in a hyperspectral image and isolate them as accurate as possible. This purpose presents two main problems: Firstly, the fact of using hyperspectral imaging not only give us a huge amount of information, but we also have to face the problem of selecting somehow the information avoiding redundancies.
Secondly, the problem of segmentation strictly speaking is still a challenging question whatever the input image would be.
This thesis is focused on solving the whole process by means of building an image processing method that analyses and optimises the information acquired by a multispectral device. After that, it detects the main regions that are present in the scene in an image segmentation procedure. Therefore, this work will be divided into two parts. In the first part, an approach for selecting the most relevant subset of input bands will be presented. In the second part, this reduced representation of the initial bands will be the input data of a segmentation method.
Finally, the main contributions of this PhD work could be briefly summarised as follows. On the one hand, we have proposed a pre-processing stage with an unsupervised band selection approach based on information measures that reduces considerably the amount of data. This approach has been successfully compared with well-known algorithms of the literature, showing its good performance with regard to pixel image classification tasks. On the other hand, after the band selection stage, two unsupervised segmentation procedures for detecting the main parts in multispectral images have been also developed. Regarding to this segmentation part, we have mainly contributed with two measures of similarity among regions. An objective functional for selecting an optimal (or close to optimal) partition of the image is another relevant contribution too.

Ce travail est dédié à la résolution de plusieurs problèmes de grand intérêt en traitement d’images : segmentation, choix de modèle et estimation de paramètres, pour le cas spécifique d’images texturées indirectement observées (convoluées et bruitées). Dans ce contexte, les contributions de cette thèse portent sur trois plans différents : modéle, méthode et algorithmique.Du point de vue modélisation de la texture, un nouveaumodèle non-gaussien est proposé. Ce modèle est défini dans le domaine de Fourier et consiste en un mélange de Gaussiennes avec une Densité Spectrale de Puissance paramétrique.Du point de vueméthodologique, la contribution est triple –troisméthodes Bayésiennes pour résoudre de manière :–optimale–non-supervisée–des problèmes inverses en imagerie dans le contexte d’images texturées ndirectement observées, problèmes pas abordés dans la littérature jusqu’à présent.Plus spécifiquement,1. la première méthode réalise la déconvolution myope non-supervisée et l’estimation des paramètres de la texture,2. la deuxième méthode est dédiée à la déconvolution non-supervisée, le choix de modèle et l’estimation des paramètres de la texture et, finalement,3. la troisième méthode déconvolue et segmente une image composée de plusieurs régions texturées, en estimant au même temps les hyperparamètres (niveau du signal et niveau du bruit) et les paramètres de chaque texture.La contribution sur le plan algorithmique est représentée par une nouvelle version rapide de l’algorithme Metropolis-Hastings. Cet algorithme est basé sur une loi de proposition directionnelle contenant le terme de la ”direction de Newton”. Ce terme permet une exploration rapide et efficace de l’espace des paramètres et, de ce fait, accélère la convergence
This thesis is addressing a series of inverse problems of major importance in the fieldof image processing (image segmentation, model choice, parameter estimation, deconvolution)in the context of textured images. In all of the aforementioned problems theobservations are indirect, i.e., the textured images are affected by a blur and by noise. Thecontributions of this work belong to three main classes: modeling, methodological andalgorithmic. From the modeling standpoint, the contribution consists in the development of a newnon-Gaussian model for textures. The Fourier coefficients of the textured images are modeledby a Scale Mixture of Gaussians Random Field. The Power Spectral Density of thetexture has a parametric form, driven by a set of parameters that encode the texture characteristics.The methodological contribution is threefold and consists in solving three image processingproblems that have not been tackled so far in the context of indirect observationsof textured images. All the proposed methods are Bayesian and are based on the exploitingthe information encoded in the a posteriori law. The first method that is proposed is devotedto the myopic deconvolution of a textured image and the estimation of its parameters.The second method achieves joint model selection and model parameters estimation froman indirect observation of a textured image. Finally, the third method addresses the problemof joint deconvolution and segmentation of an image composed of several texturedregions, while estimating at the same time the parameters of each constituent texture.Last, but not least, the algorithmic contribution is represented by the development ofa new efficient version of the Metropolis Hastings algorithm, with a directional componentof the proposal function based on the”Newton direction” and the Fisher informationmatrix. This particular directional component allows for an efficient exploration of theparameter space and, consequently, increases the convergence speed of the algorithm.To summarize, this work presents a series of methods to solve three image processingproblems in the context of blurry and noisy textured images. Moreover, we present twoconnected contributions, one regarding the texture models andone meant to enhance theperformances of the samplers employed for all of the three methods

This thesis proposes new, efficient methodologies for supervised and unsupervised image segmentation based on texture information. For the supervised case, a technique for pixel classification based on a multi-level strategy that iteratively refines the resulting segmentation is proposed. This strategy utilizes pattern recognition methods based on prototypes (determined by clustering algorithms) and support vector machines. In order to obtain the best performance, an algorithm for automatic parameter selection and methods to reduce the computational cost associated with the segmentation process are also included. For the unsupervised case, the previous methodology is adapted by means of an initial pattern discovery stage, which allows transforming the original unsupervised problem into a supervised one. Several sets of experiments considering a wide variety of images are carried out in order to validate the developed techniques.
Esta tesis propone metodologías nuevas y eficientes para segmentar imágenes a partir de información de textura en entornos supervisados y no supervisados. Para el caso supervisado, se propone una técnica basada en una estrategia de clasificación de píxeles multinivel que refina la segmentación resultante de forma iterativa. Dicha estrategia utiliza métodos de reconocimiento de patrones basados en prototipos (determinados mediante algoritmos de agrupamiento) y máquinas de vectores de soporte. Con el objetivo de obtener el mejor rendimiento, se incluyen además un algoritmo para selección automática de parámetros y métodos para reducir el coste computacional asociado al proceso de segmentación. Para el caso no supervisado, se propone una adaptación de la metodología anterior mediante una etapa inicial de descubrimiento de patrones que permite transformar el problema no supervisado en supervisado. Las técnicas desarrolladas en esta tesis se validan mediante diversos experimentos considerando una gran variedad de imágenes.

Brain tumors are part of a group of common, non-communicable, chronic and potentially lethal diseases affecting mostfamilies in Europe. Imaging plays a central role in brain tumor management, from detection and classification to staging andcomparison.
Increasingly, magnetic resonance imaging (MRI) scan is being used for suspected brain tumors, because in addition tooutline the normal brain structures in great detail, has a high sensitivity for detecting the presence of, or changes within, a tumor.Currently most of the process related to brain tumors such as diagnosis, therapy, and surgery planning are based on its previoussegmentation from MRI. Brain tumor segmentation from MRI is a challenging task that involves various disciplines. The tumors to besegmented are anatomical structures, which are often non-rigid and complex in shape, vary greatly in size and position, and exhibitconsiderable variability from patient to patient. Moreover, the task of labeling brain tumors in MRI is highly time consuming and thereexists significant variation between the labels produced by different experts.
The challenges associated with automated brain tumor segmentation have given rise to many different segmentationapproaches. Although the reported accuracy of the proposed methods is promising, these approaches have not gained wide acceptance among the neuroscientists for every day clinical practice. Two of the principal reasons are the lack of standardizedprocedures, and the deficiency of the existing methods to assist medical decision following a technician way of work.
For a brain tumor segmentation system has acceptance among neuroscientists in clinical practice, it should supportmedical decision in a transparent and interpretable way emulating the role of a technician, considering his experience and knowledge. This includes knowledge of the expected appearance, location, variability of normal anatomy, bilateral symmetry, andknowledge about the expected intensities of different tissues. The image related problems and the variability in tissue distribution among individuals in the human population makes that some degree of uncertainty must be considered together with segmentationresults.
A possible solution for designing complex systems, in which it is required to incorporate the experience of an expert, or the related concepts appear uncertain, is the use of soft computing techniques such as fuzzy systems. An important advantage of fuzzysystems is their ability for handling vague information.
In this work, it is proposed the development of a method to assist the specialists in the process of segmenting braintumors. The main objective is to develop a system that can follow a technician way of work, considering his experience andknowledge. More concretely, it is presented a fully automatic and unsupervised segmentation method, which considers humanknowledge. The method successfully manages the ambiguity of MR image features being capable of describing knowledge about thetumors in vague terms. The method was developed making use of the powerful tools provided by fuzzy set theory.
This thesis presents a step-by-step methodology for the automatic MRI brain tumor segmentation. For achieving the fullyautomatic and unsupervised segmentation, objective measures are delineated by means of adaptive histogram thresholds for defining the non-tumor and tumor populations. For defining the tumor population a symmetry analysis is conducted.
The proposed approach introduces a new way to automatically define the membership functions from the histogram. The proposed membership functions are designed to adapt well to the MRI data and efficiently separate the populations. Since any post-processing is needed, and the unique pre-processing operation is the skull stripping, the proposed segmentation technique reduces the computational times. The proposed approach is quantitatively comparable to the most accurate existing methods, even thoughthe segmentation is done in 2D.

Un sonar latéral de cartographie enregistre les signaux qui ont été rétrodiffusés par le fond marin sur une large fauchée. Les signaux sont ainsi révélateurs de l'interaction entre l'onde acoustique émise et le fond de la mer pour une large plage de variation de l'angle de rasance. L'analyse des statistiques de ces signaux rétrodiffusés montre une dépendance à ces angles de rasance, ce qui pénalise fortement la segmentation des images en régions homogènes. Pour améliorer cette segmentation, l'approche classique consiste à corriger les artefacts dus à la formation de l'image sonar (géométrie d'acquisition, gains variables, etc.) en considérant un fond marin plat et en estimant des lois physiques (Lambert, Jackson, etc.) ou des modèles empiriques. L'approche choisie dans ce travail propose de diviser l'image sonar en bandes dans le sens de la portée ; la largeur de ces bandes étant suffisamment faible afin que l'analyse statistique de la rétrodiffusion puisse être considérée indépendante de l'angle de rasance. Deux types d'analyse de texture sont utilisés sur chaque bande de l'image. La première technique est basée sur l'estimation d'une matrice des cooccurrences et de différents attributs d'Haralick. Le deuxième type d'analyse est l'estimation d'attributs spectraux. La bande centrale localisée à la moitié de la portée du sonar est segmentée en premier par un réseau de neurones compétitifs basé sur l'algorithme SOFM (Self-Organizing Feature Maps) de Kohonen. Ensuite, la segmentation est réalisée successivement sur les bandes adjacentes, jusqu'aux limites basse et haute de la portée sonar. A partir des connaissances acquises sur la segmentation de cette première bande, le classifieur adapte sa segmentation aux bandes voisines. Cette nouvelle méthode de segmentation est évaluée sur des données réelles acquises par le sonar latéral Klein 5000. Les performances de segmentation de l'algorithme proposé sont comparées avec celles obtenues par des techniques classiques.

Les cellules sont les éléments constitutifs de base de tout organisme vivant. Tous les organismes vivants partagent des processus vitaux tels que croissance, développement, mouvement, nutrition, excrétion, reproduction, respiration et réaction à l’environnement. En biologie cellulaire, comprendre la structure et fonction des cellules est essentielle pour développer et tester de nouveaux médicaments. Par ailleurs, cela aide aussi à l’étude du développement embryonnaire. Enfin, cela permet aux chercheurs de mieux comprendre les effets des mutations et de diverses maladies. La vidéo-microscopie (Time Lapse Fluorescence Microscopie) est l’une des techniques d’imagerie les plus utilisées afin de quantifier diverses caractéristiques des processus cellulaires, à savoir la survie, la prolifération, la migration ou la différenciation cellulaire. En vidéo-microscopie, non seulement les informations spatiales sont disponibles, mais aussi les informations temporelles en réitérant l’acquisition de l’échantillon, et enfin les informations spectrales, ce qui génère des données dites « cinq dimensions » (X, Y, Z + temps + canal). En règle générale, les jeux de données générés consistent en plusieurs (centaines ou milliers) d’images, chacune contenant des centaines ou milliers d’objets à analyser. Pour effectuer une quantification précise et à haut débit des processus cellulaires, les étapes de segmentation et de suivi des noyaux cellulaires doivent être effectuées de manière automatisée. Cependant, la segmentation et le suivi des noyaux sont des tâches difficiles dû notamment au bruit intrinsèque dans les images, à l’inhomogénéité de l’intensité, au changement de forme des noyaux ainsi qu’à un faible contraste des noyaux. Bien que plusieurs approches de segmentation des noyaux aient été rapportées dans la littérature, le fait de traiter le bruit intrinsèque reste la partie la plus difficile de tout algorithme de segmentation. Nous proposons un nouvel algorithme de débruitage 3D, basé sur l’apprentissage d’un dictionnaire non supervisé et une représentation parcimonieuse, qui à la fois améliore la visualisation des noyaux très peu contrastés et bruités, mais aussi détecte simultanément la position de ces noyaux avec précision. De plus, notre méthode possède un nombre limité de paramètres, un seul étant critique, à savoir la taille approximative des objets à traiter. Le cadre de la méthode proposée comprend le débruitage d’images, la détection des noyaux et leur segmentation. Dans l’étape de débruitage, un dictionnaire initial est construit en sélectionnant des régions (patches) aléatoires dans l’image originale, puis une technique itérative est implémentée pour mettre à jour ce dictionnaire afin d’obtenir un dictionnaire dont les éléments mis à jour présentent un meilleur contraste. Ensuite, une carte de détection, basée sur le calcul des coefficients du dictionnaire utilisés pour débruiter l’image, est utilisée pour détecter le centre approximatif des noyaux qui serviront de marqueurs pour la segmentation. Ensuite, une approche basée sur le seuillage est proposée pour obtenir le masque de segmentation des noyaux. Finalement, une approche de segmentation par partage des eaux contrôlée par les marqueurs est utilisée pour obtenir le résultat final de segmentation des noyaux. Nous avons créé des images synthétiques 3D afin d’étudier l’effet des paramètres de notre méthode sur la détection et la segmentation des noyaux, et pour comprendre le mécanisme global de sélection et de réglage de ces paramètres significatifs sur différents jeux de données
Cells are the basic building blocks of all living organisms. All living organisms share life processes such as growth and development, movement, nutrition, excretion, reproduction, respiration and response to the environment. In cell biology research, understanding cells structure and function is essential for developing and testing new drugs. In addition, cell biology research provides a powerful tool to study embryo development. Furthermore, it helps the scientific research community to understand the effects of mutations and various diseases. Time-Lapse Fluorescence Microscopy (TLFM) is one of the most appreciated imaging techniques which can be used in live-cell imaging experiments to quantify various characteristics of cellular processes, i.e., cell survival, proliferation, migration, and differentiation. In TLFM imaging, not only spatial information is acquired, but also temporal information obtained by repeating imaging of a labeled sample at specific time points, as well as spectral information, that produces up to five-dimensional (X, Y, Z + Time + Channel) images. Typically, the generated datasets consist of several (hundreds or thousands) images, each containing hundreds to thousands of objects to be analyzed. To perform high-throughput quantification of cellular processes, nuclei segmentation and tracking should be performed in an automated manner. Nevertheless, nuclei segmentation and tracking are challenging tasks due to embedded noise, intensity inhomogeneity, shape variation as well as a weak boundary of nuclei. Although several nuclei segmentation approaches have been reported in the literature, dealing with embedded noise remains the most challenging part of any segmentation algorithm. We propose a novel 3D denoising algorithm, based on unsupervised dictionary learning and sparse representation, that can both enhance very faint and noisy nuclei, in addition, it simultaneously detects nuclei position accurately. Furthermore, our method is based on a limited number of parameters, with only one being critical, which is the approximate size of the objects of interest. The framework of the proposed method comprises image denoising, nuclei detection, and segmentation. In the denoising step, an initial dictionary is constructed by selecting random patches from the raw image then an iterative technique is implemented to update the dictionary and obtain the final one which is less noisy. Next, a detection map, based on the dictionary coefficients used to denoise the image, is used to detect marker points. Afterward, a thresholding-based approach is proposed to get the segmentation mask. Finally, a marker-controlled watershed approach is used to get the final nuclei segmentation result. We generate 3D synthetic images to study the effect of the few parameters of our method on cell nuclei detection and segmentation, and to understand the overall mechanism for selecting and tuning the significant parameters of the several datasets. These synthetic images have low contrast and low signal to noise ratio. Furthermore, they include touching spheres where these conditions simulate the same characteristics exist in the real datasets. The proposed framework shows that integrating our denoising method along with classical segmentation method works properly in the context of the most challenging cases. To evaluate the performance of the proposed method, two datasets from the cell tracking challenge are extensively tested. Across all datasets, the proposed method achieved very promising results with 96.96% recall for the C.elegans dataset. Besides, in the Drosophila dataset, our method achieved very high recall (99.3%)

碩士
國立成功大學
電機工程學系碩博士班
96
As digital audio and video databases increasing, how to effectively manage all the more important. CBIR (content based image retrieval) of video content-based image retrieval system. The images have to cut in CBIR system is also very important, all the Query image will implement this step. So my main goal is to make a picture from the background and prospects, than to make CBIR can improve efficiency and better results.

Image segmentation is the process of partitioning an image into meaningful parts. Image segmentation is used to locate objects and boundaries in images. It is the process of assigning a label to every pixel in an image such that pixels with the same label share certain visual characteristics. The need for accurate segmentation tools in medical applications is driven by the increased capacity of the imaging devices. Due to high resolutions and a large number of image slices CT and MRI generated images cannot be examined manually. Furthermore, it is very difficult to visualize complex structures in three-dimensional image volumes without cutting away large portions of, perhaps important, data. Tools, such as segmentation, can aid the medical staff in browsing through such large images by highlighting objects of particular importance. In addition, segmentation in particular can output models of organs, tumors, and other structures for further analysis, quantification or simulation. We have used k means, fuzzy c means for better performance we map the input space onto a self-organising map and then the low dimensional input is clustered using the above methods. A self-organising map (SOM) is a type of artificial neural network that is trained using unsupervised learning to produce a low-dimensional (typically two-dimensional), discretized representation of the input space of the training samples, called a map. This thesis is devoted to medical image segmentation techniques and their applications in clinical and research settings.

碩士
國立清華大學
資訊工程學系
104
Image segmentation is an important issue in image editing and computer vision. Due to the complexity of information in images, efficient extraction of a foreground object is a challenging problem. Recently, several approaches based on optimization by graph cuts have been developed which successfully combine the color feature with the edge information. A problem is that the segmentation results heavily depend on the seeds selection. However, it is difficult to obtaining reliable seeds automatically. To overcome this problem, we propose an automatic scheme for image segmentation. Compare to the classical binary-label graph cuts, the results by the multi-label graph cuts do not heavily depend on the seeds selection. Our method uses the multi-label graph cuts to separate an image into multiple segments, and then classify the segments into the object and the background. We introduce the standard deviation to adapt the importance between the properties in our method. Experiments show that the proposed method yields more accurate segmentation results than the previous automatic approach and is comparable to the interactive approach.

碩士
國立清華大學
資訊工程學系
101
"Co-segmentation" can increase the accuracy of object recognition. The concept of co-segmentation is the problem of simultaneously dividing multiple images into common object, reference each other to segment similar region as an object. In recent years, the problem of image co-segmentation has been widely discussed. In our paper, we believe that each image pre-processing can be divided to many appropriate segments, and then co-segmentation will get the better results. At beginning, we segment each image into number of suerpixels, and extract their color histogram features. And we follow the concept of hierarchical clustering, we merge pair of superpixels which most similar with each other into one superpixel in each iteration until the appropriate threshold. This is not only ensure the superpixels which merge together have same material, but also effective in reducing the amount of computation. In addition, each superpixel records the maximum relative distance. The value can be used as a range to increase the accuracy of our co-matching method. Finally, we use GrowCut to get the final result. The results show that our method can not only achieve better results, but also we don’t need to add any setting, it is a good way for user that produces results automatically.

碩士
國立清華大學
資訊工程學系
104
Image segmentation is an important and difficult issue in computer vision and image processing. It categorized into two categories, supervised image segmentation and unsupervised image segmentation. The supervised methods need some interactions of users. It makes those methods inconvenient. Recently, most of segmentation methods usually use similarity which is defined by color difference or histogram. Every similarity has its weak side. In this paper, we proposed an unsupervised method. It uses defocus map, edge and color as similarity of pixels or superpixels. We generate an edge strength map. Then, we construct a minimum spanning tree with the superpixels and the edge map to divide the image to foreground and background. In our experiment, out method doesn’t need user interaction and the performance is better than previous superpixels grouping method.

碩士
國立清華大學
資訊工程學系
100
Co-segmentation is a new topic in computer vision, which has been discussed lively in many literatures. It is defined as the task of jointly segmenting the common objects in a given set of images. Due to there are some limitations in previous methods, this thesis presents a game theoretic unsupervised approach by using the concept of heat diffusion and saliency to solve co-segmentation problem without these limitations. Our method is divided into two stages. First, the common objects discovery task is modeled by a cooperative game. In this game, each image is treated as player. All players want to maximize the overall payoffs (i.e. the gain of heat) by putting the heat sources appropriately. Note that we must ensure that no one will be likely to uncooperative. So we define some collaborative strategies. For each input image, the game structure generates corresponding labeled image which identifies the common objects and background. Then we use cooperative cut to solve energy minimization problem in the second stage. Our method takes advantage of cooperative game theory, which enables us to discover the common objects automatically and accurately. Experimental results demonstrate that in many cases the proposed method can perform much better than state-of-the-art co-segmentation method.

Image segmentation is a classical problem in computer vision and is of paramount importance to medical imaging. Medical image segmentation is an essential step for most subsequent image analysis task. The segmentation of anatomic structure in the brain plays a crucial role in neuro imaging analysis. The study of many brain disorders involves accurate tissue segmentation of brain magnetic resonance (MR) images. Manual segmentation of the brain tissues, namely white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF) in MR images by an human expert is tedious for studies involving larger database. In addition, the lack of clearly defined edges between adjacent tissue classes deteriorates the significance of the analysis of the resulting segmentation. The segmentation is further complicated by the overlap of MR intensities of different tissue classes and by the presence of a spatially and smoothly varying intensity in-homogeneity. The prime objective of this dissertation is to develop strategies and methodologies for an automated brain MR image segmentation scheme.

碩士
國立清華大學
資訊系統與應用研究所
105
Image saliency detection is a process to pop out the most salient part in the image, and shows up with image saliency map. However, some image saliency maps are not accurate enough to separate foreground and background from images with low contrast; dark channel prior (DCP) can transform these image into a clear image. In this paper, we first apply DCP in image saliency detection to emphasize foreground from image with low contrast saliency. Moreover, we propose a simple cutting method on image saliency. We convert the saliency map into a histogram and use a first degree polynomial to smooth the histogram. The deepest and widest valley of the smoothed histogram is chosen as the cutting threshold. The part higher than threshold is identified as foreground, and the other is background. In our experiment, it proves that the proposed method successfully segments the foreground and background from the image.

碩士
國立交通大學
資訊科學與工程研究所
95
Building an automatic system for news story segmentation is an important and challenging task. A news story is composed of an anchorperson shot and a news footage shot, we can segment a news video into several stories if we know when the anchorperson shows up. This paper presents a method for anchorperson detection based on face detection. First, detecting human faces region in every news frame. Then, extracting features by the face region, and clustering on all features. Suppose that the biggest cluster is presented for anchorperson。This method would not effected by the complex background because it focuses only on the face region. And because of its unsupervised nature, the algorithm does not need to adjust model for different anchorpersons. The efficacy of the proposed method is tested on 5 h of news programs. Moreover, we integrate the proposed method to an existed news video library system and segmenting on the ETT news programs successfully.

碩士
國立交通大學
資訊科學與工程研究所
101
Recently, an unsupervised image segmentation framework, Segmentation by Aggregating Superpixels (SAS) [2] is proposed and shown to be very promising. However, the texture cues, which have been shown to be very effective in many researches [14-18], are absent in [2]. In this thesis, we propose an effective method for incorporating texture information into the SAS framework, using superpixels. To extract texture information, our algorithm first uses texture filtering and subsequent Gaussian Mixture Models (GMM) clustering which is modified from [16]. Then, we develop an edge-aware low-pass filtering to generate multiple-scale texture superpixels from GMM clustering results. Finally, by joining texture superpixels with the superpixel set originally used in [2], the incorporation of texture information is accomplished. Our method achieves superior performance on the well-known Berkeley Segmentation Dataset (BSDS) under multiple prevailing region-based evaluation criteria when compared to other benchmark algorithms.

碩士
國立交通大學
電子研究所
105
In this thesis, we present an unsupervised hierarchical clustering algorithm based on a split-and-merge scheme. Using image segmentation as an example of the applications, we propose an unsupervised image segmentation algorithm which outperforms the existing algorithms. In the split phase, we propose an efficient partition algorithm, named Just-Noticeable-Difference Bayesian Sequential Partitioning (JND-BSP), to partition image pixels into a few regions, within which the color variations are perceived to be smoothly changing without apparent color differences. In the merge phase, we proposed a Probability Based Sequential Merging algorithm to sequentially construct a hierarchical structure that represents the relative similarity among these partitioned regions. Instead of generating a segmentation result with a fixed number of segments, the new algorithm produces an entire hierarchical representation of the given image in a single run. This hierarchical representation is informative and can be very useful for subsequent processing, like object recognition and scene analysis. To demonstrate the effectiveness and efficiency of our method, we compare our new segmentation algorithm with several existing algorithms. Experiment results show that our new algorithm can not only offers a more flexible way to segment images but also provides segmented results close to human’s visual perception. The proposed algorithm can also be widely used on applications analyzing other types of data, and can be used to analyze Big Data with high dimension efficiently.

A thesis submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in fulfilment of the requirements for the degree of Doctor of Philosophy. Johannesburg, March 1994.
Region segmentation of digital imges by unsupervised thresholding is a common, conceptually simple and important branch of image processing and analysis. Its applications range from that of simple binarization to serving as a useful pre-processing stage for operations such as pattern recognition and image restoration. While many different algorithms have been proposed for the automatic selection of the "correct" threshold the results vary widely in their general usefulness. A class of selection schemes is based on the principle of maximum entropy. This formalism, While effective, is usually involed without reference to its origins or its relationship to images. This thesis attempts to clarify the definition of what is meant by the entropy of an image, to which end various image and Image segmentation models are discussed and proposed. Some apparent shortcomings related to the use of the Shannon entropy formula are addressed and the outcome of the research is applied to the problem of threshold selection. The results indicate a marked improvement in performance of methods using some form(s) of context-related information over those which simply apply the entropy formula without regard to its spatially insensitive nature. Evaluation of results and processes is usually based

博士
國立臺灣科技大學
電機工程系
102
In this thesis, we proposed to perform content-based image retrieval (CBIR) on Internet scale databases connected through peer-to-peer (P2P) networks, abbreviated as P2P-CBIR, which utilizes an intelligent preprocessing to identify the object regions and provides scalable retrieval function. For preprocessing, we proposed a dual multiScalE Graylevel mOrphological open and close recoNstructions (SEGON) algorithm, and utilized edge coverage rate to segment foreground (FG) object regions in one image. To improve FG object segmentation accuracy, a background (BG) gray-level variation mesh is built. The SEGON was developed from a macroscopic perspective on image BG gray levels and implemented through regular procedures to deal with large-scale database images. To evaluate the segmentation accuracy, the probability of coherent segmentation labeling, i.e., the normalized probability random index (PRI), between a computer-segmented image and the hand-labeled one is computed for comparisons. Experiments showed that the proposed object segmentation method outperforms others in the PRI performance. The normalized correlation coefficient of features among query samples was computed to integrate the similarity ranks of different features in order to perform multi-instance queries with multiple features (MIMF). Retrieval precision–recall (PR) and rank performances, with and without SEGON, were compared. Performing SEGON-enabled CBIR on large-scale databases yields higher PR performance. For performing Internet scale CBIR, a P2P-CBIR system has been proposed, which helps to effectively explore the large-scale image database distributed over connected peers. The decentralized unstructured P2P network topology is adopted to compromise with the structured one, and informed-like instead of blind-like searching approach enables flexible routing control when peers join/leave or network fails. The P2P-CBIR adopts MIMF to reduce average network traffics while maintaining high retrieval accuracy on the query peer. In addition, scalable retrieval control can also be developed based on the P2P-CBIR framework, which can adapt the query scope and progressively refine the accuracy during the retrieval process. We also proposed to record instant local database characteristics of peers for the P2P-CBIR system to update peer linking information. By reconfiguring system at each regular interval time, we can effectively reduce trivial peer routing and retrieval operations due to imprecise configurations. We also proposed to optimally configure the P2P-CBIR system such that, under a certain number of online users, which would yield the highest recall rate. Experiments show that the average recall rate of the proposed P2P-CBIR method with reconfiguration is higher than that of the one without, and the latter outperforms previous methods, under the same retrieval scope, i.e., same time-to-live (TTL) settings. Furthermore, simulations demonstrate that, with the optimal configuration, recall rates can be improved while the network traffic of each peer is reduced, under the same number of on-line users.

Breast cancer is among the well-researched type compared to other common types of cancer. However, there still remain important open issues for investigation. One of these issues is the clarification of the importance of certain biological factors, such as histological tumour grade and estrogens reception (ER) status, to clinical management of the disease. Until now, histological grading and ER status assessment is based on the visual evaluation of breast tissue specimens under the microscope. More specifically, grading is determined on the visual estimation of certain histological features, on H&E (Hematoxylin & Eosin) stained specimens according to the World Health Organization (WHO) guidelines, whereas ER-status is assessed as the percentage of expressed nuclei on immunohistochemically stained (IHC) specimens as suggested by the American Society of Clinical Oncology (ASCO) protocol. Recent studies have attempted to examine whether histological tumour grade relates to ER status. Such a relation seems to be of importance in the various treatment strategies followed in breast tumours. However, the quantification of ER status presents certain weaknesses: a) there is a lack of consensus among experts regarding the protocol to be followed for calculating the ER status; b) an exact estimate of the ER status is difficult to be obtained, since the latter would require manual counting of positively expressed nuclei. In clinical practice often a gross estimate is obtained by the histopathologists through visual inspection on representative specimen areas. Consequently, the evaluation of ER status, which has been considered by previous studies as the key measure for assessing the correlation between ERs and tumour grade, is prone to the physician’s subjective estimation. Therefore, more reliable methods are needed. This thesis has been carried out in the search of such alternative, more reliable, methods. Accordingly, the aims of the present thesis are: (i) to develop a reliable segmentation methodology for detection of ER-expressed nuclei in breast cancer tissue images stained with IHC, (ii) to objectively quantify ER status in breast cancer tissue images stained with IHC, (iii) to investigate potential correlation between ER status and histological grade by combining information from IHC and H&E stained breast cancer tissue images obtained from the same patient, (iv) to establish evidence for linking chromatin texture variations with textural variations on ER-expressed nuclei, (v) to investigate the potential of the proposed hybrid supervised pattern recognition strategies to other challenging fields of medical image processing and analysis. To address the above issues and in search of reliable methods for quantitatively assessing ER status and its correlation with histological grade based, a novel hybrid (unsupervised-supervised) pattern recognition methodology has been designed, developed and implemented for the analysis of breast cancer tissue images. Moreover, it will be shown that proper modification of the proposed methodology may result to generalize pixel classification approach suitable for processing and analysis of medical images other than microscopic such as Computed Tomography Angiography images.
Σε σχέση με άλλες μορφές καρκίνου, ο καρκίνος του μαστού είναι μεταξύ των ευρέως μελετημένων τύπων καρκίνου, ωστόσο, υπάρχουν ακόμη σημαντικά ανοικτά ζητήματα προς διερεύνηση. Ένα από αυτά τα είναι ο προσδιορισμός της σπουδαιότητας ορισμένων βιολογικών παραγόντων, όπως ο βαθμός διαφοροποίησης της κακοήθειας (ΒΔΚ) του όγκου και το επίπεδο έκφρασης των Οιστρογονικών Υποδοχέων (ΟΥ), στην κλινική διαχείριση της νόσου. Μέχρι τώρα, η εκτίμηση του ΒΔΚ του όγκου και της έκφρασης των ΟΥ είναι βασισμένη στην οπτική αξιολόγηση ιστολογικών δειγμάτων, τα οποία λαμβάνονται από αντιπροσωπευτικές περιοχές του μαστού, στο μικροσκόπιο. Συγκεκριμένα, σύμφωνα με τις οδηγίες του Παγκόσμιου Οργανισμού Υγείας, ο ΒΔΚ του όγκου καθορίζεται από την οπτική εκτίμηση ορισμένων ιστολογικών χαρακτηριστικών γνωρισμάτων σε ιστολογικά δείγματα που έχουν υποστεί χρώση Αιματοξυλίνης - Ηωσίνης (Heamatoxylin & Eosin-Η&Ε), ενώ σύμφωνα με τις οδηγίες της Αμερικάνικης Εταιρείας Κλινικής Ογκολογίας, η έκφραση των ΟΥ πρέπει να εκτιμάται ως το εκατοστιαίο ποσοστό των εκφρασμένων πυρήνων σε δείγματα βαμμένα με ανοσοϊστοχημικές τεχνικές (Immunohistochemistry-IHC). Πρόσφατες μελέτες έχουν προσπαθήσει να εντοπίσουν εάν υπάρχει σύνδεση μεταξύ του ΒΔΚ του όγκου και της έκφρασης των ΟΥ στον όγκο, συσχετίζοντας τον ΒΔΚ από εικόνες με χρώση H&E με τον ποσοστό των εκφρασμένων ΟΥ σε δείγματα IHC. Αυτή η συσχέτιση φαίνεται να είναι σημαντική στις διάφορες ακολουθούμενες στρατηγικές για τη θεραπεία του καρκίνου του μαστού. Εντούτοις, ο προσδιορισμός της έκφρασης των ΟΥ παρουσιάζει ορισμένες αδυναμίες: α) υπάρχει σημαντική μεταβλητότητα μεταξύ των ειδικών σχετικά με το πρωτόκολλο που ακολουθείται για τον υπολογισμό της έκφρασης των ΟΥ, β) είναι δύσκολο να εκτιμηθεί με ακρίβεια η έκφραση των ΟΥ, δεδομένου ότι θα απαιτούσε τη μέτρηση του συνόλου των θετικά εκφρασμένων πυρήνων από τον ειδικό ιστοπαθολόγο. Στην κλινική πράξη, λαμβάνεται συνήθως μια χονδρική εκτίμηση από τον ιστοπαθολόγο, μέσω μικροσκοπίου, παρατηρώντας αντιπροσωπευτικές περιοχές των δειγμάτων όπου υπάρχει μεγάλη συγκέντρωση εκφρασμένων πυρήνων σε ΟΥ. Ως εκ τούτου, η αξιολόγηση της έκφρασης των ΟΥ, που έχει θεωρηθεί από προηγούμενες μελέτες ως βασική μέτρηση για τη συσχέτιση μεταξύ ΟΥ και του βαθμού διαφοροποίησης των όγκων, είναι επιρρεπής στην υποκειμενικότητα του ειδικού. Για τον λόγο αυτό απαιτούνται πιο αξιόπιστες μέθοδοι. Η παρούσα διατριβή πραγματοποιήθηκε σε αναζήτηση εναλλακτικών, πιο αξιόπιστων μεθόδων. Έτσι οι στόχοι της παρούσας διατριβής είναι: (i) η ανάπτυξη μιας αξιόπιστης μεθοδολογίας τμηματοποίησης ιστολογικών εικόνων μικροσκοπίας επεξεργασμένες με χρώση IHC για τον εντοπισμό των πυρήνων που εκφράζουν τους ΟΥ για την αντικειμενική ποσοτικοποίηση της έκφρασης των ΟΥ στον καρκίνο του μαστού, (ii) η διερεύνηση ενδεχόμενης σχέσης μεταξύ της έκφρασης των ΟΥ και του ΒΔΚ του όγκου, συνδυάζοντας την πληροφορία των ιστολογικών δειγμάτων, που προέρχονται από τον καρκινικό ιστό του ίδιου ασθενούς και έχουν υποστεί επεξεργασία με ανοσοϊστοχημική χρώση και με χρώση H&E, (iii) η διερεύνηση πιθανής συσχέτισης στις μεταβολές της υφής της χρωματίνης με τις μεταβολές στην υφή των πυρήνων που εκφράζουν τους ΟΥ, και (iv) η διερεύνηση της δυνατότητας της προτεινόμενης μεθοδολογίας σε άλλους τομείς επεξεργασίας και ανάλυσης ιατρικών εικόνων. Για την εκπλήρωση των ανωτέρω στόχων και σε αναζήτηση αξιόπιστων μεθόδων για την ποσοτικοποίηση της έκφρασης των ΟΥ και της σύνδεσή της με το ΒΔΚ του όγκου, σχεδιάστηκε, αναπτύχθηκε και εφαρμόστηκε μια νέα μεθοδολογία βασισμένη στην αναγνώριση προτύπων ημι-εποπτευόμενης μάθησης για την ανάλυση ιστοπαθολογικής εικόνας. Επιπλέον, η κατάλληλη τροποποίηση της προτεινόμενης μεθόδου μπορεί να οδηγήσει στη γενίκευση της μεθοδολογικής προσέγγισης της ταξινόμησης εικονοστοιχείων για την επεξεργασία και την ανάλυση ιατρικών εικόνων, πέρα αυτών της μικροσκοπίας, όπως εικόνες από Aγγειογραφία Υπολογιστικής Τομογραφίας.

碩士
國立臺灣大學
生醫電子與資訊學研究所
105
After the surgery being taken, the after care of the surgical wound has a great impact toward the patients’ prognosis. It’s often takes few days even few weeks for the wound to stabilize. It’s is a great cost of health care and nursing resources. The advance of image process and machine learning improves the accuracy of wound assessment and analysis and there are some recent works started on this field of wound analysis. In our tele-health scenario, we hope the user can use their mobile device to obtain an accurate result without using high-end camera. In this literature, we proposed an image segmentation algorithm based on edge detection and Hough transform. We further developed an optimization method based on unsupervised image segmentation evaluation and genetic algorithm. The result was evaluated by the image provided by NTUH, division of surgery. We also implemented an analysis system cooperate with NTUH telehealth center, which has been used on pacemaker implantation patient. The result of performing this segmentation algorithm on the data set provided by NTUH, Division of cardiovascular surgery, achieve the accuracy of 75.7%, after the optimization of genetic algorithm it achieves 94.3%.

碩士
朝陽科技大學
資訊管理系碩士班
100
The morbidity and mortality of cervical cancer can be reduced by the screening of the precancerous lesions. Pap smears, colposcopy and biopsy are the most common screening tools. Pap smear is the first-line tool because of its high specificity and low cost. But its false-positive rate is too high and must be confirmed by other tools. Biopsy is a deterministic examination for cervical neoplasia. However, it is not suitable for the high probability of false-positive. Digital colposcopy is a promising technology for the detection of cervical intraepithelial neoplasia. However, there are no quantitative criteria for the differential of precancerous lesions and it is subjected to the variation of inter-observer and intra-observer. Therefore, automated image analysis of colposcopic images is thus necessary for the improvement of diagnosis of colposcopy. The segmenetation of the lession from digital colposcopyic image is a key issue of this analysis. Our goal is to develop a segmentation policy that can separate images into regions which contain the lesion areas. These areas can be provided to the analysis system and help doctor to make the diagnosis.

The current trend in visual recognition research is to place a strict division between the supervised and unsupervised learning paradigms, which is problematic for two main reasons. On the one hand, supervised methods require training data for each and every category that the system learns; training data may not always be available and is expensive to obtain. On the other hand, unsupervised methods must determine the optimal visual cues and distance metrics that distinguish one category from another to group images into semantically meaningful categories; however, for unlabeled data, these are unknown a priori. I propose a visual category discovery framework that transcends the two paradigms and learns accurate models with few labeled exemplars. The main insight is to automatically focus on the prevalent objects in images and videos, and learn models from them for category grouping, segmentation, and summarization. To implement this idea, I first present a context-aware category discovery framework that discovers novel categories by leveraging context from previously learned categories. I devise a novel object-graph descriptor to model the interaction between a set of known categories and the unknown to-be-discovered categories, and group regions that have similar appearance and similar object-graphs. I then present a collective segmentation framework that simultaneously discovers the segmentations and groupings of objects by leveraging the shared patterns in the unlabeled image collection. It discovers an ensemble of representative instances for each unknown category, and builds top-down models from them to refine the segmentation of the remaining instances. Finally, building on these techniques, I show how to produce compact visual summaries for first-person egocentric videos that focus on the important people and objects. The system leverages novel egocentric and high-level saliency features to predict important regions in the video, and produces a concise visual summary that is driven by those regions. I compare against existing state-of-the-art methods for category discovery and segmentation on several challenging benchmark datasets. I demonstrate that we can discover visual concepts more accurately by focusing on the prevalent objects in images and videos, and show clear advantages of departing from the status quo division between the supervised and unsupervised learning paradigms. The main impact of my thesis is that it lays the groundwork for building large-scale visual discovery systems that can automatically discover visual concepts with minimal human supervision.
text

Coastal upwelling is a process when cold and nutrient-rich water dynamically appears over the surface of the ocean by replacing the warm water. The oceanographers are interested to detect the upwelling regions and corresponding boundaries but to examine the whole process of upwelling they have to work manually on each image, therefore; it increases the workload. The main purpose of this application is to automatically detect the upwelling regions, monitoring environmental changes and the study of fishery resources. The Seed Expanding Clustering algorithm (SEC) (Nascimento et al., 2015) is a thresholding clustering method for automatic detection of upwelling and delineation of its fronts. The self‐tuning thresholding is derived from the clustering criterion and serves as a boundary regularizer of the growing clusters. The SEC algorithm is shown more than 80% of accuracy rate on the unsupervised automatic recognition of the phenomenon. The main contribution of this dissertation is threefold. First, the development of a sequential extraction version of the SEC algorithm with a stop condition that takes advantage of the knowledge domain to select seeds and model extracted features. Second, the development of an explosion control procedure to detect the so-called leakage problem. Third, the development of a fusion scheme of unsupervised clustering validation measures. The experimental comparison of the new iterative version of the SEC algorithm with a new developed iterative version of Adams & Bischof SRG on the unsupervised segmentation of upwelling regions on SST images from different regions of the globe show their competitiveness comparing to other conventional SRG methods.

Pattern recognition has as its objective to classify objects into different categories and classes. It is a fundamental component of artificial intelligence and computer vision. This thesis investigates the application of an efficient optimization method, known as Particle Swarm Optimization (PSO), to the field of pattern recognition and image processing. First a clustering method that is based on PSO is proposed. The application of the proposed clustering algorithm to the problem of unsupervised classification and segmentation of images is investigated. A new automatic image generation tool tailored specifically for the verification and comparison of various unsupervised image classification algorithms is then developed. A dynamic clustering algorithm which automatically determines the "optimum" number of clusters and simultaneously clusters the data set with minimal user interference is then developed. Finally, PSO-based approaches are proposed to tackle the color image quantization and spectral unmixing problems. In all the proposed approaches, the influence of PSO parameters on the performance of the proposed algorithms is evaluated.
Thesis (PhD)--University of Pretoria, 2006.
Computer Science
unrestricted

The expression of genetic information, in all organisms, might be characterized as in a constant state of flux with only a fraction of the gene within a genome being expressed at any given time. The genes’ expression pattern reflects the response of cells to stimuli that control growth, development and signal environmental changes. Understanding genes’ expression at the level of transcription and/or other stages of gene regulation at the mRNA level (half life of mRNA, RNA production from primary transcript) might reveal insights into the genes expression mechanisms that control these changes. With the DNA microarray technology researchers are now able to determine, in a single experiment, the gene expression profiles of hundreds to tens of thousands of genes in tissue, tumors, cells or biological fluids. Accordingly, and since the patterns of gene expression are strongly functionally correlated, microarrays might provide unprecedented information both on basic research (e.g. expression profiles of different tissues) and on applied research (e.g. human diseases, drug and hormone action etc). While the simultaneous measurement of thousands of gene expression levels potentially serves as source of profound knowledge, genes quantification (i.e. extraction of the genes expression levels) is confounded by various types of noise originating both from the microarray experimental procedure (e.g. sample preparation) and the probabilistic characteristics of the microarray detection process (e.g. scanning errors). The “noisy” nature of the measured gene expression levels obscures some of the important characteristics of the biological processes of interest. The latter, as a direct effect, renders the extraction of biological meaningful conclusions through microarray experiments difficult and affects the accuracy of the biological inference. Thus, as a major challenge in DNA microarray analysis, and especially for the accurate extraction of genes expression levels, might be considered the effective separation of “true” gene expression values from noise. Noise reduction is an essential process, which has to be incorporated into the microarray image analysis pipeline in order to minimize the “errors” that propagate throughout the microarray analysis pipeline and, consequently, affect the extracted gene expression levels. A possible solution, as proposed in previous studies, for addressing microarray image noise is image enhancement. Results of these studies have indicated a superior quality of the enhanced images, without however examining whether enhancement leads to more accurate spot segmentation or reduces the variability of the extracted gene expression levels. As foresaid, noise also complicates the extraction of meaningful biological conclusions. While more advanced methods have been introduced [28-32] that attempt to prevent the noisy set of genes from being grouped, there is a lack of consensus among experts on the selection of a single method for determining meaningful clusters of genes. The latter, directly affects the biological inference, since different number of clusters are produced when different clustering techniques or either different parameters in the clustering algorithms are utilized. Thus, it turns up that it is not only important to assess the performance of each analysis stage independently (i.e. whether the techniques employed in the microarray analysis pipeline provide accurate extracted gene expression levels or the clustering techniques group biologically related genes) but it is also necessary to ensure an acceptable performance of all steps, as a whole, in terms of biologically meaningful information. This thesis has been carried out towards the development of a complete microarray image processing and analysis framework in order to improve the extraction and, consequently, the quantification of gene expression levels on spotted complementary DNA (cDNA) microarray images. The aims of the present thesis are: a) to model and address the effects of cDNA microarray image noise in such a way that it will increase the accuracy of the extracted gene expression levels, b) to investigate the impact of noise and facilitate genes expression data analysis in order to allow biologists to develop an integrated understanding of the process being studied, c) to introduce a semi-supervised biologically informed criterion for the detection of meaningful biological clusters of genes that answer specific biological questions, d) to investigate the performance and the impact of various state-of-art and novel cDNA microarray image segmentation techniques in the quantification of genes expression levels For exploring all of these aspects, a complete and robust framework of microarray image processing and analysis techniques was designed, built and implemented. The framework incorporated in the microarray analysis pipeline a novel combination of image processing and analysis techniques originating from the comprehensive quantitative investigation of the impact of noise on spot segmentation, intensity extraction and data mining. Additionally, novel formulations of known image segmentation techniques have been introduced, implemented and evaluated in the task of microarray image segmentation. The usefulness of the proposed methods has been validated experimentally on both simulated and real cDNA microarray images.
Η έκφραση της γενετικής πληροφορίας, σε όλους τους οργανισμούς, χαρακτηρίζεται από μια σταθερή κατάσταση «ροής» στην οποία όμως μόνο ένα μέρος του γονιδίου μέσα στο γονιδίωμα (genome) εκφράζεται ανά χρονική στιγμή. Το γονιδιακό μοτίβο έκφρασης (gene expression pattern or gene expression profile) θα μπορούσαμε να πούμε ότι αντανακλά την αντίδραση των κυττάρων στα διάφορα εξωτερικά ερεθίσματα. Για να μπορέσουν να απαντηθούν ερωτήματα σχετικά με τους μηχανισμούς που επηρεάζουν και μεταβάλλουν τη γονιδιακή έκφραση ανάλογα με το εξωτερικό ερέθισμα είναι απαραίτητη η μελέτη της γονιδιακής έκφρασης σε μεταγραφικό επίπεδο (transcription level) ή/και άλλα στάδια (παράγοντες) που ρυθμίζουν τη γονιδιακή έκφραση (gene regulation) σε επίπεδο mRNA. Με τη χρήση της τεχνολογίας των μικροσυστοιχιών, οι ερευνητές έχουν πλέον τη δυνατότητα να μελετήσουν ταυτόχρονα την γονιδιακή έκφραση δεκάδων ή και εκατοντάδων χιλιάδων γονιδίων σε ιστούς, κύτταρα όγκους κλπ με τη χρήση ενός και μόνο πειράματος. Κατά συνέπεια, και από τη στιγμή που τα γονιδιακά μοτίβα έκφρασης συσχετίζονται έντονα λειτουργικά (functionally correlated), η τεχνολογία των μικροσυστοιχιών παρέχει ανεκτίμητης αξίας πληροφορίες που μπορούν να δώσουν ώθηση τόσο στην ανάπτυξη της βασικής έρευνας π.χ. μελέτη των γονιδιακών προφίλ έκφρασης διαφορετικών ιστών όσο και στην ανάπτυξη της εφαρμοσμένης έρευνας π.χ. μελέτη ασθενειών, δράση φαρμάκων και ορμονών κλπ. Παρά τη δυνατότητα που παρέχει η τεχνολογία των μικροσυστοιχιών για την ταυτόχρονη μέτρηση των επιπέδων έκφρασης χιλιάδων γονιδίων, η ποσοτικοποίηση της γονιδιακής έκφρασης (δηλ. η εξαγωγή των επιπέδων έκφρασης των γονιδίων), επηρεάζεται από τους διάφορους τύπους θορύβου που υπεισέρχονται τόσο κατά την πειραματική διαδικασία κατασκευής των μικροσυστοιχιών (π.χ. προετοιμασία δειγμάτων) όσο και από τα πιθανοκρατικά χαρακτηριστικά που διέπουν τη διαδικασία ανίχνευσης (microarray scanning procedure) των μικροσυστοιχιών (π.χ. λάθη ανίχνευσης). Η «θορυβώδης» φύση των γονιδίων και κατά συνέπεια των μετρούμενων γονιδιακών εκφράσεων «κρύβει» (obscure) μερικά από τα πιο σημαντικά χαρακτηριστικά των βιολογικών διαδικασιών ενδιαφέροντος και καθιστά δύσκολη την εξαγωγή χρήσιμων βιολογικών συμπερασμάτων. Από τα παραπάνω διαφαίνεται ότι η μείωση του θορύβου είναι μια πολύ σημαντική διαδικασία η οποία θα πρέπει να ενσωματωθεί στην αλγοριθμική μεθοδολογία που μέχρι στιγμής χρησιμοποιείται για την εξαγωγή των γονιδιακών εκφράσεων από τις εικόνες μικροσυστοιχιών. Με αυτό τον τρόπο θα ελαχιστοποιηθούν τα πιθανά «λάθη» τα οποία μεταφέρονται (propagate) κατά τη διαδικασία εξαγωγής των εντάσεων (μέσω της χρησιμοποιούμενης αλγοριθμικής μεθοδολογίας) και τελικά επηρεάζουν την «ακριβή» εξαγωγή των γονιδιακών εκφράσεων. ‘Ως πιθανή λύση για την αντιμετώπιση του θορύβου στις εικόνες μικροσυστοιχιών, έχει προταθεί στη διεθνή βιβλιογραφία η χρήση τεχνικών αναβάθμισης εικόνας. Τα αποτελέσματα αυτών των επιστημονικών εργασιών συμπεραίνουν ότι με τη χρήση τεχνικών αναβάθμισης η ποιότητα των επεξεργασμένων εικόνων είναι σαφώς καλύτερη. Ωστόσο, καμία από αυτές τις εργασίες δεν μελετάει εάν οι τεχνικές αναβάθμισης οδηγούν στον ακριβέστερο προσδιορισμό των παρυφών των κουκίδων (spot) από τις οποίες εξάγονται οι γονιδιακές εκφράσεις ή εάν βοηθάνε στη μείωση της μεταβλητότητας (variability) των εξαγόμενων γονιδιακών εκφράσεων. Επιπρόσθετα, όπως έχει ήδη προαναφερθεί, ο θόρυβος παρεμποδίζει την εξαγωγή χρήσιμων βιολογικών συμπερασμάτων. Παρά το μεγάλο πλήθος εξελιγμένων μεθόδων που έχουν προταθεί στη διεθνή βιβλιογραφία για την αποτροπή της ομαδοποίησης γονιδίων που χαρακτηρίζονται ως «θορυβώδη», δεν έχει καθοριστεί ακόμα (από τους ειδικούς) μια ενιαία μέθοδος που να βρίσκει και να ομαδοποιεί γονίδια τα οποία θα παρέχουν βιολογικά χρήσιμες πληροφορίες. Αποτέλεσμα αυτής της «ασυμφωνίας» μεταξύ των ειδικών αποτελεί η εξαγωγή διαφορετικών βιολογικών συμπερασμάτων ανάλογα α) με τον αριθμό των δημιουργούμενων γονιδιακών ομάδων (που εξαρτάται άμεσα από τη διαφορετική μέθοδο ομαδοποίησης (clustering)) και β) με τις διαφοροποιήσεις που μπορεί να έχουμε στις παραμέτρους των διαφόρων μεθόδων ομαδοποίησης. H παρούσα διατριβή στοχεύει στη δημιουργία ενός ολοκληρωμένου πλαισίου για την επεξεργασία και ανάλυση εικόνων μικροσυστοιχιών με σκοπό την βελτιστοποίηση της εξαγωγής και κατά συνέπεια της ποσοτικοποίησης των γονιδιακών εντάσεων από εικόνες μικροσυστοιχιών κουκίδων (spotted cDNA microarray images). Οι στόχοι της παρούσας διατριβής συνοψίζονται ως εξής: α) μοντελοποίηση και περιορισμός των επιδράσεων του θορύβου σε εικόνες μικροσυστοιχιών κουκίδων κατά τέτοιο τρόπο ώστε να αυξηθεί η ακρίβεια των εξαγόμενων γονιδιακών εκφράσεων, β) μελέτη της επίδρασης του θορύβου και βελτιστοποίηση των μεθόδων ανάλυσης των γονιδιακών εκφράσεων με σκοπό τη διευκόλυνση των βιολόγων στην εξαγωγής βιολογικών συμπερασμάτων και την καλύτερη κατανόηση της βιολογικής διεργασίας που μελετάται, γ) εισαγωγή ενός ημιεποπτευόμενου (semi-supervised) κριτηρίου που στηριζόμενο σε βιολογικές πληροφορίες θα αποσκοπεί στην ανεύρεση βιολογικά σημαντικών ομάδων γονιδίων τα οποία ταυτόχρονα θα απαντούν σε συγκεκριμένα βιολογικά ερωτήματα ,δ) μελέτη της επίδρασης και της απόδοσης διαφόρων τεχνικών κατάτμησης εικόνων μικροσυστοιχιών κουκίδων, τόσο ανωτάτου επιπέδου (state-of-art) όσο και νέων, στην ποσοτικοποίηση γονιδιακών εκφράσεων. Για την πραγματοποίηση των παραπάνω στόχων σχεδιάστηκε και κατασκευάστηκε μια πλήρως δομημένη μεθοδολογία (a complete and robust framework) που περιελάμβανε αλγοριθμους επεξεργασίας και ανάλυσης εικόνας κουκίδων μικροσυστοιχιών Η προτεινόμενη μεθοδολογία ενσωμάτωσε στην ήδη υπάρχουσα αλγοριθμική μεθοδολογία (microarray analysis pipeline) έναν πρωτότυπο συνδυασμό τεχνικών επεξεργασίας και ανάλυσης εικόνας βασισμένο στην εις βάθος ποσοτική έρευνα της επίδρασης του θορύβου στην κατάτμηση κουκίδων (spot segmentation), στην εξαγωγή εντάσεων και στην εξόρυξη δεδομένων (data mining). Επιπρόσθετα, κατά την παρούσα διατριβή προτάθηκαν, κατασκευάστηκαν και αξιολογήθηκαν νέες τεχνικές κατάτμησης εικόνας από μικροσυστοιχές κουκίδων. Η χρησιμότητα των προτεινόμενων μεθοδολογιών αξιολογήθηκε τόσο σε εικονικές (simulated) όσο και σε πραγματικές εικόνες μικροσυστοιχιών κουκίδων.

A malária é uma doença infecciosa complexa, que resulta do “vírus” plasmodium, e manifesta-se sob cinco tipos distintos de espécies protozoários (plasmodium vivax, plasmodium ovale, plasmodium falciparum, plasmodium malariae e plasmodium Knowlesi), atacando sobretudo os glóbulos vermelhos. Considerada a quinta maior causa de morte por doenças infecciosas em todo o mundo após doenças respiratórias, VIH/SIDA, doenças diarreicas e tuberculose, no continente africano, a malária é considerada a segunda causa do aumento da mortalidade, após VIH/SIDA. No caso particular da Guiné-Bissau, esta constitui a principal causa do incremento da morbilidade e da mortalidade naquele país, onde, em 2012 foram notificados 129.684 casos de paludismo, dos quais 370 resultaram em óbitos. Partindo da realidade acima constatada, em particular, da complexidade e o impacto global da doença associada a uma forte mortalidade e morbilidade, concluiu-se ser necessário abordar esta temática, utilizando os SIG e a DR no sentido de determinar as regiões de elevado risco. Entendeu-se serem necessárias novas abordagens e novas ferramentas de análise dos dados epidemiológicos e consequentemente novas metodologias que possibilitem a determinação de áreas de risco por malária. O presente estudo, pretende demonstrar o papel dos SIG e DR na determinação das regiões de risco por malária. A metodologia utilizada centrou-se numa abordagem quantitativa baseada na hierarquização das variáveis. Pretende-se, assim abordar os impactos da malária e simultaneamente demonstrar as potencialidades dos SIG e das ferramentas de Análise Espacial no estudo da disseminação da mesma na Guiné-Bissau.