Дисертації з теми "Segmentation"

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2005.
Includes bibliographical references (p. 105-110).
This thesis describes the SANE (Segmentation According to Natural Examples) algorithm for learning to segment objects in static images from video data. SANE uses background subtraction to find the segmentation of moving objects in videos. This provides object segmentation information for each video frame. The collection of frames and segmentations forms a training set that SANE uses to learn the image and shape properties that correspond to the observed motion boundaries. Then, when presented with new static images, the model infers segmentations similar to the observed motion segmentations. SANE is a general method for learning environment-specific segmentation models. Because it is self-supervised, it can adapt to a new environment and new objects with relative ease. Comparisons of its output to a leading image segmentation algorithm demonstrate that motion-defined object segmentation is a distinct problem from traditional image segmentation. The model outperforms a trained local boundary detector because it leverages the shape information it learned from the training data.
by Michael Gregory Ross.
Ph.D.

The segmentation of medical images is a difficult task due to the inhomogeneous intensity variations that occurs during digital image acquisition, the complicated shape of the object, and the medical expert’s lack of semantic knowledge. Automated segmentation algorithms work well for some medical images, but no algorithm has been general enough to work for all medical images. In practice, most of the time the segmentation results are corrected by the experts before the actual use. In this work, we are motivated to determine how to make use of manually segmented data in automatic segmentation. The key idea is to transfer the ground truth segmentation from the database of train images to a given test image. The ground truth segmentation of MR images is done by experts. The process includes a hierarchical image decomposition approach that performs the shape matching of test images at several levels, starting with the image as a whole (i.e. level 0) and then going through a pyramid decomposition (i.e. level 1, level 2, etc.) with the database of the train images and the given test image. The goal of pyramid decomposition is to find the section of the training image that best matches a section of the test image of a different level. After that, a re-composition approach is taken to place the best matched sections of the training image to the original test image space. Finally, the ground truth segmentation is transferred from the best training images to their corresponding location in the test image. We have tested our method on a hip joint MR image database and the experiment shows successful results on level 0, level 1 and level 2 re-compositions. Results improve with deeper level decompositions, which supports our hypotheses.

In this thesis, a new algorithm has been proposed to segment the foreground of the fingerprint from the image under consideration. The algorithm uses three features, mean, variance and coherence. Based on these features, a rule system is built to help the algorithm to efficiently segment the image. In addition, the proposed algorithm combine split and merge with modified Otsu. Both enhancements techniques such as Gaussian filter and histogram equalization are applied to enhance and improve the quality of the image. Finally, a post processing technique is implemented to counter the undesirable effect in the segmented image. Fingerprint recognition system is one of the oldest recognition systems in biometrics techniques. Everyone have a unique and unchangeable fingerprint. Based on this uniqueness and distinctness, fingerprint identification has been used in many applications for a long period. A fingerprint image is a pattern which consists of two regions, foreground and background. The foreground contains all important information needed in the automatic fingerprint recognition systems. However, the background is a noisy region that contributes to the extraction of false minutiae in the system. To avoid the extraction of false minutiae, there are many steps which should be followed such as preprocessing and enhancement. One of these steps is the transformation of the fingerprint image from gray-scale image to black and white image. This transformation is called segmentation or binarization. The aim for fingerprint segmentation is to separate the foreground from the background. Due to the nature of fingerprint image, the segmentation becomes an important and challenging task. The proposed algorithm is applied on FVC2000 database. Manual examinations from human experts show that the proposed algorithm provides an efficient segmentation results. These improved results are demonstrating in diverse experiments.

The objective of this thesis is to detect high level semantic ideas to help to impose a structure on television talk shows. Indexing TV-shows is a subject that, to our knowledge, is rarely talked about in the scientific community.There is no common understanding of what this imposed structure should look like. We can say that the purpose is to organise the audiovisual content into sections that convey a specific information. It thus encompasses issues as diverse as scene segmentation, speech noise detection, speaker identification, etc. The basic problem of structuring is the gap between the information extracted from visual data flow and human interpretation made by the user of these data. Numerous studies have examined the organisation of highly structured video content. Thus, the state of the art has many studies on sport or newscast transmissions. Our goal is to detect key audiovisual events using a variety of descriptors and generic classifiers. We propose a generic approach that is able to assess all TV-show indexing problems. This enables an operator to use one single tool to infer a logical structure. Our approach can be considered as ``semi-automatic'' in the sense that the training data is collected on the fly by the operator who is asked to arbitrarily select one video excerpt of each concept involved. We have assessed a wide selection of audio and video features, used MKL as a feature selection algorithm and then built various content detectors and segmentors useful for imposing broad semantic classes on television data.This master's thesis was set forth by TELECOM ParisTech and was begun there March 1, 2010. This final report was submitted to TELECOM ParisTech, NTNU and Institute EURECOM August 29, 2010.

The problem of unsupervised texture segmentation was studied and a texture segmentation algorithm was developed making use of the minimum number of prior assumptions. In particular, no prior information about the type of textures, the number of textures and the appropriate scale of analysis for each texture was required. The texture image was analysed by the multiresolution Gabor expansion. The Gabor expansion generates a large number of features for each image and the most suitable feature space for segmentation needs to be determined automatically. The two-point correlation function was used to test the separability of the distributions in each feature space. A measure was developed to evaluate evidence of multiple clusters from the two-point correlation function, making it possible to determine the most suitable feature space for clustering. Thus, at a given resolution level, the most appropriate feature space was selected and used to segment the image. Due to inherent ambiguities and limitations of the two-point correlation function, this feature space exploration and segmentation was performed several times at the same resolution level until no further evidence of multiple clusters was found, at which point, the process was repeated at the next finer resolution level. In this way, the image was progressively segmented, proceeding from coarse to fine Gabor resolution levels without any knowledge of the actual number of textures present. In order to refine the region-labelled image obtained at the end of the segmentation process, two postprocessing pixel-level algorithms were developed and implemented. The first was the mixed pixel classification algorithm which is based on the analysis of the effect of the averaging window at the boundary between two regions and re-assigns the pixel labels to improve the boundary localisation. Multiresolution probabilistic relaxation is the second postprocessing algorithm which we developed. This algorithm incorporates contextual evidence to relabel pixels close to the boundary in order to smooth it and improve its localisation. The results obtained were quantified by known error measures, as well as by new error measures which we developed. The quantified results were compared to similar results by other authors and show that our unsupervised algorithm performs as well as other methods which assume prior information.

Image segmentation is an important area in the general field of image processing and computer vision. It is a fundamental part of the 'low level' aspects of computer vision and has many practical applications such as in medical imaging, industrial automation and satellite imagery. Traditional methods for image segmentation have approached the problem either from localisation in class space using region information, or from localisation in position, using edge or boundary information. More recently, however, attempts have been made to combine both region and boundary information in order to overcome the inherent limitations of using either approach alone. In this thesis, a new approach to image segmentation is presented that integrates region and boundary information within a multiresolution framework. The role of uncertainty is described, which imposes a limit on the simultaneous localisation in both class and position space. It is shown how a multiresolution approach allows the trade-off between position and class resolution and ensures both robustness in noise and efficiency of computation. The segmentation is based on an image model derived from a general class of multiresolution signal models, which incorporates both region and boundary features. A four stage algorithm is described consisting of: generation of a low-pass pyramid, separate region and boundary estimation processes and an integration strategy. Both the region and boundary processes consist of scale-selection, creation of adjacency graphs, and iterative estimation within a general framework of maximum a posteriori (MAP) estimation and decision theory. Parameter estimation is performed in situ, and the decision processes are both flexible and spatially local, thus avoiding assumptions about global homogeneity or size and number of regions which characterise some of the earlier algorithms. A method for robust estimation of edge orientation and position is described which addresses the problem in the form of a multiresolution minimum mean square error (MMSE) estimation. The method effectively uses the spatial consistency of output of small kernel gradient operators from different scales to produce more reliable edge position and orientation and is effective at extracting boundary orientations from data with low signal-to-noise ratios. Segmentation results are presented for a number of synthetic and natural images which show the cooperative method to give accurate segmentations at low signal-to-noise ratios (0 dB) and to be more effective than previous methods at capturing complex region shapes.

This thesis investigates the evaluation of automatic and manual text segmentation. Text segmentation is the process of placing boundaries within text to create segments according to some task-dependent criterion. An example of text segmentation is topical segmentation, which aims to segment a text according to the subjective definition of what constitutes a topic. A number of automatic segmenters have been created to perform this task, and the question that this thesis answers is how to select the best automatic segmenter for such a task. This requires choosing an appropriate segmentation evaluation metric, confirming the reliability of a manual solution, and then finally employing an evaluation methodology that can select the automatic segmenter that best approximates human performance. A variety of comparison methods and metrics exist for comparing segmentations (e.g., WindowDiff, Pk), and all save a few are able to award partial credit for nearly missing a boundary. Those comparison methods that can award partial credit unfortunately lack consistency, symmetricity, intuition, and a host of other desirable qualities. This work proposes a new comparison method named boundary similarity (B) which is based upon a new minimal boundary edit distance to compare two segmentations. Near misses are frequent, even among manual segmenters (as is exemplified by the low inter-coder agreement reported by many segmentation studies). This work adapts some inter-coder agreement coefficients to award partial credit for near misses using the new metric proposed herein, B. The methodologies employed by many works introducing automatic segmenters evaluate them simply in terms of a comparison of their output to one manual segmentation of a text, and often only by presenting nothing other than a series of mean performance values (along with no standard deviation, standard error, or little if any statistical hypothesis testing). This work asserts that one segmentation of a text cannot constitute a “true” segmentation; specifically, one manual segmentation is simply one sample of the population of all possible segmentations of a text and of that subset of desirable segmentations. This work further asserts that an adapted inter-coder agreement statistics proposed herein should be used to determine the reproducibility and reliability of a coding scheme and set of manual codings, and then statistical hypothesis testing using the specific comparison methods and methodologies demonstrated herein should be used to select the best automatic segmenter. This work proposes new segmentation evaluation metrics, adapted inter-coder agreement coefficients, and methodologies. Most important, this work experimentally compares the state-or-the-art comparison methods to those proposed herein upon artificial data that simulates a variety of scenarios and chooses the best one (B). The ability of adapted inter-coder agreement coefficients, based upon B, to discern between various levels of agreement in artificial and natural data sets is then demonstrated. Finally, a contextual evaluation of three automatic segmenters is performed using the state-of-the art comparison methods and B using the methodology proposed herein to demonstrate the benefits and versatility of B as opposed to its counterparts.

Congenital heart disease affects just under one percentage of all live births [1]. Those defects that manifest themselves as changes to the cardiac chamber volumes are the motivation for the research presented in this thesis. Blood volume measurements in vivo require delineation of the cardiac chambers and manual tracing of foetal cardiac chambers is very time consuming and operator dependent. This thesis presents a multi region based level set snake deformable model applied in both 2D and 3D which can automatically adapt to some extent towards ultrasound noise such as attenuation, speckle and partial occlusion artefacts. The algorithm presented is named Mumford Shah Sarti Collision Detection (MSSCD). The level set methods presented in this thesis have an optional shape prior term for constraining the segmentation by a template registered to the image in the presence of shadowing and heavy noise. When applied to real data in the absence of the template the MSSCD algorithm is initialised from seed primitives placed at the centre of each cardiac chamber. The voxel statistics inside the chamber is determined before evolution. The MSSCD stops at open boundaries between two chambers as the two approaching level set fronts meet. This has significance when determining volumes for all cardiac compartments since cardiac indices assume that each chamber is treated in isolation. Comparison of the segmentation results from the implemented snakes including a previous level set method in the foetal cardiac literature show that in both 2D and 3D on both real and synthetic data, the MSSCD formulation is better suited to these types of data. All the algorithms tested in this thesis are within 2mm error to manually traced segmentation of the foetal cardiac datasets. This corresponds to less than 10% of the length of a foetal heart. In addition to comparison with manual tracings all the amorphous deformable model segmentations in this thesis are validated using a physical phantom. The volume estimation of the phantom by the MSSCD segmentation is to within 13% of the physically determined volume.

Thesis (M.S.)--West Virginia University, 2004.
Title from document title page. Document formatted into pages; contains vi, 52 p. : ill. (some col.). Includes abstract. Includes bibliographical references (p. 50-52).

Systeme der Computer Vision spielen in der Automatisierung vieler Prozesse eine wichtige Rolle. Die wichtigste Aufgabe solcher Systeme ist die Automatisierung des visuellen Erkennungsprozesses und die Extraktion der relevanten Information aus Bildern oder Bildsequenzen. Eine wichtige Komponente dieser Systeme ist die Bildsegmentierung, denn sie bestimmt zu einem großen Teil die Qualitaet des Gesamtsystems. Fuer die Segmentierung von Bildern und Bildsequenzen werden neue Algorithmen vorgeschlagen. Das Konzept der Multiresolution wird als eigenstaendig dargestellt, es existiert unabhaengig von der Wavelet-Transformation. Die Wavelet-Transformation wird zur Verarbeitung von Bildern und Bildsequenzen zu einer 2D- bzw. 3D-Wavelet- Transformation erweitert. Fuer die Segmentierung von Bildern wird der Algorithmus Resolution Mosaic Expectation Maximization (RM-EM) vorgeschlagen. Das Ergebnis der Vorverarbeitung sind unterschiedlich aufgeloesten Teilbilder, das Aufloesungsmosaik. Durch dieses Mosaik lassen sich raeumliche Korrelationen zwischen den Pixeln ausnutzen. Die Verwendung unterschiedlicher Aufloesungen beschleunigt die Verarbeitung und verbessert die Ergebnisse. Fuer die Extraktion von bewegten Objekten aus Bildsequenzen werden neue Algorithmen vorgeschlagen, die auf der 3D-Wavelet-Transformation und auf der Analyse mit 3D-Wavelet-Packets beruhen. Die neuen Algorithmen haben den Vorteil, dass sie sowohl die raeumlichen als auch die zeitlichen Bewegungsinformationen beruecksichtigen. Wegen der geringen Berechnungskomplexitaet der Wavelet-Transformation ist fuer den ersten Segmentierungsschritt Hardware auf der Basis von FPGA entworfen worden. Aktuelle Anwendungen werden genutzt, um die Algorithmen zu evaluieren: die Segmentierung von Magnetresonanzbildern des menschlichen Gehirns und die Detektion von bewegten Objekten in Bildsequenzen von Verkehrsszenen. Die neuen Algorithmen sind robust und fuehren zu besseren Segmentierungsergebnissen.
More and more computer vision systems take part in the automation of various applications. The main task of such systems is to automate the process of visual recognition and to extract relevant information from the images or image sequences acquired or produced by such applications. One essential and critical component in almost every computer vision system is image segmentation. The quality of the segmentation determines to a great extent the quality of the final results of the vision system. New algorithms for image and video segmentation based on the multiresolution analysis and the wavelet transform are proposed. The concept of multiresolution is explained as existing independently of the wavelet transform. The wavelet transform is extended to two and three dimensions to allow image and video processing. For still image segmentation the Resolution Mosaic Expectation Maximization (RM-EM) algorithm is proposed. The resolution mosaic enables the algorithm to employ the spatial correlation between the pixels. The level of the local resolution depends on the information content of the individual parts of the image. The use of various resolutions speeds up the processing and improves the results. New algorithms based on the 3D wavelet transform and the 3D wavelet packet analysis are proposed for extracting moving objects from image sequences. The new algorithms have the advantage of considering the relevant spatial as well as temporal information of the movement. Because of the low computational complexity of the wavelet transform an FPGA hardware for the primary segmentation step was designed. Actual applications are used to investigate and evaluate all algorithms: the segmentation of magnetic resonance images of the human brain and the detection of moving objects in image sequences of traffic scenes. The new algorithms show robustness against noise and changing ambient conditions and gave better segmentation results.

Spotting patterns of interest in an input signal is a very useful task in many different fields including medicine, bioinformatics, economics, speech recognition and computer vision. Example instances of this problem include spotting an object of interest in an image (e.g., a tumor), a pattern of interest in a time-varying signal (e.g., audio analysis), or an object of interest moving in a specific way (e.g., a human's body gesture). Traditional spotting methods, which are based on Dynamic Time Warping or hidden Markov models, use some variant of dynamic programming to register the pattern and the input while accounting for temporal variation between them. At the same time, those methods often suffer from several shortcomings: they may give meaningless solutions when input observations are unreliable or ambiguous, they require a high complexity search across the whole input signal, and they may give incorrect solutions if some patterns appear as smaller parts within other patterns. In this thesis, we develop a framework that addresses these three problems, and evaluate the framework's performance in spotting and recognizing hand gestures in video. The first contribution is a spatiotemporal matching algorithm that extends the dynamic programming formulation to accommodate multiple candidate hand detections in every video frame. The algorithm finds the best alignment between the gesture model and the input, and simultaneously locates the best candidate hand detection in every frame. This allows for a gesture to be recognized even when the hand location is highly ambiguous. The second contribution is a pruning method that uses model-specific classifiers to reject dynamic programming hypotheses with a poor match between the input and model. Pruning improves the efficiency of the spatiotemporal matching algorithm, and in some cases may improve the recognition accuracy. The pruning classifiers are learned from training data, and cross-validation is used to reduce the chance of overpruning. The third contribution is a subgesture reasoning process that models the fact that some gesture models can falsely match parts of other, longer gestures. By integrating subgesture reasoning the spotting algorithm can avoid the premature detection of a subgesture when the longer gesture is actually being performed. Subgesture relations between pairs of gestures are automatically learned from training data. The performance of the approach is evaluated on two challenging video datasets: hand-signed digits gestured by users wearing short sleeved shirts, in front of a cluttered background, and American Sign Language (ASL) utterances gestured by ASL native signers. The experiments demonstrate that the proposed method is more accurate and efficient than competing approaches. The proposed approach can be generally applied to alignment or search problems with multiple input observations, that use dynamic programming to find a solution.

3D polygonal models are increasingly being deployed in a wide range of Virtual Reality (VR) applications, including scientific visualization, collaborative design, entertainment, e-commerce, and remote education, etc. These models, produced by the 3D laser scanning systems, are often represented as complex polyhedral meshes with hundreds of thousands or millions of vertices and triangles. Although this representation can achieve high level of realism, these models usually demand a huge amount of storage space and/or transmission bandwidth in the raw data format. Also, a polygonal mesh does not capture a high-level structure, which is useful for managing data in applications, such as object registration, object retrieval and indexing, feature modeling, etc. One way to impose such a high-level description is through mesh segmentation. Therefore, mesh simplification, segmentation and compression have recently been the main areas in 3D mesh processing. In this dissertation, we present an efficient 3D mesh multi-resolution modeling algorithm, which can output arbitrary resolutions of an input model. This algorithm considers edge curvature and neighborhood face area change as the error metrics for the edge collapse operation. Compared with most of the existing simplification algorithms, the proposed method is simple and effective. We also present an efficient and robust neighborhood based algorithm for 3D mesh segmentation. This approach uses discrete Gaussian curvature and concaveness estimation to detect the boundary vertices between the distinct regions of a given mesh. To capture more accurate and relevant geometric information of a vertex on the mesh surface, we enlarge the 1-ring neighborhood to an eXtended Multiple-Ring (XMR) neighborhood. After feature detection, a fast marching watershed method is deployed, followed by an efficient region merging scheme. Simulation results show that this algorithm is efficient and robust to high-resolution models. Finally, we propose a segmentation based mesh compression scheme. Most of the existing 3D mesh coding algorithms compress the model as a whole. Our algorithm separately compresses the partitioned regions that are obtained by the segmentation method described above. The compressed data are put into one stream part by part. Each part is independent of the others. The boundary strips between the different regions are also encoded and appended to the end of the stream. In an inactive or selective application, the users can get the interested parts or the whole object after all the parts and the boundary strips that can connect all the parts together have been received.

Thèse doctorat : Images et Systèmes : Villeurbanne, INSA : 2003.
Thèse rédigée en anglais. Résumé en français en début de chaque chapitre. Titre provenant de l'écran-titre. Bibliogr. p. 216-232. Publications de l'auteur p. 214-215.

Ce travail traite de la segmentation d'images médicales: ce type de problème est difficile à résoudre à cause de la piètre qualité du signal, et de la fréquente nécessité d'avoir une connaissance spécifique. Aussi se restreint-on à des objets de type volumique courants, tels que les os ou des organes massifs. On modélise une surface fermée à l'aide de produits tensoriels de B-splines (utilisés en CAO), et on propose deux formes de représentation: soit un cylindre à pôles, défini par une seule grille de contrôle, soit un graphe de contrôle, reliant continûment plusieurs grilles. Ces représentations sont adaptées aux déformations globales ou locales, et aux changements de topologie. La déformation d'une telle surface est modélisée par le dégonflage d'un ballon: partant d'une forme englobant la scène, on laisse se dégonfler la surface pour atteindre successivement l'enveloppe convexe, puis l'enveloppe localement convexe, et enfin la surface externe de la scène. L'apport principal de cette méthode consiste en la possibilité d'appréhender un objet complexe, par sa topologie, ou par le nombre de ses composantes connexes. La méthode de segmentation utilise alors cette technique sur l'ensemble des indices de contours détectés dans les images. On peut voir cette méthode comme un outil modélisant l'interpolation naturelle de l'œil. Ce travail se place ainsi parmi les opérateurs de niveau intermédiaire, en apportant la mise en cohérence de certaines propriétés géométriques de la fonction densité associée aux images. Ceci ne résoud cependant pas le problème de connaissance a priori, une interface utilisateur pouvant y remédier dans les cas critiques

Motion segmentation is the process of dividing video frames into regions which have different motions, providing a cut-out of the moving objects. Such a segmentation is a necessary first stage in many video analysis applications, but providing an accurate, efficient motion segmentation still presents a challenge. This dissertation proposes a novel approach to motion segmentation, using the image edges in a frame. Using edges, a motion can be calculated for each object. Edges provide good motion information, and it is shown that a set of edges, labelled according to the object motion that they obey, is sufficient to completely determine the labelling of the whole frame, up to unresolvable ambiguities. The areas of the frame between edges are divided into regions, grouping together pixels of similar colour, and these regions can each be assigned to different motion layers by reference to the edges. The depth ordering of these layers can also be deduced. A Bayesian framework is presented, which determines the most likely region labelling and depth ordering, given edges labelled with their probability of obeying each of the object motions. An efficient implementation of this framework is presented, initially for segmenting two motions (foreground and background) using two frames. The ExpectationMaximisation algorithm is used to determine the two motions and calculate the label probability for each edge. The frame is then segmented into regions. The best motion labelling for these regions is determined using simulated annealing. Extensions of this simple implementation are then presented. It is demonstrated how, by tracking the edges into further frames, the statistics may be accumulated to provide an even more accurate and robust segmentation. This also allows a complete sequence to be segmented. It is then demonstrated that the framework can be extended to a larger number of motions. A new hierarchical method of initialising the Expectation-Maximisation algorithm is described, which also determines the best number of motions. These techniques have been extensively tested on thirty-four real sequences, covering a wide range of genres. The results demonstrate that the proposed edge-based approach is an accurate and efficient method of obtaining a motion segmentation.

This study is dedicated to the problem of segmenting monochrome images into distinct homogeneous regions by texture properties. The principle of the approaches to texture segmentation adopted in this thesis is mapping a textured image into a grey level image so that conventional segmentation techniques by intensity can be applied. Three novel approaches to texture segmentation have been developed in this thesis. They are called the Local Feature Statistics Approach (LFS), the Local Spectral Mapping Approach (LSM) and the Multichannel Spatial Filtering Approach (MSF). In the LFS approach, a multiresolution scheme for extracting texture features is introduced. This scheme produces features which can describe texture characteristics at different resolution levels. The gradient vector at each resolution level is used as the local texture feature. Based on the population statistics of gradient magnitude and direction in a local observation window, two novel texture measures, named as the Linear Gradient Magnitude Enhancement Measure (LGME) and the Linear Gradient Direction Enhancement Measure (LGDE), are developed to enhance different texture characteristics. In the LSM approach, the new scheme for the extraction of local texture features is based on performing transformations on the power spectra of local regions. The power spectrum of a local region is divided into a number of rings or wedges, and local spectral vectors are formed by summing the energy in these rings or wedges as vector elements. Two new texture measures, named as the Linear Radial Feature Enhancement Measure (LRFE) and the Linear Angular Feature Enhancement Measure (LAFE), are developed to highlight different texture characteristics. The MSF approach is based on the Multichannel Spatial Filtering Model (MSFM) for the human visual cortex. It is assumed in this approach that a texture can be characterised by its principal spatial frequency components, and that these components can be captured by a number of narrowband spatial filters. A new class of filters, called the Gaussian-Smoothed Fan (GSF) filters, is developed to perform channel filtering operations. The passband characteristic of these GSF filters is flatter than that of the Gabor filters, thus their bandwidths are inherently better defined. Computational algorithms based on these three new approaches are implemented and applied to a set of textured images. Good segmentation results are obtained, with more than 92% of the pixel population of each of the test images (derived from Brodatzs texture album) being correctly classified by all the three approaches. By comparison, the newly-developed GSF filters used in the MSF approach have an important advantage over the Gabor filters in that they can produce better defined boundaries between texture regions.

L'utilisation de systèmes multi-caméras est de plus en plus populaire et il y a un intérêt croissant à résoudre les problèmes de vision par ordinateur dans ce contexte particulier. L'objectif étant de ne pas se limiter à l'application des méthodes monoculaires mais de proposer de nouvelles approches intrinsèquement orientées vers les systèmes multi-caméras. Le travail de cette thèse a pour objectif une meilleure compréhension du problème de segmentation multi-vues, pour proposer une nouvelle approche qui tire meilleur parti de la redondance d'information inhérente à l'utilisation de plusieurs points de vue. La segmentation multi-vues est l'identification de l'objet observé simultanément dans plusieurs caméras et sa séparation de l'arrière-plan. Les approches monoculaires classiques raisonnent sur chaque image de manière indépendante et ne bénéficient pas de la présence de plusieurs points de vue. Une question clé de la segmentation multi-vues réside dans la propagation d'information sur la segmentation entres les images tout en minimisant la complexité et le coût en calcul. Dans ce travail, nous investiguons en premier lieu l'utilisation d'un ensemble épars d'échantillons de points 3D. L'algorithme proposé classe chaque point comme "vide" s'il se projette sur une région du fond et "occupé" s'il se projette sur une région avant-plan dans toutes les vues. Un modèle probabiliste est proposé pour estimer les modèles de couleur de l'avant-plan et de l'arrière-plan, que nous testons sur plusieurs jeux de données de l'état de l'art. Deux extensions du modèle sont proposées. Dans la première, nous montrons la flexibilité de la méthode proposée en intégrant les mélanges de Gaussiennes comme modèles d'apparence. Cette intégration est possible grâce à l'utilisation de l'inférence variationelle. Dans la seconde, nous montrons que le modèle bayésien basé sur les échantillons 3D peut aussi être utilisé si des mesures de profondeur sont présentes. Les résultats de l'évaluation montrent que les problèmes de robustesse, typiquement causés par les ambigüités couleurs entre fond et forme, peuvent être au moins partiellement résolus en utilisant cette information de profondeur. A noter aussi qu'une approche multi-vues reste meilleure qu'une méthode monoculaire utilisant l'information de profondeur. Les différents tests montrent aussi les limitations de la méthode basée sur un échantillonnage éparse. Cela a montré la nécessité de proposer un modèle reposant sur une description plus riche de l'apparence dans les images, en particulier en utilisant les superpixels. L'une des contributions de ce travail est une meilleure modélisation des contraintes grâce à un schéma par coupure de graphes liant les régions d'images aux échantillons 3D. Dans le cas statique, les résultats obtenus rivalisent avec ceux de l'état de l'art mais sont obtenus avec beaucoup moins de points de vue. Les résultats dans le cas dynamique montrent l'intérêt de la propagation de l'information de segmentation à travers la géométrie et le mouvement. Enfin, la dernière partie de cette thèse explore la possibilité d'améliorer le suivi dans les systèmes multi-caméras non calibrés. Un état de l'art sur le suivi monoculaire et multi-caméras est présenté et nous explorons l'utilisation des matrices d'autosimilarité comme moyen de décrire le mouvement et de le comparer entre plusieurs caméras
There has been a growing interest for multi-camera systems and many interesting works have tried to tackle computer vision problems in this particular configuration. The general objective is to propose new multi-view oriented methods instead of applying limited monocular approaches independently for each viewpoint. The work in this thesis is an attempt to have a better understanding of the multi-view object segmentation problem and to propose an alternative approach making maximum use of the available information from different viewpoints. Multiple view segmentation consists in segmenting objects simultaneously in several views. Classic monocular segmentation approaches reason on a single image and do not benefit from the presence of several viewpoints. A key issue in that respect is to ensure propagation of segmentation information between views while minimizing complexity and computational cost. In this work, we first investigate the idea that examining measurements at the projections of a sparse set of 3D points is sufficient to achieve this goal. The proposed algorithm softly assigns each of these 3D samples to the scene background if it projects on the background region in at least one view, or to the foreground if it projects on foreground region in all views. A complete probabilistic framework is proposed to estimate foreground/background color models and the method is tested on various datasets from state of the art. Two different extensions of the sparse 3D sampling segmentation framework are proposed in two scenarios. In the first, we show the flexibility of the sparse sampling framework, by using variational inference to integrate Gaussian mixture models as appearance models. In the second scenario, we propose a study of how to incorporate depth measurements in multi-view segmentation. We present a quantitative evaluation, showing that typical color-based segmentation robustness issues due to color-space ambiguity between foreground and background, can be at least partially mitigated by using depth, and that multi-view color depth segmentation also improves over monocular color depth segmentation strategies. The various tests also showed the limitations of the proposed 3D sparse sampling approach which was the motivation to propose a new method based on a richer description of image regions using superpixels. This model, that expresses more subtle relationships of the problem trough a graph construction linking superpixels and 3D samples, is one of the contributions of this work. In this new framework, time related information is also integrated. With static views, results compete with state of the art methods but they are achieved with significantly fewer viewpoints. Results on videos demonstrate the benefit of segmentation propagation through geometric and temporal cues. Finally, the last part of the thesis explores the possibilities of tracking in uncalibrated multi-view scenarios. A summary of existing methods in this field is presented, in both mono-camera and multi-camera scenarios. We investigate the potential of using self-similarity matrices to describe and compare motion in the context of multi-view tracking

This thesis investigates the segmentation of data in 2D and 3D by texture analysis using Fourier domain filtering. The field of texture analysis is a well-trodden one in 2D, but many applications, such as Medical Imaging, Stratigraphy or Crystallography, would benefit from 3D analysis instead of the traditional, slice-by-slice approach. With the intention of contributing to texture analysis and segmentation in 3D, a multiresolution volumetric texture segmentation (M-VTS) algorithm is presented. The method extracts textural measurements from the Fourier domain of the data via sub-band filtering using a Second Orientation Pyramid. A novel Bhattacharyya space, based on the Bhattacharyya distance is proposed for selecting of the most discriminant measurements and produces a compact feature space. Each dimension of the feature space is used to form a Quad Tree. At the highest level of the tree, new positional features are added to improve the contiguity of the classification. The classified space is then projected to lower levels of the tree where a boundary refinement procedure is performed with a 3D equivalent of butterfly filters. The performance of M-VTS is tested in 2D by classifying a set of standard texture images. The figures contain different textures that are visually stationary. M-VTS yields lower misclassification rates than reported elsewhere ([104, 111, 124]). The algorithm was tested in 3D with artificial isotropic data and three Magnetic Resonance Imaging sets of human knees with satisfactory results. The regions segmented from the knees correspond to anatomical structures that could be used as a starting point for other measurements. By way of example, we demonstrate successful cartilage extraction using our approach.

Recent advances in technology have facilitated the prospect of automatic cardiac auscultation by using digital stethoscopes. This in turn creates the need for development of algorithms capable of automatic segmentation of the heart sound. Pediatric heart sound segmentation is a challenging task due to various factors including the significant influence of respiration on the heart sound. This project studies the application of homomorphic filtering and Hidden Markov Model for the purpose of pediatric heart sound segmentation. The efficacy of the proposed method is evaluated on a publicly available dataset and its performance is compared with those of three other existing methods. The results show that our proposed method achieves accuracy of 92.4% ±1.1% and 93.5% ±1.1% in identification of first and second heart sound components, and is superior to four other existing methods in term of accuracy or time complexity.

Visual perception of images is closely related to the recognition of the different texture areas within an image. Identifying the boundaries of these regions is an important step in image analysis and image understanding. This thesis presents supervised and unsupervised methods which allow an efficient segmentation of the texture regions within multi-texture images. The features used by the methods are based on a measure of the fractal dimension of surfaces in several directions, which allows the transformation of the image into a set of feature images, however no direct measurement of the fractal dimension is made. Using this set of features, supervised and unsupervised, statistical processing schemes are presented which produce low classification error rates. Natural texture images are examined with particular application to the analysis of sonar images of the seabed. A number of processes based on fractal models for texture synthesis are also presented. These are used to produce realistic images of natural textures, again with particular reference to sonar images of the seabed, and which show the importance of phase and directionality in our perception of texture. A further extension is shown to give possible uses for image coding and object identification.

Con lo scopo di risparmiare capitale e incrementare i profitti tramite attività di marketing sempre più mirate, conoscere le preferenze di un cliente e supportarlo nell’acquisto, sta passando dall’essere una scelta all’essere una necessità. A tal proposito, le aziende si stanno muovendo verso un approccio sempre più automatizzato per riuscire a classificare la clientela, cos`ı da ottimizzare sempre più l’esperienza d’acquisto. Tramite il Machine Learning è possibile effettuare svariati tipi di analisi che consentano di raggiungere questo scopo. L’obiettivo di questo progetto è, in prima fase, di dare una panoramica al lettore su quali siano le tecniche e gli strumenti che mette a disposizione il ML. In un secondo momento verrà descritto il problema della Customer Segmentation e quali tecniche e benefici porta con sé questo tema. Per finire, verranno descritte le varie fasi su cui si fonda il seguente progetto di ML rivolto alla classificazione della clientela, basandosi sul totale di spesa effettuata in termini monetari e la quantità di articoli acquistati.

Cette these est consacree a la segmentation statistique floue d'images. Dans le cadre de la segmentation bayesienne, une image est modelisee par deux champs aleatoires: le champ des classes representant la verite-terrain inobservable et que l'on cherche a estimer et le champ des observations. Le champ des classes est generalement suppose prendre ses valeurs dans un ensemble fini des classes thematiques. Le modele propose dans cette these vise a enrichir ce modele classique: chaque site de la verite-terrain peut soit appartenir a l'une des classes thematiques, soit resulter d'un melange de ces classes. Dans ce dernier cas, la nature du site est decrite par le vecteur des coefficients d'appartenance du site a chacune des classes. La modelisation statistique s'appuie sur une mesure contenant des masses de dirac chargeant les classes thematiques et la mesure de lebesgue par rapport a laquelle est definie la densite des vecteurs d'appartenance dans le cas d'un site melange. Une premiere partie est consacree a la segmentation proprement dite. D'une part, plusieurs methodes bayesiennes sont proposees, deduites de differentes fonctions de perte. De plus, deux methodes de segmentation est developpee dans trois versions, correspondant a differentes maniere d'exploiter l'information spatiale: la segmentation aveugle, contextuelle et adaptative. Afin de proposer des algorithmes de segmentation non supervisee, trois procedures d'estimation des parametres necessaires a la mise en uvre de la segmentation sont adaptees a ce modele: expectation- maximisation, stochastic estimation maximisation et iterative conditional estimation. Enfin, les methodes non supervisees de segmentation statistiques floues proposees sont testees sur des images reelles et de simulation

Includes bibliographical references (p. 67-72).
Texture is an important visual cue in both human and computer vision. Segmenting images into regions of constant texture is used in many applications. This work surveys a wide range of texture descriptors and segmentation methods to determine the state of the art in texture segmentation. Two types of texture descriptors are investigated: filter bank based methods and local descriptors. Filter banks deconstruct an image into several bands, each of which emphasises areas of the image with different properties. Textons are an adaptive histogram method which describes the distribution of typical feature vectors. Local descriptors calculate features from smaller neighbourhoods than filter banks. Some local descriptors calculate a scale for their local neighbourhood to achieve scale invariance. Both local and global segmentation methods are investigated. Local segmentation methods consider each pixel in isolation. Global segmentation methods penalise jagged borders or fragmented regions in the segmentation. Pixel labelling and border detection methods are investigated. Methods for measuring the accuracy of segmentation are discussed. Two data sets are used to test the texture segmentation algorithms. The Brodatz Album mosaics are composed of grayscale texture images from the Brodatz Album. The Berkeley Natural Images data set has 300 colour images of natural scenes. The tests show that, of the descriptors tested, filter bank based textons are the best texture descriptors for grayscale images. Local image patch textons are best for colour images. Graph cut segmentation is best for pixel labelling problems and edge detection with regular borders. Non-maxima suppression is best for edge detection with irregular borders. Factors affecting the performance of the algorithms are investigated.

Cette thèse est consacrée a la segmentation statistique non supervisée des images radar. L'originalité de notre étude réside dans l'adoption du système de distributions de Pearson pour modéliser les scènes naturelles homogènes. Nous généralisons les méthodes d'estimation employées pour un mélange de distributions de même loi a un mélange de distributions de lois différentes et appartenant au système de Pearson. Des algorithmes d'estimation bases sur la maximisation de la vraisemblance et sur la notion de l’Esperance conditionnelle ont été développes afin d'accomplir cette tâche. Les méthodes d'estimation développées pour les images radar mono spectrales ont été généralisées pour le cas des images radar multi spectrales. Nous étudions la robustesse des algorithmes proposes à travers la segmentation d'images de synthèse et d'images radar de natures différentes.

Dans cette these, nous nous sommes interesse a plusieurs problemes lies a l'imagerie discrete 3d d'une part, et a la parallelisation d'algorithmes de segmentation d'images d'autres part. Dans une premiere partie, divers problemes d'imagerie discrete 3d sont presentes. L'approche des complexes cellulaires dont l'objectif est d'offrir une modelisation coherente en imagerie volumique discrete est introduite. Une notion de surface discrete composee d'elements bidimensionnels de l'espace 3d est rappele. Plusieurs resultats theoriques sur ce sujet sont rassembles dans un rapport de recherche fourni en annexe. Les principaux resultats personnels sont repris dans la these: une extension aux objets 26-connexes de la definition des surfaces discretes qui verifie la propriete fondamentale qu'il existe une surface connectee entre chaque pair de composante objet, composante du fond. Enfin, on propose egalement un algorithme d'extraction de surfaces sur pram. Dans une deuxieme partie, une bibliographie sur le sujet de la segmentation d'image est presentee. Des algorithmes de segmentation d'image par fusion de regions sont introduits, et on compare des algorithmes a caracteres intrinsequement sequentiels et paralleles existant. Dans ce cadre la, une etude de la parallelisation de l'etiquetage en composantes connexes des images binaires 3d est presentee. Une bibliographie du sujet decris les principaux algorithmes sequentiels et les diverses approches paralleles existantes. Nous introduisons notre contribution personnelle qui est un algorithme parallele sur hypercule ipsc860. La derniere partie de la these presente divers outils qui ont ete developpes pour la parallelisation du traitement d'images volumiques: la librairie ppcm qui definie un standard de communication pour des machines paralleles mimd a memoire distribuee et une structure de donnee qui inclue des fonctions d'equilibrage de charges

Cette thèse concerne la segmentation des données discrètes 3D, appelées images de profondeur, fournies par des capteurs de profondeur actifs. Ces informations tridimensionnelles sont une représentation point par point des surfaces observées par ces capteurs de vision. Les différents méthodes et capteurs permettant l'acquisition d'informations tridimensionnelles sont décrits. Le problème de la segmentation d'images de profondeur est défini, puis les principales techniques de segmentation sont présentées. Une analyse de l'état de l'art dans le domaine nous amène à justifier la direction de nos travaux, en particulier à dissocier la segmentation d'images de profondeur en deux problèmes selon le type de scènes analysées, constituées d'objets uniquement polyedriques ou non. Une méthode de segmentation d'images de profondeur en surfaces planes est tout d'abord proposée, puis elle est étendue à la segmentation en régions planes et courbes. Les originalités de la méthode reposent sur le procédé d'estimation des attributs différentiels et sur le choix des germes de la croissance de régions. Les régions extraites sont représentées par un graphe d'adjacence et forment la base d'une pyramide de graphes. Les données initiales sont filtrées. Le choix du filtre réducteur de bruit est realisé lors d'une étude comparative de plusieurs opérateurs existants. L'étape suivante consiste à analyser les différents profils 1D de l'image, constitués par ses lignes et ses colonnes, afin d'estimer les derivées directionnelles du premier et du second ordre. La normale et la courbure moyenne à la surface en chaque pixel sont ensuite déduites, tout en tenant compte des discontinuités de profondeur. Les pixels sont alors regroupés en régions homogènes au sens de ces attributs. A partir du graphe d'adjacence des régions ainsi obtenues, une stratégie pyramidale de fusion, parallélisable, est mise en oeuvre pour aboutir au résultat de la segmentation. Les performances de l'algorithme proposé sont caracterisées. Les résultats, obtenus à partir d'un jeu important d'images réelles issues de capteurs différents, sont présentés et comparés avec ceux fournis par d'autres méthodes connues. Cette évaluation quantitative est menée sur la base de critères calculés à partir d'une connaissance de la segmentation idéale. Les mesures effectuées montrent que l'approche que nous avons developpée fournit, dans des temps de calcul acceptables, des résultats de qualité similaire, sinon supérieure, à ceux obtenus par d'autres techniques.

My thesis deals with clothing habits of young people. The first part descibes the process of market segmentation, the second part abou the lifestyle of young people and about their clothing habits. I used a questionaire and with the help of PASW Statistics I revaled four segments.

Cette thèse porte sur les problèmes de segmentation d’objets et la segmentation sémantique qui visent soit à séparer des objets du fond, soit à l’attribution d’une étiquette sémantique spécifique à chaque pixel de l’image. Nous proposons deux approches pour la segmentation d’objets, et une approche pour la segmentation sémantique. La première approche est basée sur la détection de saillance. Motivés par notre but de segmentation d’objets, un nouveau modèle de détection de saillance est proposé. Cette approche se formule dans le modèle de récupération de la matrice de faible rang en exploitant les informations de structure de l’image provenant d’une segmentation ascendante comme contrainte importante. La segmentation construite à l’aide d’un schéma d’optimisation itératif et conjoint, effectue simultanément, d’une part, une segmentation d’objets basée sur la carte de saillance résultant de sa détection et, d’autre part, une amélioration de la qualité de la saillance à l’aide de la segmentation. Une carte de saillance optimale et la segmentation finale sont obtenues après plusieurs itérations. La deuxième approche proposée pour la segmentation d’objets se fonde sur des images exemples. L’idée sous-jacente est de transférer les étiquettes de segmentation d’exemples similaires, globalement et localement, à l’image requête. Pour l’obtention des exemples les mieux assortis, nous proposons une représentation nouvelle de haut niveau de l’image, à savoir le descripteur orienté objet, qui reflète à la fois l’information globale et locale de l’image. Ensuite, un prédicteur discriminant apprend en ligne à l’aide les exemples récupérés pour attribuer à chaque région de l’image requête un score d’appartenance au premier plan. Ensuite, ces scores sont intégrés dans un schéma de segmentation du champ de Markov (MRF) itératif qui minimise l’énergie. La segmentation sémantique se fonde sur une banque de régions et la représentation parcimonieuse. La banque des régions est un ensemble de régions générées par segmentations multi-niveaux. Ceci est motivé par l’observation que certains objets peuvent être capturés à certains niveaux dans une segmentation hiérarchique. Pour la description de la région, nous proposons la méthode de codage parcimonieux qui représente chaque caractéristique locale avec plusieurs vecteurs de base du dictionnaire visuel appris, et décrit toutes les caractéristiques locales d’une région par un seul histogramme parcimonieux. Une machine à support de vecteurs (SVM) avec apprentissage de noyaux multiple est utilisée pour l’inférence sémantique. Les approches proposées sont largement évaluées sur plusieurs ensembles de données. Des expériences montrent que les approches proposées surpassent les méthodes de l’état de l’art. Ainsi, par rapport au meilleur résultat de la littérature, l’approche proposée de segmentation d’objets améliore la mesure d F-score de 63% à 68,7% sur l’ensemble de données Pascal VOC 2011
This thesis focuses on the problems of object segmentation and semantic segmentation which aim at separating objects from background or assigning a specific semantic label to each pixel in an image. We propose two approaches for the object segmentation and one approach for semantic segmentation. The first proposed approach for object segmentation is based on saliency detection. Motivated by our ultimate goal for object segmentation, a novel saliency detection model is proposed. This model is formulated in the low-rank matrix recovery model by taking the information of image structure derived from bottom-up segmentation as an important constraint. The object segmentation is built in an iterative and mutual optimization framework, which simultaneously performs object segmentation based on the saliency map resulting from saliency detection, and saliency quality boosting based on the segmentation. The optimal saliency map and the final segmentation are achieved after several iterations. The second proposed approach for object segmentation is based on exemplar images. The underlying idea is to transfer segmentation labels of globally and locally similar exemplar images to the query image. For the purpose of finding the most matching exemplars, we propose a novel high-level image representation method called object-oriented descriptor, which captures both global and local information of image. Then, a discriminative predictor is learned online by using the retrieved exemplars. This predictor assigns a probabilistic score of foreground to each region of the query image. After that, the predicted scores are integrated into the segmentation scheme of Markov random field (MRF) energy optimization. Iteratively finding minimum energy of MRF leads the final segmentation. For semantic segmentation, we propose an approach based on region bank and sparse coding. Region bank is a set of regions generated by multi-level segmentations. This is motivated by the observation that some objects might be captured at certain levels in a hierarchical segmentation. For region description, we propose sparse coding method which represents each local feature descriptor with several basic vectors in the learned visual dictionary, and describes all local feature descriptors within a region by a single sparse histogram. With the sparse representation, support vector machine with multiple kernel learning is employed for semantic inference. The proposed approaches have been extensively evaluated on several challenging and widely used datasets. Experiments demonstrated the proposed approaches outperform the stateofthe- art methods. Such as, compared to the best result in the literature, the proposed object segmentation approach based on exemplar images improves the F-score from 63% to 68.7% on Pascal VOC 2011 dataset

A fundamental automatic target recognition (ATR) system can be composed of an object segmentation stage, followed by feature extraction from those objects produced by segmentation, and finally classification of these object features. The capability of such a system in terms of classification success is therefore limited not only by the quality of the feature extraction and classification methods used, but also by the quality of the initial object segmentation. In this thesis, a novel architecture is described which uses two stages of segmentation. This allows image features derived after a primary segmentation stage to influence the parameters of a secondary segmentation stage which is applied to the same image area. This is aimed at allowing improved, and locally optimised, segmentation of those objects which were poorly segmented by the primary segmentation stage. To enable the implementation of the system, a probability density estimate function is used as a method of detecting novelty in objects presented for classification. This is found to be a non-ideal solution, although useful in the context of the application concerned. The development of all the system components, and ultimately the full ATR system, is described with experimental results derived from real-world infrared imagery. From this work, conclusions are drawn as to the usefulness of a such a two-stage segmentation architecture; specifically, the clutter rejection flexibility and the potential ability for the system to locally optimise segmentation on a per object basis are highlighted.

The analysis of glandular morphology within histopathology images is a crucial step in determining the stage of cancer. Manual annotation is a very laborious task. It is time consuming and suffers from the subjectivity of the specialists that label the glands. One of the aims of computational pathology is developing tools to automate gland segmentation. Such an algorithm would improve the efficiency of cancer diag- nosis. This is a complex task as there is a large variability in glandular morphologies and staining techniques. So far, specialised models have given promising results focusing on only one organ. This work investigated the idea of a cross domain ap- proximation. Unlike parenchymae the stroma tissue that lies between the glands is similar throughout all organs in the body. Creating a model able to precisely seg- ment the stroma would pave the way for a cross organ model. It would be able to segment the tissue and therefore give access to gland morphologies of different organs. To address this issue, we investigated different new and former architec- tures such as the MILD-net which is the currently best performing algorithm of the GlaS challenge. New architectures were created based on the promising U shaped network as well as Xception and the ResNet for feature extraction. These networks were trained on colon histopathology images focusing on glands and on the stroma. The comparision of the different results showed that this initial cross domain ap- proximation goes into the right direction and incites for further developments.

Attention has been directed towards audience segmentation due to the added value that it may deliver to a social marketing program. Segmentation offers one means to better utilise limited resources to enhance reach in groups most at need in social marketing projects. While the importance of segmentation is well known, its application in social marketing practice remains limited, and research considering different segmentation approaches is inadequate. In this social marketing project, qualitative and quantitative segmentation approaches were specifically applied to the water use market of residents in the United Arab Emirates. In addition, this project critically evaluated qualitative and quantitative segmentation approaches using cost benefits analysis from the perspective of executive decision makers. Three studies were conducted. First, a qualitative study was conducted using focus groups with participants representative of the water use market of the UAE. The purpose of this study was to segment the water use market using a qualitative approach and generate insights into factors that influence water consumption behaviours. The qualitative study identified four distinct segments on two segmentation bases: participants’ habits and beliefs. These segments were characterised as (1) comfort users; (2) careless users; (3) contradictory users; and (4) price sensitive users. This study further identified factors that influence residents’ water consumption behaviour such as education, accessibility and restrictions, technology, and pricing. In the second study, survey data were gathered to segment the water use market using a quantitative approach. A sample of 1,350 respondents was obtained (875 online, and 475 paper survey). Two-step cluster analysis was employed to segment the water use market based on 19 segmentation variables drawn from demographic, geographic, psychographic, and behavioural bases. A new, augmented model of the Theory of Interpersonal Behaviour (TIB) was used to guide the quantitative segmentation study. This model incorporated newly identified constructs such as religiosity, policy, price, and accessibility in addition to TIB’s main constructs (awareness/knowledge, attitude, facilitating factors, social norms, emotions, habits). Three main segments were generated based on the main criterion of the participants' consumption habits, namely (1) normal users (25% of the total sample, who shared the characteristics of low income, single, living with family, good attitude, low emotional affect, and low water consumption habits); (2) conscious users (half of the sample, characterised by high income, high education level, living inside the campus, high attitude and emotion, religiosity, and average consumption habits); and (3) careless users, characteristically young single students living in campus dormitories, having good attitudes, influenced by friends, and consuming large amounts of water. In the third formative study, semi-structured interviews were conducted to capture executives’ opinions on the more effective segmentation approach to the water use market in an assessment of the costs and benefits of qualitative and quantitative methods. Results indicated that decision makers found the quantitative, over the qualitative, segmentation approach offered greater accuracy in segmentation, and deeper insights about the characteristics of each segment, which, in generating segments for a larger pool of participants, justified its higher cost. This social marketing project makes several theoretical, contextual, and method contributions. It employed an augmented model of the TIB in the field of social marketing by adding new constructs, such as religiosity, and identified the degree to which the TIB constructs explained the variance in water consumption behaviour within a defined community. Furthermore, social marketing principles were utilised in a developing country (the UAE) applying and assessing a qualitative approach to market segmentation and empirically evaluating data-driven segmentation that provided insights into the influencing factors of water consumption behaviours for differentiated segments. Lastly, this research imported a managerial tool, cost benefit analysis, to evaluate the efficacy of qualitative vs. quantitative segmentation approach contributing to executives’ decision-making capabilities.
Thesis (PhD Doctorate)
Doctor of Philosophy (PhD)
Dept of Marketing
Griffith Business School
Full Text

En aquesta tesi s’estudia el problema de la segmentació del moviment. La tesi presenta una revisió dels principals algoritmes de segmentació del moviment, s’analitzen les característiques principals i es proposa una classificació de les tècniques més recents i importants. La segmentació es pot entendre com un problema d’agrupament d’espais (manifold clustering). Aquest estudi aborda alguns dels reptes més difícils de la segmentació de moviment a través l’agrupament d’espais. S’han proposat nous algoritmes per a l’estimació del rang de la matriu de trajectòries, s’ha presenta una mesura de similitud entre subespais, s’han abordat problemes relacionats amb el comportament dels angles canònics i s’ha desenvolupat una eina genèrica per estimar quants moviments apareixen en una seqüència. L´ultima part de l’estudi es dedica a la correcció de l’estimació inicial d’una segmentació. Aquesta correcció es du a terme ajuntant els problemes de la segmentació del moviment i de l’estructura a partir del moviment.
IN THIS STUDY THE PROBLEM OF MOTION SEGMENTATION IS DISCUSSED. MOTION SEGMENTATION STATE OF THE ART IS PRESENTED, THE MAIN FEATURES OF MOTION SEGMENTATION ALGORITHMS ARE ANALYSED, AND A CLASSIFICATION OF THE RECENT AND MOST IMPORTANT TECHNIQUES IS PROPOSED. THE SEGMENTATION PROBLEM COULD BE CAST INTO A MANIFOLD CLUSTERING PROBLEM. IN THIS STUDY SOME OF THE MOST CHALLENGING ISSUES RELATED TO MOTION SEGMENTATION VIA MANIFOLD CLUSTERING ARE TACKLED. NEW ALGORITHMS FOR THE RANK ESTIMATION OF THE TRAJECTORY MATRIX ARE PROPOSED. A MEASURE OF SIMILARITY BETWEEN SUBSPACES IS PRESENTED. THE BEHAVIOUR OF PRINCIPAL ANGLES IS DISCUSSED. A GENERIC TOOL FOR THE ESTIMATION OF THE NUMBER OF MOTIONS IS DEVELOPED. THE LAST PART OF THE STUDY IS DEDICATED TO THE DEVELOPMENT OF AN ALGORITHM FOR THE CORRECTION OF AN INITIAL MOTION SEGMENTATION SOLUTION. SUCH A CORRECTION IS ACHIEVED BY BRINGING TOGETHER THE PROBLEMS OF MOTION SEGMENTATION AND STRUCTURE FROM MOTION.

Els avanços en neurociència han portat el desenvolupament de models complexes de neurones artificials semblants a les seves contraparts biològiques. Aquests models han millorat la funcionalitat de les xarxes neuronals artificials i han aparegut noves aplicacions. No obstant, la complexitat dels nous models de neurones fa que la seva simulació sigui difícil i requereixi molt de temps i un elevat consum de potència. Aquests no són un problemes importants per aplicacions que no tenen restriccions de consum i en la complexitat del sistema com són la simulació de neurones naturals o la recerca de les diferents habilitats de les xarxes neuronals artificials. Però hi ha altres aplicacions, com les de temps real en sistemes portables, que requereixen sistemes ràpids, petits i de baix consum pels que la simulació de complexes neurones artificials no és una bona solució. Trobar una solució viable a aquest problema és l'objectiu d'aquest treball.
Aquesta tesi descriu i analitza el model hardware d'una xarxa neuronal artificial basada en oscil.ladors acoblats que ha estat adaptada per ajustar-se als requeriments d'una realització VLSI i a la seva aplicació a tasques de segmentació d'imatge. Per reduir l'àrea ocupada i el consum de potència, les neurones modelades com oscil.ladors astables, es realitzen en un ASIC full-custom en lloc de ser simulades en una arquitectura hardware standard. La realització física de l'oscil.lador en lloc de simular-lo, permet al sistema portar a terme les mateixes tasques i reduir el consum de potència comparat amb el requeriments que necessita un ordinador per simular la xarxa.
Primer, es modela un oscil.lador astable en mode corrent com a un integrador i un comparador amb histèresi. Després, s'utilitza aquest esquema per estudiar algebraicament i numèricament la sincronització d'oscil.ladors acoblats amb excitacions amb i sense inhibició i mismatch. Després, es repeteix l'anàlisi amb un model millorat que consta de dos integradors amb escales de temps diferents. Això ens permet simular els efectes secundaris com la capacitat de sortida de l'oscil.lador. A partir d'aquests resultats, s'estudia el comportament de matrius de una i dues dimensions d'oscil.ladors acoblats i després la xarxa és usada per segmentar imatges sintètiques.
Basat en els resultats de l'anàlisi matemàtica, es dissenya una xarxa microelectrònica en un ASIC CMOS amb tecnologia 0.8µm i doble capa de polisilici. El circuit és descrit i simulat extensivament per tal de comprovar la seva funcionalitat com a element de segmentació. Després, els resultats experimentals validen la funcionalitat de la xarxa com a element de segmentació i confirmen la importància del efectes secundaris modelats en la secció de l'anàlisi matemàtica. Finalment, la tesi acaba amb una estimació del nivell de complexitat del procés i una comparació amb d'altres mètodes, exposa les conclusions i explora noves possibilitats en la realització hardware d'algorismes neuromorphics per a la segmentació.
L'anàlisi matemàtica i les simulacions demostren que els oscil.ladors astables poden ser usats com a cel.les bàsiques per a xarxes de segmentació. També demostren que els retards deguts a la capacitat de sortida combinada amb el mismatch dels dispositius han de ser limitats per tal que la xarxa treballi correctament. La realització física del model de neurona basat en un oscil.lador no linial demostra que és possible realitzar físicament un sistema de segmentació més ràpid que la seva simulació en ordinadors molt més potents sense perdre funcionalitat.
Noves línies de recerca són l'estudi en més detall dels mecanismes de sincronització amb acoblaments més febles combinat amb el mismatch dels dispositius, l'ús de comparador en mode corrent més ràpids de baix consum i l'ús d'imatges en nivell de grisos.
Advances in neurosciences have induced the development of complex models of artificial neurons closer to their biological counterparts. These models improved functionality of artificial neural networks and novel applications have appeared. Nevertheless, complexity of new neuron models makes their simulation difficult, and time and power consuming. This is not a major drawback for applications that have no restrictions on power consumption and system complexity as simulation of natural neurons or exploration of different abilities of artificial neural networks. But, there are other applications, as found in real time portable systems, that require fast and small systems and low power requirements for which simulating complex artificial neurons is not a good solution. Finding a feasible solution to this problem is the aim of this work.
This dissertation describes and analyzes a hardware model of an artificial neural network based on coupled oscillators that have been adapted to fit VLSI requirements and its applications to scene segmentation tasks. To reduce area overhead and power consumption, neurons, which are modeled as astable oscillators, are implemented on a full custom ASIC instead of being simulated on a standard hardware architecture. The implementation of a physical oscillator instead of their simulation, allows the system to perform the same tasks and reduce power consumption compared to requirements needed for a computer to simulate the network.
First, a current-mode astable oscillator is modeled as an integrator and a hysteresis comparator. Then, this scheme is used to study algebraically and numerically the synchronization of excitatory coupled oscillators with and without external inhibition and mismatch. After this, the analysis is repeated with an improved model composed of two integrators with different timescales. This allows us to simulate secondary effects as oscillator output capacitance. From these results, the behavior of one-dimensional and two-dimensional arrays of coupled oscillators is studied and then, the network is applied to synthetic image segmentation.
Based on results of the mathematical analysis, a microelectronic network is designed on a double-poly 0.8µm CMOS ASIC. This circuit is described and extensively simulated to check its functionality as a segmentation layer. Then, experimental results validate the network functionality as a segmentation network and confirm the importance of secondary effects modeled in the mathematical analysis section. Finally, this dissertation ends with an estimation of the scheme complexity, compares it to other methods, sets out concluding remarks and explores future trends on implementation of neuromorphic segmentation schemes.
Mathematical analysis and simulations demonstrate that astable oscillators can be used as basic cells of segmentation networks. They also demonstrate that delays due to cell output capacitance combined with device mismatch have to be limited below certain boundaries for the network to work properly. The physical implementation of a neuron model based on a non-linear oscillator demonstrates that it is possible to implement an oscillatory segmentation scheme that runs much faster that its simulated counterpart on powerful computers.
Future lines of research are the deeper study of synchronization mechanisms with weaker coupling strength combined with device mismatch, the use of faster current comparators with low power consumption figures and the use of gray level input images.

Es considera que la segmentació de subtítols, és a dir, la forma en què es divideix el text en un subtítol de dues línies, constitueix un dels paràmetres que influeix en la llegibilitat dels subtítols. Durant més de dues dècades, els experts en subtitulat han sostingut que les línies dels subtítols han de dividir-se seguint les regles sintàctiques per facilitar la lectura del text. No obstant això, la indústria audiovisual no té en compte sempre aquestes regles a l'hora de crear els subtítols. Existeixen dues raons que podrien justificar per què no sempre es posen en pràctica aquestes regles: el temps i l'esforç que requereixen els subtituladors per editar els subtítols, així com la necessitat de condensar el text per conservar les unitats de sentit en la mateixa línia. Les investigacions empíriques que s'han dut a terme fins ara no han aportat proves concloents sobre l'impacte directe que té la segmentació sintàctica en la lectura dels subtítols. L'objectiu d'aquesta tesi doctoral és esclarir l'impacte de la segmentació de subtítols duent a terme noves investigacions amb elements que no es van prendre anteriorment en compte: un major nombre de perfils d'usuaris, dispositius amb diferents mides de pantalla i més mesures en el disseny experimental. Es van realitzar tres estudis empírics per determinar si la segmentació de subtítols és un element clau en l'accessibilitat als mitjans audiovisuals. Els dos primers estudis van investigar la importància de respectar les regles sintàctiques en espectadors amb diferents llengües maternes i diferents nivells de pèrdua auditiva. Es van prendre mesures de seguiment ocular, càrrega cognitiva, comprensió i preferències. El tercer estudi va analitzar la recepció de subtítols en dispositius amb diferents mides de pantalla. Es van avaluar els nivells de comprensió i les preferències de presentació de subtítols (centrant l'atenció en estils de segmentació) en cada dispositiu. En general, els resultats d'aquests estudis semblen indicar que la segmentació de subtítols no és un factor determinant en l'accessibilitat dels mitjans audiovisuals. A pesar que els subtítols que no es divideixen sintàcticament augmenten en general la càrrega cognitiva i els moviments oculars, no afecten de manera negativa a la comprensió. Els espectadors aconsegueixen adaptar les estratègies de lectura independentment de l'enfocament emprat per segmentar els subtítols o de la mida de pantalla. Els resultats dels moviments oculars indiquen que les unitats lingüístiques es llegeixen de manera diferent segons la seva divisió a la pantalla, la categoria lingüística a la qual pertanyen i el perfil de l'espectador. Els resultats d'aquesta tesi doctoral analitzen els efectes de la segmentació en la interpretació dels subtítols i l'experiència de l'espectador en el panorama audiovisual actual en constant canvi. S'espera que aquesta tesi doni suport la necessitat de basar les recomanacions i les pràctiques actuals de subtitulat en investigacions empíriques per millorar la qualitat de l'accessibilitat dels mitjans audiovisuals.
Se considera que la segmentación de subtítulos, es decir, la forma en la que se divide el texto en un subtítulo de dos líneas, constituye uno de los parámetros que influye en la legibilidad de los subtítulos. Durante más de dos décadas, los expertos en subtitulado han sostenido que las líneas de los subtítulos deben dividirse siguiendo las reglas sintácticas para facilitar la lectura del texto. Sin embargo, la industria audiovisual no tiene en cuenta siempre estas reglas a la hora de crear los subtítulos. Existen dos razones que podrían justificar por qué no siempre se ponen en práctica dichas reglas: el tiempo y el esfuerzo que requieren los subtituladores para editar los subtítulos, así como la necesidad de condensar el texto para conservar las unidades de sentido en la misma línea. Las investigaciones empíricas que se han llevado a cabo hasta ahora no han aportado pruebas concluyentes sobre el impacto directo que tiene la segmentación sintáctica en la lectura de los subtítulos. El objetivo de esta tesis doctoral es esclarecer el impacto de la segmentación de subtítulos llevando a cabo nuevas investigaciones con elementos que no se tomaron anteriormente en cuenta: un mayor número de perfiles de usuarios, dispositivos con diferentes tamaños de pantalla y más medidas en el diseño experimental. Se realizaron tres estudios empíricos para determinar si la segmentación de subtítulos es un elemento clave en la accesibilidad a los medios audiovisuales. Los dos primeros estudios investigaron la importancia de respetar las reglas sintácticas en espectadores con distintas lenguas maternas y diferentes niveles de pérdida auditiva. Se tomaron medidas de seguimiento ocular, carga cognitiva, comprensión y preferencias. El tercer estudio analizó la recepción de subtítulos en dispositivos con diferentes tamaños de pantalla. Se evaluaron los niveles de comprensión y las preferencias de presentación de subtítulos (centrando la atención en estilos de segmentación) en cada dispositivo. En general, los resultados de estos estudios parecen indicar que la segmentación de subtítulos no es un factor determinante en la accesibilidad de los medios audiovisuales. A pesar de que los subtítulos que no se dividen sintácticamente aumentan por lo general la carga cognitiva y los movimientos oculares, no afectan de manera negativa a la comprensión. Los espectadores consiguen adaptar las estrategias de lectura independientemente del enfoque empleado para segmentar los subtítulos o del tamaño de pantalla. Los resultados de los movimientos oculares indican que las unidades lingüísticas se leen de manera diferente según su división en la pantalla, la categoría lingüística a la que pertenecen y el perfil del espectador. Los resultados de esta tesis doctoral analizan los efectos de la segmentación en la interpretación de los subtítulos y la experiencia del espectador en el panorama audiovisual actual en constante cambio. Se espera que esta tesis respalde la necesidad de basar las recomendaciones y las prácticas actuales de subtitulado en investigaciones empíricas para mejorar la calidad de la accesibilidad de los medios audiovisuales.
Subtitle segmentation, i.e. the way text is divided in a two-line subtitle, is believed to be one of the features that influences the readability of subtitles. For over two decades, experts in subtitling claimed that subtitle lines should be split according to syntactic rules to facilitate the reading process. However, the subtitling industry does not always implement these syntactic rules when creating subtitles. Two reasons could explain why these rules are not always applied: human time and effort to edit subtitles, as well as considerable text reduction to keep units of meaning together in the same line. Previous empirical research on this topic has not provided conclusive evidence as to whether syntactic segmentation has a direct impact on the subtitle reading process. This PhD thesis aims to shed more light on the impact of subtitle segmentation by conducting further research with elements that had not previously be included: a wider range of user profiles, devices with different screen size and more measures in the experimental design. Three empirical studies were carried out to determine whether subtitle segmentation is a key element in Media Accessibility. The first two studies examined the relevance of following syntactic segmentation among viewers with different native languages and hearing statuses, measuring cognitive load, comprehension scores, eye-tracking variables and preferences in line breaks. The third study assessed the reception of subtitles across devices with different screen size, analysing viewers' subtitle layout (specifically focusing on line-break styles) preferences and comprehension. Overall, the results of these studies seem to indicate that subtitle segmentation is not a critical factor in Media Accessibility. Although non-syntactically segmented subtitles generally induce higher cognitive load and more eye movements, they do not negatively affect comprehension. Viewers are able to adapt their reading strategies regardless of the subtitle segmentation approach or the screen size. Eye tracking results demonstrate that linguistic units are processed differently depending on the way they are split on the screen, their linguistic category and the viewers' profile. The results of this PhD thesis discuss the effects of segmentation on subtitle processing and the viewer experience in the context of today’s changing audiovisual landscape. It is hoped that this thesis provides support for the need to base guidelines and current subtitling practices on empirical research evidence to enhance the quality of Media Accessibility.

La fotografia congela en un instant les dades que després es poden extreure, interpretar i transformar amb el pas del temps per comunicar informació en diferents formats. Fer mapes a partir de fotografies va ser una revolució en la cartografia. Els avenços en Computer Vision ajuden a provocar la propera revolució d’aquesta disciplina, que té com a objectiu la informació geogràfica cada cop més detallada que es requereix en períodes de temps més curts. D’aquesta manera, el procés que va de la imatge a un mapa ha esdevingut cada cop més automàtic. Les imatges, ja capturades amb càmeres digitals d’alta resolució, se situen automàticament en la posició correcta del terreny com si fossin un full que el cobreixi, gràcies als models digitals del terreny, obtenint així ortofotomapes. En aquestes circumstàncies, l’única càrrega que queda per alleugerir és l’extracció dels elements topogràfics, sense perdre la precisió i la qualitat d’interpretació que fins ara han proporcionat els operadors humans. Aquesta investigació se centra en el desenvolupament de nous mètodes informàtics que facilitin aquestes tasques d’extreure informació d’imatges aèries. Comencem amb el desenvolupament d’una estratègia d’extracció semi-automàtica de parcel·les de les imatges. Aquest enfocament utilitza la resposta gairebé homogènia dels camps i com aquesta resposta es diferencia de la que proporcionen els veïns. El procés es porta a terme mitjançant el mètode en el qual les regions adjacents competeixen per posseir un píxel. Quan també es tenen en compte les línies de contrast de les imatges, és possible ampliar la metodologia anterior per extreure carreteres. En ambdós casos cal guiar tot el procés, no només pels punts donats per un operador, sinó pel model de l’element que s’ha d’extreure. El model ajuda a perfeccionar els resultats obtinguts. Quan Deep Learning va irrompre a l'escena de la Visió per Computador, es van millorar tots els processos de classificació d'imatges. Així doncs, proposem una aventura conjunta entre una xarxa profunda i un mètode de minimització d'energia adreçat per un model que millora els beneficis de cada component. Aquest enfocament redueix al mínim la necessitat d’interacció humana i obté resultats fiables.
La fotografía congela en un instante los datos que luego se pueden extraer, interpretar y transformar con el tiempo para comunicar información en diferentes formatos. Hacer mapas a partir de fotografías fue una revolución en la cartografía. Los avances en la visión por computador están ayudando a lograr la próxima revolución en esta disciplina, que apunta a una información geográfica cada vez más detallada que se requiere en períodos de tiempo más cortos. De esta manera, el proceso que va de la imagen a un mapa se ha vuelto cada vez más automático. Las imágenes ya capturadas con cámaras digitales de alta resolución se colocan automáticamente en la posición correcta del terreno como si fuera una hoja que lo cubre, gracias a los modelos digitales del terreno, obteniendo así ortofotomapas. En estas circunstancias, la única carga que queda por aligerar es la extracción de los elementos topográficos, sin perder la precisión y la calidad de la interpretación que hasta ahora ha sido proporcionada por operadores humanos. Esta investigación se centra en el desarrollo de nuevos métodos por ordenador que facilitan estas tareas de extracción de información de imágenes aéreas. Comenzamos con el desarrollo de una estrategia para extraer parcelas semi-automáticamente de las imágenes. Este enfoque utiliza la respuesta casi homogénea de las parcelas y cómo esta respuesta difiere de la obtenida de sus vecinas. El proceso se lleva a cabo mediante el método en el que las regiones adyacentes compiten para poseer un píxel. Cuando las líneas de contraste de las imágenes también se tienen en cuenta, es posible ampliar la metodología anterior para extraer carreteras. En ambos casos es necesario guiar todo el proceso, no sólo por los puntos dados por un operador, sino por el modelo del elemento a extraer. El modelo ayuda a refinar los resultados obtenidos. Cuando Deep Learning irrumpió en la escena de Visión por Computador, todos los procesos de clasificación de imágenes mejoraron. Proponemos una aventura conjunta entre una red profunda y un método de minimización de energía guiado por un modelo que mejore los beneficios de cada componente. Este enfoque reduce al mínimo la necesidad de interacción humana y obtiene buenos resultados.
Photography freezes in an instant the data that can later be extracted, interpreted and transformed over time to communicate information in different formats. Making maps from photographs was a revolution in cartography. Advances in Computer Vision are helping to bring about the next revolution in this discipline, which aims at more and more detailed geographic information which is required in shorter periods of time. In this way, the process that goes from image to a map has become increasingly automatic. The images already captured with high-resolution digital cameras are automatically placed in the correct position of the terrain as if they were a sheet that covers it, thanks to the digital terrain models, thus obtaining orthophotomaps. In these circumstances, the only burden that remains to be lightened is the extraction of the topographic elements, without losing the precision and quality of interpretation that until now has been provided by human operators. This research focuses on the development of new computerized methods that facilitate these tasks of extracting information from aerial images. We start with the development of a strategy to semi-automatically extract fields from the images. This approach uses the almost homogeneous response of the fields and how this response differs from that obtained from their neighbors. The process is carried out by means of the method in which adjacent regions compete to own a pixel. When the contrast lines of the images are also taken into account, it is possible to extend the previous methodology to extract roads. In both cases it is necessary to guide the entire process, not only by the points given by an operator, but by the model of the element to be extracted. The model helps to refine the results obtained. When Deep Learning burst onto the Computer Vision scene, all the processes of image classification were upended. So, we propose a joint venture between a deep network and an energy-minimization model-guided radiometric method that improves the benefits of each component. This approach reduces to a minimum the need for human interaction and obtains reliable results.