Thesis (PhD)--University of Stellenbosch, 2001. / ENGLISH ABSTRACT: In this thesis we examine some aspects of automatic face recognition, with specific reference to the eigenface technique. We provide a thorough theoretical analysis of this technique which allows us to explain many of the results reported in the literature. It also suggests that clustering can improve the performance of the system and we provide experimental evidence of this. From the analysis, we also derive an efficient algorithm for updating the eigenfaces. We demonstrate the ability of an eigenface-based system to represent faces efficiently (using at most forty values in our experiments) and also demonstrate our updating algorithm. Since we are concerned with aspects of face recognition, one of the important practical problems is locating the face in a image, subject to distortions such as rotation. We review two well-known methods for locating faces based on the eigenface technique.e These algorithms are computationally expensive, so we illustrate how the Synthetic Discriminant Function can be used to reduce the cost. For our purposes, we propose the concept of a linearly interpolating SDF and we show how this can be used not only to locate the face, but also to estimate the extent of the distortion. We derive conditions which will ensure a SDF is linearly interpolating. We show how many of the more popular SDF-type filters are related to the classic SDF and thus extend our analysis to a wide range of SDF-type filters. Our analysis suggests that by carefully choosing the training set to satisfy our condition, we can significantly reduce the size of the training set required. This is demonstrated by using the equidistributing principle to design a suitable training set for the SDF. All this is illustrated with several examples. Our results with the SDF allow us to construct a two-stage algorithm for locating faces. We use the SDF-type filters to obtain initial estimates of the location and extent of the distortion. This information is then used by one of the more accurate eigenface-based techniques to obtain the final location from a reduced search space. This significantly reduces the computational cost of the process. / AFRIKAANSE OPSOMMING: In hierdie tesis ondersoek ons sommige aspekte van automatiese gesigs- herkenning met spesifieke verwysing na die sogenaamde eigengesig ("eigen- face") tegniek. ‘n Deeglike teoretiese analise van hierdie tegniek stel ons in staat om heelparty van die resultate wat in die literatuur verskyn te verduidelik. Dit bied ook die moontlikheid dat die gedrag van die stelsel sal verbeter as die gesigte in verskillende klasse gegroepeer word. Uit die analise, herlei ons ook ‘n doeltreffende algoritme om die eigegesigte op te dateer. Ons demonstreer die vermoë van die stelsel om gesigte op ‘n doeltreffende manier te beskryf (ons gebruik hoogstens veertig eigegesigte) asook ons opdateringsalgoritme met praktiese voorbeelde. Verder ondersoek ons die belangrike probleem om gesigte in ‘n beeld te vind, veral as rotasie- en skaalveranderinge plaasvind. Ons bespreek twee welbekende algoritmes om gesigte te vind wat op eigengesigte gebaseer is. Hierdie algoritme is baie duur in terme van numerise berekeninge en ons ontwikkel n koste-effektiewe metode wat op die sogenaamde "Synthetic Discriminant Functions" (SDF) gebaseer is. Vir hierdie doel word die begrip van lineêr interpolerende SDF’s ingevoer. Dit stel ons in staat om nie net die gesig te vind nie, maar ook ‘n skatting van sy versteuring te bereken. Voorts kon ons voorwaardes aflei wat verseker dat ‘n SDF lineêr interpolerend is. Aangesien ons aantoon dat baie van die gewilde SDF-tipe filters aan die klassieke SDF verwant is, geld ons resultate vir ‘n hele verskeidenheid SDF- tipe filters. Ons analise toon ook dat ‘n versigtige keuse van die afrigdata mens in staat stel om die grootte van die afrigstel aansienlik te verminder. Dit word duidelik met behulp van die sogenaamde gelykverspreidings beginsel ("equidistributing principle") gedemonstreer. Al hierdie aspekte van die SDF’s word met voorbeelde geïllustreer. Ons resultate met die SDF laat ons toe om ‘n tweestap algoritme vir die vind van ‘n gesig in ‘n beeld te ontwikkel. Ons gebruik eers die SDF-tipe filters om skattings vir die posisie en versteuring van die gesig te kry en dan verfyn ons hierdie skattings deur een van die teknieke wat op eigengesigte gebaseer is te gebruik. Dit lei tot ‘n aansienlike vermindering in die berekeningstyd.

This thesis relates to the design, implementation and evaluation of statistical face recognition techniques. In particular, the use of Hidden Markov Models in various forms is investigated as a recognition tool and critically evaluated. Current face recognition techniques are very dependent on issues like background noise, lighting and position of key features (ie. the eyes, lips etc.). Using an approach which specifically uses an embedded Hidden Markov Model along with spectral domain feature extraction techniques, shows that these dependencies may be lessened while high recognition rates are maintained.

We have a huge amount of video data from extensively available surveillance cameras and increasingly growing technology to record the motion of a moving object in the form of trajectory data. With proliferation of location-enabled devices and ongoing growth in smartphone penetration as well as advancements in exploiting image processing techniques, tracking moving objects is more flawlessly achievable. In this work, we explore some domain-independent qualitative and quantitative features in raw trajectory (spatio-temporal) data in videos captured by a fixed single wide-angle view camera sensor in outdoor areas. We study the efficacy of those features in classifying four basic high level actions by employing two supervised learning algorithms and show how each of the features affect the learning algorithms’ overall accuracy as a single factor or confounded with others.

Scene understanding attempts to produce a textual description of visible and latent concepts in an image to describe the real meaning of the scene. Concepts are either objects, events or relations depicted in an image. To recognize concepts, the decision of object detection algorithm must be further enhanced from visual similarity to semantical compatibility. Semantically relevant concepts convey the most consistent meaning of the scene. Object detectors analyze visual properties (e.g., pixel intensities, texture, color gradient) of sub-regions of an image to identify objects. The initially assigned objects names must be further examined to ensure they are compatible with each other and the scene. By enforcing inter-object dependencies (e.g., co-occurrence, spatial and semantical priors) and object to scene constraints as background information, a concept classifier predicts the most semantically consistent set of names for discovered objects. The additional background information that describes concepts is called context. In this dissertation, a framework for building context-based concept detection is presented that uses a combination of multiple contextual relationships to refine the result of underlying feature-based object detectors to produce most semantically compatible concepts. In addition to the lack of ability to capture semantical dependencies, object detectors suffer from high dimensionality of feature space that impairs them. Variances in the image (i.e., quality, pose, articulation, illumination, and occlusion) can also result in low-quality visual features that impact the accuracy of detected concepts. The object detectors used to build context-based framework experiments in this study are based on the state-of-the-art generative and discriminative graphical models. The relationships between model variables can be easily described using graphical models and the dependencies and precisely characterized using these representations. The generative context-based implementations are extensions of Latent Dirichlet Allocation, a leading topic modeling approach that is very effective in reduction of the dimensionality of the data. The discriminative contextbased approach extends Conditional Random Fields which allows efficient and precise construction of model by specifying and including only cases that are related and influence it. The dataset used for training and evaluation is MIT SUN397. The result of the experiments shows overall 15% increase in accuracy in annotation and 31% improvement in semantical saliency of the annotated concepts. / Includes bibliography. / Dissertation (Ph.D.)--Florida Atlantic University, 2016. / FAU Electronic Theses and Dissertations Collection

Two related image processing problems are addressed in this thesis. First, the problem of identification of 3-D objects in real time is explored. An algorithm to solve this problem and a hardware system for parallel implementation of this algorithm are proposed. The classification scheme is based on the "Invariant Numerical Shape Modeling" (INSM) algorithm originally developed for 2-D pattern recognition such as alphanumeric characters. This algorithm is then extended to 3-D and is used for general 3-D object identification. The hardware system is an SIMD parallel processor, designed in bit slice fashion for expandability. It consists of a library of images coded according to the 3-D INSM algorithm and the SIMD classifier which compares the code of the unknown image to the library codes in a single clock pulse to establish its identity. The output of this system consists of three signals: U, for unique identification; M, for multiple identification; and N, for non-identification of the object. Second, the problem of real time image compaction is addressed. The quad tree data structure is described. Based on this structure, a parallel processor with a tree architecture is developed which is independent of the data entry process, i.e., data may be entered pixel by pixel or all at once. The hardware consists of a tree processor containing a tree generator and three separate memory arrays, a data transfer processor, and a main memory unit. The tree generator generates the quad tree of the input image in tabular form, using the memory arrays in the tree processor for storage of the table. This table can hold one picture frame at a given time. Hence, for processing multiple picture frames the data transfer processor is used to transfer their respective quad trees from the tree processor memory to the main memory. An algorithm is developed to facilitate the determination of the connections in the circuit.

This thesis studies rapid object detection, focusing on feature-based methods. Firstly, modifications of training and detection of the Viola-Jones method are made to improve performance and overcome some of the current limitations such as rotation, occlusion and articulation. New classifiers produced by training and by converting existing classifiers are tested in face detection and hand detection. Secondly, the nature of invariant features in terms of the computational complexity, discrimination power and invariance to rotation and scaling are discussed. A new feature extraction method called Concentric Discs Moment Invariants (CDMI) is developed based on moment invariants and summed-area tables. The dimensionality of this set of features can be increased by using additional concentric discs, rather than using higher order moments. The CDMI set has useful properties, such as speed, rotation invariance, scaling invariance, and rapid contrast stretching can be easily implemented. The results of experiments with face detection shows a clear improvement in accuracy and performance of the CDMI method compared to the standard moment invariants method. Both the CDMI and its variant, using central moments from concentric squares, are used to assess the strength of the method applied to hand-written digits recognition. Finally, the parallelisation of the detection algorithm is discussed. A new model for the specific case of the Viola-Jones method is proposed and tested experimentally. This model takes advantage of the structure of classifiers and of the multi-resolution approach associated with the detection method. The model shows that high speedups can be achieved by broadcasting frames and carrying out the computation of one or more cascades in each node.

This research augments current Multiple Objective Evolutionary Algorithms with methods that dramatically reduce the time required to evolve toward a region of interest in objective space. Multiple Objective Evolutionary Algorithms (MOEAs) are superior to other optimization techniques when the search space is of high dimension and contains many local minima and maxima. Likewise, MOEAs are most interesting when applied to non-intuitive complex systems. But, these systems are often computationally expensive to calculate. When these systems require independent computations to evaluate each objective, the computational expense grows with each additional objective. This method has developed methods that reduces the time required for evolution by reducing the number of objective evaluations, while still evolving solutions that are Pareto optimal. To date, all other Multiple Objective Evolutionary Algorithms (MOEAs) require the evaluation of all objectives before a fitness value can be assigned to an individual. The original contributions of this thesis are: 1. Development of a hierarchical search space description that allows association of crossover and mutation settings with elements of the genotypic description. 2. Development of a method for parallel evaluation of individuals that removes the need for delays for synchronization. 3. Dynamical evolution of thresholds for objectives to allow partial evaluation of objectives for individuals. 4. Dynamic objective orderings to minimize the time required for unnecessary objective evaluations. 5. Application of MOEAs to the computationally expensive flare pattern design domain. 6. Application of MOEAs to the optimization of fielded missile warning receiver algorithms. 7. Development of a new method of using MOEAs for automatic design of pattern recognition systems.

This thesis consists of two parts. The first part focuses on high-dimensional classification problems in microarray experiments. The second part deals with forecasting problems with a large number of categories in predictors. Classification problems in microarray experiments refer to discriminating subjects with different biologic phenotypes or known tumor subtypes as well as to predicting the clinical outcomes or the prognostic stages of subjects. One important characteristic of microarray data is that the number of genes is much larger than the sample size. The penalized logistic regression method is known for simultaneous variable selection and classification. However, the performance of this method declines as the number of variables increases. With this concern, in the first study, we propose a new classification approach that employs the penalized logistic regression method iteratively with a controlled size of gene subsets to maintain variable selection consistency and classification accuracy. The second study is motivated by a modern microarray experiment that includes two layers of replicates. This new experimental setting causes most existing classification methods, including penalized logistic regression, not appropriate to be directly applied because the assumption of independent observations is violated. To solve this problem, we propose a new classification method by incorporating random effects into penalized logistic regression such that the heterogeneity among different experimental subjects and the correlations from repeated measurements can be taken into account. An efficient hybrid algorithm is introduced to tackle computational challenges in estimation and integration. Applications to a breast cancer study show that the proposed classification method obtains smaller models with higher prediction accuracy than the method based on the assumption of independent observations. The second part of this thesis develops a new forecasting approach for large-scale datasets associated with a large number of predictor categories and with predictor structures. The new approach, beyond conventional tree-based methods, incorporates a general linear model and hierarchical splits to make trees more comprehensive, efficient, and interpretable. Through an empirical study in the air cargo industry and a simulation study containing several different settings, the new approach produces higher forecasting accuracy and higher computational efficiency than existing tree-based methods.

This thesis tackles the visual tracking problem as a target contour estimation problem in the face of corrupted measurements. The major aim is to design robust recursive curve filters for accurate contour-based tracking. The state-space representation adopted comprises of a group component and a shape component describing the rigid motion and the non-rigid shape deformation respectively; filtering strategies on each component are then decoupled. The thesis considers two implicit curve descriptors, a classification probability field and the traditional signed distance function, and aims to develop an optimal probabilistic contour observer and locally optimal curve filters. For the former, introducing a novel probabilistic shape description simplifies the filtering problem on the infinite-dimensional space of closed curves to a series of point-wise filtering tasks. The definition and justification of a novel update model suited to the shape space, the derivation of the filtering equations and the relation to Kalman filtering are studied. In addition to the temporal consistency provided by the filtering, extensions involving distributed filtering methods are considered in order to maintain spatial consistency. For the latter, locally optimal closed curve filtering strategies involving curve velocities are explored. The introduction of a local, linear description for planar curve variation and curve uncertainty enables the derivation of a mechanism for estimating the optimal gain associated to the curve filtering process, given quantitative uncertainty levels. Experiments on synthetic and real sequences of images validate the filtering designs.

This thesis deals with image segmentation problems that arise in various computer vision related fields such as medical imaging, satellite imaging, video surveillance, recognition and robotic vision. More specifically, this thesis deals with a special class of image segmentation technique called Snakes or Active Contour Models. In active contour models, image segmentation is posed as an energy minimization problem, where an objective energy function (based on certain image related features) is defined on the segmenting curve (contour). Typically, a gradient descent energy minimization approach is used to drive the initial contour towards a minimum for the defined energy. The drawback associated with this approach is that the contour has a tendency to get stuck at undesired local minima caused by subtle and undesired image features/edges. Thus, active contour based curve evolution approaches are very sensitive to initialization and noise. The central theme of this thesis is to develop techniques that can make active contour models robust against certain classes of local minima by incorporating global information in energy minimization. These techniques lead to energy minimization with global considerations; we call these models -- 'Non-local active contours'. In this thesis, we consider three widely used active contour models: 1) Edge- and region-based segmentation model, 2) Prior shape knowledge based segmentation model, and 3) Motion segmentation model. We analyze the traditional techniques used for these models and establish the need for robust models that avoid local minima. We address the local minima problem for each model by adding global image considerations.