<p>Face detection has been a research area for more than ten years. It is a complex problem due to the high variability in faces and amongst faces; therefore it is not possible to extract a general pattern to be used for detection. This is what makes the face detection problem a challenge.</p><p>This thesis gives the reader a background to the face detection problem, where the two main approaches of the problem are described. A face detection algorithm is implemented using a context-based method in combination with an evolving neural network. The algorithm consists of two majors steps: detect possible face areas and within these areas detect faces. This method makes it possible to reduce the search space.</p><p>The performance of the algorithm is evaluated and analysed. There are several parameters that affect the performance; the feature extraction method, the classifier and the images used.</p><p>This work resulted in a face detection algorithm and the performance of the algorithm is evaluated and analysed. The analysis of the problems that occurred has provided a deeper understanding for the complexity of the face detection problem.</p>

<p>A technique to detect deviations in sets of systems in a self-organized way is described in this work. System features are extracted to allow compact representation of the system. Distances between systems are calculated by computing distances between the features. The distances are then stored in an affinity matrix. Deviating systems are detected by assuming a statistical model for the affinities. The key idea is to extract features and and identify deviating systems in a self-organized way, using nonlinear techniques for the feature extraction. The results are compared with those achieved with linear techniques, (principal component analysis).</p><p>The features are computed with principal curves and an isometric feature mapping. In the case of principal curves the feature is the curve itself. In the case of isometric feature mapping is the feature a set of curves in the embedding space. The similarity measure between two representations is either the Hausdorff distance, or the Frechet distance. The deviation detection is performed by computing the probability of each system to be observed given all the other systems. To perform reliable inference the Bootstrapping technique was used.</p><p>The technique is demonstrated on simulated and on-road vehicle cooling system data. The results show the applicability and comparison with linear techniques.</p>

<p>It is important to detect faults in bus details at an early stage. Because the driving style affects the breakdown of different details in the bus, identification of the driving style is important to minimize the number of failures in buses.</p><p>The identification of the driving style of the driver was based on the input data which contained examples of the driving runs of each class. K-nearest neighbor and neural networks algorithms were used. Different models were tested.</p><p>It was shown that the results depend on the selected driving runs. A hypothesis was suggested that the examples from different driving runs have different parameters which affect the results of the classification.</p><p>The best results were achieved by using a subset of variables chosen with help of the forward feature selection procedure. The percent of correct classifications is about 89-90 % for the k-nearest neighbor algorithm and 88-93 % for the neural networks.</p><p>Feature selection allowed a significant improvement in the results of the k-nearest neighbor algorithm and in the results of the neural networks algorithm received for the case when the training and testing data sets were selected from the different driving runs. On the other hand, feature selection did not affect the results received with the neural networks for the case when the training and testing data sets were selected from the same driving runs.</p><p>Another way to improve the results is to use smoothing. Computing the average class among a number of consequent examples allowed achieving a decrease in the error.</p>

<p>This thesis presents a study of alternative speech feature extraction methods aimed at increasing robustness of automatic speech recognition (ASR) against additive background noise. </p><p>Spectral peak positions of speech signals remain practically unchanged in presence of additive background noise. Thus, it was expected that emphasizing spectral peak positions in speech feature extraction would result in improved noise robustness of ASR systems. If frequency subbands are properly chosen, dominant subband frequencies can serve as reasonable estimates of spectral peak positions. Thus, different methods for incorporating dominant subband frequencies into speech feature vectors were investigated in this study.</p><p>To begin with, two earlier proposed feature extraction methods that utilize dominant subband frequency information were examined. The first one uses zero-crossing statistics of the subband signals to estimate dominant subband frequencies, while the second one uses subband spectral centroids. The methods were compared with the standard MFCC feature extraction method on two different recognition tasks in various background conditions. The first method was shown to improve ASR performance on both recognition tasks at sufficiently high noise levels. The improvement was, however, smaller on the more complex recognition task. The second method, on the other hand, led to some reduction in ASR performance in all testing conditions.</p><p>Next, a new method for incorporating subband spectral centroids into speech feature vectors was proposed, and was shown to be considerably more robust than the standard MFCC method on both ASR tasks. The main difference between the proposed method and the zero-crossing based method is in the way they utilize dominant subband frequency information. It was shown that the performance improvement due to the use of dominant subband frequency information was considerably larger for the proposed method than for the ZCPA method, especially on the more complex recognition task. Finally, the computational complexity of the proposed method is two orders of magnitude lower than that of the zero-crossing based method, and of the same order of magnitude as the standard MFCC method.</p>

<p>Syftet med denna uppsats är att undersöka hur svenska dagstidningsredaktioner förhåller sig till featurejournalistik i dag och vilken betydelse feature har för papperstidningens överlevnad. Undersökningen är uppdelad i en kvantitativ och en kvalitativ del. Vi har lagt fokus på fyra tidningar – Katrineholms-kuriren, Norrköpings tidningar, Sydsvenska dagbladet och Göteborgs-posten. Som kvantitativ metod har vi räknat antalet texter med featurekaraktär i respektive tidning under en vecka i november 2009. För att kunna bedöma om mängden förändrats under det senaste decenniet, har vi jämfört materialet i samma tidningar 1999. Den kvalitativa delen av studien består av intervjuer med featureansvariga på respektive tidning. Vår teoretiska ram bygger bland annat på forskning vid Oslo universitet som visar att featurejournalistiken tycks bli viktigare för papperstidningar. Detta är ett resultat av att tidningarna konkurrerar med webbredaktionerna som kan uppdatera sina nyheter i realtid, en egenskap som papperstidningen saknar.</p><p>Undersökningen visar att den allmänna tendensen är att feature har ökat i mängd samtidigt som längre nyhetsartiklar har minskat. De vi intervjuat anser att feature är viktigt för att konkurrera med tidningar på Internet och behålla attraktionsvärdet hos prenumeranterna.</p>

Comment les différents aspects d'une scène visuelle sont-ils encodés en mémoire visuelle à court-terme (MVCT) ? S'il est communément admis que la MVCT ne peut stocker simultanément qu'un nombre extrêmement limité d'informations, la question du format des représentations est actuellement loin d'être résolue. Cette thèse avait pour objectif d'étudier les influences de l'organisation perceptive des différents éléments d'une scène visuelle sur la capacité de stockage de la MVCT. Pour ce faire, dix expériences, utilisant le paradigme de détection de changement (c'est-à-dire en présentant deux séries de stimuli séparées par un court laps de temps et en demandant aux participants de détecter un éventuel changement entre ces deux séries) ont été réalisées. La principale manipulation expérimentale était la comparaison entre différentes organisations perceptives des stimuli dans le champ visuel. Les résultats de nos recherches ont contribué de façon significative à une meilleure compréhension de la façon dont les différents aspects d'une scène visuelle sont traités et encodés en MVCT et suggèrent l'existence de différents types de représentations organisées hiérarchiquement. Tout d'abord, les relations spatiales entre tous les éléments d'une scène visuelle seraient encodées sous la forme d'une configuration spatiale. Cette représentation configurale permettrait, dans un deuxième temps, l'analyse et l'encodage de l'identité individuelle des différents éléments. A ce niveau, nous avons montré que le nombre d'éléments pouvant être simultanément encodés et maintenus en MVCT pourrait dépendre considérablement de la manière dont ils sont présentés dans le champ visuel des participants. Ainsi, lorsque différents traits (couleur, forme, texture, orientation, etc.) occupent la même localisation spatiale, formant ainsi une même partie d'un objet, ils seraient traités et encodés aussi efficacement qu'un simple trait. Dans ce cas, la capacité de stockage de la MVCT doit être comprise en terme d'objets intégrés. Par contre, lorsque les traits d'un objet sont spatialement séparés les uns des autres, constituant différentes parties d'objets, le nombre d'éléments pouvant être simultanément encodés en MVCT est fonction du nombre de traits présents dans le champ visuel, et non du nombre d'objets. L'organisation perceptive des informations visuelles a donc une influence directe sur le nombre d'éléments pouvant être simultanément encodés en MVCT en modifiant la nature des représentations./ How are the different aspects of a visual scene encoded in visual short-term memory (VSTM)? Although it is acknowledged that only a small number of information can be simultaneously stored in VSTM, the format of representation is far to be understood. The goal of the present thesis was to study the perceptual influences of the items in a visual scene on the VSTM storage capacity. Ten experiments, using the change detection paradigm (i.e., two stimuli arrays were presented and separated by a short period of time, and the task was to detect a possible change between the two arrays) were conducted. The major experimental manipulation was to contrast different perceptual organisations of the stimuli in an array. The results of those experiments have significantly contributed to a better understanding about how the different aspects of a visual scene are processed and encoded in VSTM and suggest the existence of different types of representation, hierarchically organised. Firstly, the spatial relations between items in a visual scene would be encoded though spatial configuration. Then, this configural representation would allow the processing and the encoding of the identities of each individual item. Here, the number of items that can be simultaneously encoded in VSTM might greatly depend on the way they are presented in the visual field. Indeed, we provided evidence that visual features (colour, shape, texture, orientation), that share the same spatial location, are encoded in VSTM as accurately as single features. In that case, the VSTM storage capacity must be understood in terms of integrated objects rather than in terms of single features. In contrast, when visual features are spatially distributed over different spatial locations, the limitation in the storage capacity is function of the number of features. The perceptual organisation of the visual information has thus a direct influence on the number of items that can be simultaneously encoded in VSTM by modulating the nature of representations.

This work presents a novel technique for 2D human motion capture using a single non calibrated camera. The user's five extremities (head, hands and feet) are extracted, labelled and tracked after silhouette segmentation. As they are the minimal number of points that can be used in order to enable whole body gestural interaction, we will henceforth refer to these features as crucial points. The crucial point candidates are defined as the local maxima of the geodesic distance with respect to the center of gravity of the actor region which lie on the silhouette boundary. In order to disambiguate the selected crucial points into head, left and right foot, left and right hand classes, we propose a Bayesian method that combines a prior human model and the intensities of the tracked crucial points. Due to its low computational complexity, the system can run at real-time paces on standard Personal Computers, with an average error rate range between 2% and 7% in realistic situations, depending on the context and segmentation quality.

Feature matching in images plays an important role in computer vision such as for 3D reconstruction, motion analysis, object recognition, target tracking and dynamic scene analysis. In this paper, we present a robust cooperative strategy to establish the correspondence of the contours between two uncalibrated images based on the recovered epipolar geometry. We take into account two representations of contours in image as contour points and contour chains. The method proposed in the paper is composed of the following two consecutive steps: (1) The first step uses the LMedS method to estimate the fundamental matrix based on Hartley’s 8-point algorithm, (2) The second step uses a new robust cooperative strategy to match contours. The presented approach has been tested with various real images and experimental results show that our method can produce more accurate contour correspondences. / Singapore-MIT Alliance (SMA)

We consider the problem of detecting a large number of different classes of objects in cluttered scenes. Traditional approaches require applying a battery of different classifiers to the image, at multiple locations and scales. This can be slow and can require a lot of training data, since each classifier requires the computation of many different image features. In particular, for independently trained detectors, the (run-time) computational complexity, and the (training-time) sample complexity, scales linearly with the number of classes to be detected. It seems unlikely that such an approach will scale up to allow recognition of hundreds or thousands of objects. We present a multi-class boosting procedure (joint boosting) that reduces the computational and sample complexity, by finding common features that can be shared across the classes (and/or views). The detectors for each class are trained jointly, rather than independently. For a given performance level, the total number of features required, and therefore the computational cost, is observed to scale approximately logarithmically with the number of classes. The features selected jointly are closer to edges and generic features typical of many natural structures instead of finding specific object parts. Those generic features generalize better and reduce considerably the computational cost of an algorithm for multi-class object detection.

Information representation is a critical issue in machine vision. The representation strategy in the primitive stages of a vision system has enormous implications for the performance in subsequent stages. Existing feature extraction paradigms, like edge detection, provide sparse and unreliable representations of the image information. In this thesis, we propose a novel feature extraction paradigm. The features consist of salient, simple parts of regions bounded by zero-crossings. The features are dense, stable, and robust. The primary advantage of the features is that they have abstract geometric attributes pertaining to their size and shape. To demonstrate the utility of the feature extraction paradigm, we apply it to passive navigation. We argue that the paradigm is applicable to other early vision problems.