VISION

Author: LLADÓ BARDERA XAVIER.
Year: 2003.
University: GIRONA [www.udg.es].
Place of defense: ESCUELA POLITÉCNICA SUPERIOR.
Place of preparation: ESCUELA POLITÉCNICA SUPERIOR.

Summary: The vision is probably dominate our sense from which derive most of the information world around us. Through the vision we see how things are, where they are and as mueven.En images that we see with our vision system can extract features of color, texture and shape, and with the help of this information we are able to recognize objects even when they are seen in totally different conditions. For example, different viewpoints to the observer distances, lighting conditions, etc.. The Computer Vision tries to emulate the human vision system through a system of capturing images, a computer and software package. The desired goal is nothing less than the desrrollar a sistma able to understand an image in a similar manner as would a person. This thesis focuses on elanálisis the pair perform texture recognition superdficies.La primary motivation is to solve the problem of classifying textured surfaces when they have been caught under sistintas condiciones.En This paper presents a detailed recognition system based on a 3D model of the surface (which includes information about color and shape), which is then used to generate new images 2D textures under new condiciones.Estas virtual images that are generated are the foundation of our system of recognition, as it is used as models referenia for classifier textures. The proposed recognition system combines the matrices Coocurrencia for extracting features of texture, with the use of a Nearest Nieghbour Classifer. This classifier allows us to recognize the various textures and at the same time get an approximation of the direction of illumination in the images used to validate the system. The results obtained in different experiments demonstrate the feasibility of the system genración texture, as well as the recognition system.

Author: Valerio Cascajo Roberto.
Year: 2004.
University: POLITÉCNICA DE MADRID [www.upm.es].
Place of defense: E.T.S. INGENIEROS TELECOMUNICACION.
Place of preparation: E.T.S.I. Telecomunicación.

Summary: In this thesis we develop a schema representation of multi-inspired images in the treatment of non-linear information in the human visual system. This part of the work of Simoncelli and collaborators showing that the primary stages of this process, consisting of a linear filtering stage multiscale (Gabor, wavelet, etc.). Followed by a normalization divisive non-linear, could lead to answers neural statistically independent, which is a very desirable property in a representation of image. Eltrabajo done has been to first place in a systematic analysis of the model proposed by these authors, and a rigorous formulation using mutual information (MI) as metrics of the statistical unit. As a result, we have demonstrated that the responses predicted by this model are not totally statistically independent of each other but, surprisingly, the departures appear to be independent of almost all entries. Although we have seen that it is not possible to achieve a complete independence between neighboring responses, we found that in practice the minimum condition of the IM appears to be very close to zero. Following this analysis, we have solved the problem of roughly optimize the parameters of the model free, ie calculate values that minimize dependency statistics (IM) between neighboring answers. This has particularized is the general expression for a gaussian model, which we have previously verified empirically with a series of natural images. The schema representation resulting image is extremely robust and flexible, it supports various modifications suboptimal enhancing some of its features for applications that require it. All this has been empirically patent through the corresponding implementations and numerical results. Once tested and implemented normalization divisive one of the key contributions has been to solve the problem of reversing the non-linear transformation. This has been proposed and implemented a scheme investable directly, which can be obtained by relaxing slightly status of statistical independence, initially imposed. By be resolved invertibilidad, schema representation can be seen as multipurpose, with clear advantages given their greater relevance and compatibility perceptual and statistical independence between neighboring samples. We have focused on two specific applications: (1) development of a metrics of perceptual image quality and (2) inclusion of non-linear phase of a codec JPEG 2000 to improve the visual quality of the reconstruction.

Author: VERDÚ MONEDERO RAFAEL.
Year: 2005.
University: POLITÉCNICA DE CARTAGENA [www.upct.es].
Place of defense: POLITÉCNICA DE CARTAGENA..
Place of preparation: ANTIGUO HOSPITAL DE MARINA, CAMPUS MURALLA DEL MAR UPCT.

Summary: The contours assets parametric or snakes, are a special case of deformable models embedded in the plane of the image. Their fundamentaos mathematical represent the confluence of the geometry, physics and theory of Approximation. Since its debut in 1987, the usefulness of snakes has been tested and proven in medical image analysis, image segmentation, in tracking moving objects in video sequences, and so on. A major problem to be addressed is the dynamic behavior of the snake when it moves towards its final resolution of balance. This analysis of convergence is particularly necessary when the governing functional active contour depends on parameters or characteristics, usually non-linear, both internal and external forces to contour as inflation and stretched dependent on the structure itself. The design of dynamic behavior total is necessary to control the definition of the functional power to ensure that the existing feedback does not give rise to unstable behavior. In this PhD thesis has been revised formulating the contours active in the domain space, including parametric models, models and adaptable to the topology based on the joint level (level sets). We have described the problems posed by the implementation of the classic contours assets (initialization sensitivity, robustness against noise, selectivity in the segmentation and tracking objects, stopping condition in the iterative procedure, etc.) and possible solutions to current are the art. The design space of deformable models has been moved to the frequency domain and has been used to analyze the speed of convergence. From this analysis provides a method to the rules of designing the optimal parameters of a dynamic contour active for the segmentation of objects in images. The method of dynamic optimal design parameters of active contour has been tested in applications segmentation and tracking of objects in image sequences for accelerating the speed of convengencia of emph () snake. The method has been incorporated in the implementation of movement, characterization mechanical artificial muscles and filtering vector movement in a non-rigid registration method using deformable models. The main contribution of this theory is the analysis of the rate of convergence of the contours assets and a parametric method for establishing the values of their dynamic parameters that control the evolution of the contour in a problem of image segmentation. Another contribution of this thesis is the translation of the classical formulation of deformable models, in the spatial domain, in the frequency domain, providing a new perspective for the design and analysis of these. The extension of the analysis to the case of deformable surfaces to accelerate the speed of convergence, as well as the export of the ideas raised in the analysis for application in joint technical level, are the main lines of future research.

Author: Ila Viorela Simona.
Year: 2005.
University: GIRONA [www.udg.es].
Place of defense: Universidad de Girona.
Place of preparation: Universidad de Girona.

Summary: The work carried out in this thesis provides innovative solutions in the field of image processing submarine. In this environment, the task of image processing is complicated by the lack of well-defined contours due to the smearing of the images on the one hand and the need for a system of artificial lighting that produces a non-uniform illumination. The estimate of the movement of the vehicle and its location are two fundamental problems in underwater robotics. One way to solve these problems is by using a system of computer vision. The vision systems are characterized by their high-resolution, low cost and the fact provide a great deal of information. The estimate of the movement is from the correlation between two images acquired by a camera mounted on the vehicle and oriented seafloor. The correlation can be obtained using techniques of "matching". This thesis proposes an algorithm that can detect correlation between consecutive images in real time. The two main contributions of this thesis is on the one hand, a method that improves the algorithm "matching" by providing it with greater robot, and secondly, the implementation of the algorithm in hardware for the purpose of obtaining a performance in real time. From the point of view algorithmic, the thesis proposes the use of texture features to eliminate false correlation (called "outliers") between two pictures to improve the robustness of the algorithm "matching" and allowing improve the outcome of the motion estimation algorithm that it is very sensitive to the false correlation. The technique proposed in this thesis has been obtained by an extensive study with a large number of experiments to select the operator texture more suited for image processing submarine. Compared with the existing methods, the new proposal has a much lower computational cost and eliminates the need for an a priori estimate of the movement. To make the implementation of the correlation detection algorithm in hardware has designed a parallel architecture that accelerates performance for the purpose of producing a profit for the speed video. The design of the architecture has been made on the basis of a study prepared in VLSI architectures used for the estimation of movement in multimedia applications video encoding. In the particular case of underwater views, it has been determined that there must be a correlation approach based on a similarity measure more complex, taking into account the average value of the intensities for each point. This approach is called "Normalized Cross Correlation Mean" and has the advantage of being robust in the case of a non-uniform illumination. The proposed implementation done consists of two main parts: one is an algorithm in hardware to select points of interest in real-time and on the other side of a parallel architecture to detect the correlation between items belonging to consecutive images. The verification of the implementation has been carried out using platforms based devices reprogramables FPGA. The proposed architecture is characterized by its high flexibility, allowing the change of parameters, and its great effectiveness in relation resources / runtime.

Author: Fischer Sylvain.
Year: 2006.
University: POLITÉCNICA DE MADRID [www.upm.es].
Place of defense: E.T.S. DE INGE. DE TELECOMUNICACION.
Place of preparation: ESCUELA TECNICA SUPERIOR INGENIEROS TELECOMUNICACIÓN.

Summary: This thesis aims to explore some parallels between architecture functional areas and the primary visual image processing. A first objective is to improve existing models of biological vision based on information theory. A second is the development of new algorithms for image processing based on the natural vision. Available data on the visual system covering physiological and psychophysical studies, Gestalt psychology and statistics of natural images. The thesis focuses primarily on the representations sobrecompletas (ie representations that increase the dimensionality of the data) for the following reasons. First, they allow for significant disadvantages exceed the orthogonal transformation, and second because the models of biological vision often need to be sobrecompletos and third because construct representations sobrecompletas efficient and relevant mathematical problems involving novel, in particular the problem of scattered approximations. The first thesis proposes a transformation in ondículas log-Gabor auto-inversible inspired receptive field and the organization multiresolución of the cells simple primary visual cortex (V1). This transformation offers promising results for the elimination of noise. Secondly, the interactions observed between cells V1 consisting of lateral inhibition and facilitation between cell line have shown efficient to extract the natural edges of the images. Thirdly, the redundancy introduced by the transformation sobrecompleta is reduced thanks to a dedicated approximation algorithm which builds a sparse representation sparse images based on their edges. For a decorrelación further and to achieve higher compression rates, the edges aligned along continuous contours are coded so predictive by chains coefficients, which provides an efficient representation of the contours. Finally, it presents a study on closing contours using the methodology tensor voting. We propose the use of iterations and information of curvature to enhance the robustness and perceptual quality of existing methods.

Author: Redondo Tejedor Rafael.
Year: 2006.
University: POLITÉCNICA DE MADRID [www.upm.es].
Place of defense: ETSI TELECOMUNICACIÓN.
Place of preparation: ETSI Telecomunicación.

Summary: The joint representations have experienced a marked swing in recent decades, so much so that there is no area in the signal processing that have not been used. Within the sea of representations exist in the literature, one of which concerns the present work: the implementation log-Gabor proposed in [70, 68]. Its low overlap, high sensitivity orientation and scalability, invarianza to Carrier, auto-invertibilidad and defining complex confer efficiency, versatility and robustness against noise and the appearance of artifacts. Beyond the close resemblance of the filters sobrecompletos log-Gabor the area cortical V1 together with the modeling of behavior inhibition of neuronal / facilitation and poor coding allow a approximación to get the image based on the extraction of the most salient features normally coindidentes with the contours. This kind of representation, based on contours multiscale charts a new path for resolving taréas image processing, namely compression and image fusion. A new paradigm posits a high compression efficiency if the characteristic features of the images are encoded separately, such as luminance, shape or texture [19, 145, 240]. Following this paradigm, in this thesis has proposed a new compression method based on codify these contours multiscale extracted from the processing low log-Gabor. Given the nature of such traits, an encryption algorithm chains has been specially designed according to the stochastic and morphological characteristics of these contours. To this end, different techniques and predictive codes prefjos and arithmetic have been combined in accordance with each alphabet. Furthermore, the proposed algorithm offers a complete compression scheme including the codification of the residue paso-bajo and placement cabezeras of the plot. Such consolidation is based on models of the primary visual cortex to mitigate distortions typically produced by the compression of JPEG compression standards or JPEG2000. The descomposiciones multiresolución have demonstrated their superiority against other traditional techniques of fusion imaging. However, there is no evidence of hegemony, often due to lack of a reference image. In this thesis, various types of wavelets are compared with log-Gabor filters successfully, which had never been used previously because of its traditional lack of accurate reconstruction. Furthermore, an algorithm for schemes multiresolución called windows multitamaño has been proposed, which adapts the window size to the local features in the image by exploiting the advantages of both windows, small or large or accurate and robust, significantly reducing the errors in the maps decision contrary to the traditional techniques fixed-size window. Finally, a new method oriented contours has been proposed to incorporate the contours multiscale scheme merger multiresolución. This algorithm based on traits reduces sensibiliad noise, effects of blurring and artifacts of alignment.

Author: PONSA MUSSARRA DANIEL.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: ESCOLA TÈCNICA SUPERIOR D'ENGINYERIA.
Place of preparation: ESCOLA TÈCNICA SUPERIOR D'ENGINYERIA (ETSE-UAB).

Summary: This thesis focuses the analysis of video sequences, applying model-based techniques for extracting quantitative information. In particular, we make several proposals in two application areas: shape tracking based on contour models, and detection and tracking of vehicles in images acquired by a camera installed on a mobile platform. The work devoted to shape tracking follows the paradigm of active contours, from which we present a review of the existent approaches. First, we measure the performance of the most common algorithms (Kalman based filters and particle filters), and then we evaluate its implementation aspects trough an extensive experimental study, where several synthetic sequences are considered, distorted with different degrees of noise. Thus, we determine the best way to implement in practice these classical tracking algorithms, and we identify its benefits and drawbacks. Next, the work is oriented towards the improvement of contour tracking algorithms based on particle filters. These algorithms reach good results provided that the number of particles is high enough, but unfortunately the required number of particles grows exponentially with the number of parameters to be estimated. Therefore, and in the context of contour tracking, we present three variants of the classical particle filter, corresponding to three new strategies to deal with this problem. First, we propose to improve the contour tracking by propagating more accurately the particles from one image to the next one. This is done by using a linear approximation of the optimal propagation function. The second proposed strategy is based in estimating part of the parameters analytically. Thus, we aim to do a more productive use of the particles, reducing the amount of model parameters that must be estimated through them. The third proposed method aims to exploit the fact that, in contour tracking applications, the parameters related to the rigid transform can be estimated accurately enough independently from the local deformation presented by the contour. This is used to perform a better propagation of the particles, concentrating them more densely in the zone where the tracked contour is located. These three proposals are validated extensively in sequences with different noise levels, on which the reached improvement is evaluated. After this study, we propose to deal directly with the origin of the previous problem by reducing the number of parameters to be estimated in order to follow a given shape of interest. To reach that, we propose to model the shape using multiple models, where each one requires a lower quantity of parameters than when using a unique model. We propose a new method to learn these models from a training set, and a new algorithm to use the obtained models for tracking the contours. The experimental results certify the validity of this proposal. Finally, the thesis focuses on the development of a system for the detection and tracking of vehicles. The proposals include: a vehicle detection module, a module devoted to the determination of the three-dimensional position and velocity of the detected vehicles, and a tracking module for updating the location of vehicles on the road in a precise and efficient manner. Several original contributions are done in these three subjects, and the performance of these is empirically evaluated.

Author: Felip Rodríguez Ramon Lluís.
Year: 2006.
University: AUTÓNOMA DE BARCELONA [www.uab.es].
Place of defense: Escola Tecnica Superior d'Enginyeries.
Place of preparation: Universidad Autónoma de Barcelona.

Summary: Most situations faced by techniques of Computer Vision must deal with data sets that are contaminated by noise and containing a large number of elements that can be considered erroneous regarding the desired results. This fact is even more evident when it comes to estimating a parametric description of the elements to describe the data. Techniques capable of working with elements afectador for errors and noise are necessary in these cases. The statistics and robust algorithms based on data densities are designed to solve such problems and enable the estimation of parametric models although reliable data presented errors. This thesis presents the development of new algorithms estiamación parameters and technical aspects as well as the implementation of algorithms classics and new proposals on Computer Vision tasks dealing with contaminated data. Specifically, this thesis proposes new techniques to analyze video clips in the MPEG domain and to process images LADAR air using robust statistical techniques basades in data densities.