Supervised localimage feature detection

This thesis is about detection of local image features. The research topic belongs to the wider area of object detection, which is a machine vision and pattern recognition problem where an object must be detected (located) in an image. State-of-the-art object detection methods often divide the problem into separate interest point detection and local image description steps, but in this thesis a different technique is used, leading to higher quality image features which enable more precise localization. Instead of using interest point detection the landmark positions are marked manually. Therefore, the quality of the image features is not limited by the interest point detection phase and the learning of image features is simplified. The approach combines both interest point detection and local description into one phase for detection. Computational efficiency of the descriptor is therefore important, leaving out many of the commonly used descriptors as unsuitably heavy. Multiresolution Gabor features has been the main descriptor in this thesis and improving their efficiency is a significant part. Actual image features are formed from descriptors by using a classifierwhich can then recognize similar looking patches in new images. The main classifier is based on Gaussian mixture models. Classifiers are used in one-class classifier configuration where there are only positive training samples without explicit background class. The local image feature detection method has been tested with two freely available face detection databases and a proprietary license plate database. The localization performance was very good in these experiments. Other applications applying the same under-lying techniques are also presented, including object categorization and fault detection.

How to describe species richness patterns for bryophyte conservation?

A large amount of data for inconspicuous taxa is stored in natural history collections; however, this information is often neglected for biodiversity patterns studies. Here, we evaluate the performance of direct interpolation of museum collections data, equivalent to the traditional approach used in bryophyte conservation planning, and stacked species distribution models (S-SDMs) to produce reliable reconstructions of species richness patterns, given that differences between these methods have been insufficiently evaluated for inconspicuous taxa. Our objective was to contrast if species distribution models produce better inferences of diversity richness than simply selecting areas with the higher species numbers. As model species, we selected Iberian species of the genus Grimmia (Bryophyta), and we used four well-collected areas to compare and validate the following models: 1) four Maxent richness models, each generated without the data from one of the four areas, and a reference model created using all of the data and 2) four richness models obtained through direct spatial interpolation, each generated without the data from one area, and a reference model created with all of the data. The correlations between the partial and reference Maxent models were higher in all cases (0.45 to 0.99), whereas the correlations between the spatial interpolation models were negative and weak (-0.3 to -0.06). Our results demonstrate for the first time that S-SDMs offer a useful tool for identifying detailed richness patterns for inconspicuous taxa such as bryophytes and improving incomplete distributions by assessing the potential richness of under-surveyed areas, filling major gaps in the available data. In addition, the proposed strategy would enhance the value of the vast number of specimens housed in biological collections.

Detecting Heterogeneous Risk Attitudes with Mixed Gambles

We propose a task for eliciting attitudes toward risk that is close to real-world risky decisions which typically involve gains and losses. The task consists of accepting or rejecting gambles that provide a gain with probability p and a loss with probability 1−p . We employ finite mixture models to uncover heterogeneity in risk preferences and find that (i) behavior is heterogeneous, with one half of the subjects behaving as expected utility maximizers, (ii) for the others, reference-dependent models perform better than those where subjects derive utility from final outcomes, (iii) models with sign-dependent decision weights perform better than those without, and (iv) there is no evidence for loss aversion. The procedure is sufficiently simple so that it can be easily used in field or lab experiments where risk elicitation is not the main experiment.

Framework for Traffic Pattern Identification: Required Step for Short-term Forecasting

In the world of transport management, the term ‘anticipation’ is gradually replacing ‘reaction’. Indeed, the ability to forecast traffic evolution in a network should ideally form the basis for many traffic management strategies and multiple ITS applications. Real-time prediction capabilities are therefore becoming a concrete need for the management of networks, both for urban and interurban environments, and today’s road operator has increasingly complex and exacting requirements. Recognising temporal patterns in traffic or the manner in which sequential traffic events evolve over time have been important considerations in short-term traffic forecasting. However, little work has been conducted in the area of identifying or associating traffic pattern occurrence with prevailing traffic conditions. This paper presents a framework for detection pattern identification based on finite mixture models using the EM algorithm for parameter estimation. The computation results have been conducted taking into account the traffic data available in an urban network.

Modelos ridge em planejamentos de misturas: uma aplicação na extração da polpa de pequi

Mixture Models can be used in experimental situations involving areas related to food science and chemistry. Some problems of a statistical nature can be found, such as effects of multicollinearity that result in uncertainty in the optimization of a dependent variable. This study proposes the application of the ridge model adapted for mixture planning considering the Kronecker (K-model) and Scheffe (S-Model) methods applied to response surfaces. The method determined the proportions of hexane, acetone and alcohol proportions that resulted in the maximum response of percentage of extracted pequi (Caryocar brasiliense) pulp oil.

Speaker diarization: Segmentation and clustering of speeches

Speaker diarization is the process of sorting speeches according to the speaker. Diarization helps to search and retrieve what a certain speaker uttered in a meeting. Applications of diarization systemsextend to other domains than meetings, for example, lectures, telephone, television, and radio. Besides, diarization enhances the performance of several speech technologies such as speaker recognition, automatic transcription, and speaker tracking. Methodologies previously used in developing diarization systems are discussed. Prior results and techniques are studied and compared. Methods such as Hidden Markov Models and Gaussian Mixture Models that are used in speaker recognition and other speech technologies are also used in speaker diarization. The objective of this thesis is to develop a speaker diarization system in meeting domain. Experimental part of this work indicates that zero-crossing rate can be used effectively in breaking down the audio stream into segments, and adaptive Gaussian Models fit adequately short audio segments. Results show that 35 Gaussian Models and one second as average length of each segment are optimum values to build a diarization system for the tested data. Uniting the segments which are uttered by same speaker is done in a bottom-up clustering by a newapproach of categorizing the mixture weights.

Soil penetration resistance analysis by multivariate and geostatistical methods

The penetration resistance (PR) is a soil attribute that allows identifies areas with restrictions due to compaction, which results in mechanical impedance for root growth and reduced crop yield. The aim of this study was to characterize the PR of an agricultural soil by geostatistical and multivariate analysis. Sampling was done randomly in 90 points up to 0.60 m depth. It was determined spatial distribution models of PR, and defined areas with mechanical impedance for roots growth. The PR showed a random distribution to 0.55 and 0.60 m depth. PR in other depths analyzed showed spatial dependence, with adjustments to exponential and spherical models. The cluster analysis that considered sampling points allowed establishing areas with compaction problem identified in the maps by kriging interpolation. The analysis with main components identified three soil layers, where the middle layer showed the highest values of PR.

Statistical segmentation methods and color variance analysis of retinal images

In this research, the effectiveness of Naive Bayes and Gaussian Mixture Models classifiers on segmenting exudates in retinal images is studied and the results are evaluated with metrics commonly used in medical imaging. Also, a color variation analysis of retinal images is carried out to find how effectively can retinal images be segmented using only the color information of the pixels.

Heterotachy and long-branch attraction in phylogenetics

Affiliation: Département de Biochimie, Faculté de médecine, Université de Montréal

Analyse bayésienne et classification pour modèles continus modifiés à zéro

Les modèles à sur-représentation de zéros discrets et continus ont une large gamme d'applications et leurs propriétés sont bien connues. Bien qu'il existe des travaux portant sur les modèles discrets à sous-représentation de zéro et modifiés à zéro, la formulation usuelle des modèles continus à sur-représentation -- un mélange entre une densité continue et une masse de Dirac -- empêche de les généraliser afin de couvrir le cas de la sous-représentation de zéros. Une formulation alternative des modèles continus à sur-représentation de zéros, pouvant aisément être généralisée au cas de la sous-représentation, est présentée ici. L'estimation est d'abord abordée sous le paradigme classique, et plusieurs méthodes d'obtention des estimateurs du maximum de vraisemblance sont proposées. Le problème de l'estimation ponctuelle est également considéré du point de vue bayésien. Des tests d'hypothèses classiques et bayésiens visant à déterminer si des données sont à sur- ou sous-représentation de zéros sont présentées. Les méthodes d'estimation et de tests sont aussi évaluées au moyen d'études de simulation et appliquées à des données de précipitation agrégées. Les diverses méthodes s'accordent sur la sous-représentation de zéros des données, démontrant la pertinence du modèle proposé. Nous considérons ensuite la classification d'échantillons de données à sous-représentation de zéros. De telles données étant fortement non normales, il est possible de croire que les méthodes courantes de détermination du nombre de grappes s'avèrent peu performantes. Nous affirmons que la classification bayésienne, basée sur la distribution marginale des observations, tiendrait compte des particularités du modèle, ce qui se traduirait par une meilleure performance. Plusieurs méthodes de classification sont comparées au moyen d'une étude de simulation, et la méthode proposée est appliquée à des données de précipitation agrégées provenant de 28 stations de mesure en Colombie-Britannique.

Modèle de mélange de lois multinormales appliqué à l'analyse de comportements et d'habiletés cognitives d'enfants.

Cette étude aborde le thème de l’utilisation des modèles de mélange de lois pour analyser des données de comportements et d’habiletés cognitives mesurées à plusieurs moments au cours du développement des enfants. L’estimation des mélanges de lois multinormales en utilisant l’algorithme EM est expliquée en détail. Cet algorithme simplifie beaucoup les calculs, car il permet d’estimer les paramètres de chaque groupe séparément, permettant ainsi de modéliser plus facilement la covariance des observations à travers le temps. Ce dernier point est souvent mis de côté dans les analyses de mélanges. Cette étude porte sur les conséquences d’une mauvaise spécification de la covariance sur l’estimation du nombre de groupes formant un mélange. La conséquence principale est la surestimation du nombre de groupes, c’est-à-dire qu’on estime des groupes qui n’existent pas. En particulier, l’hypothèse d’indépendance des observations à travers le temps lorsque ces dernières étaient corrélées résultait en l’estimation de plusieurs groupes qui n’existaient pas. Cette surestimation du nombre de groupes entraîne aussi une surparamétrisation, c’est-à-dire qu’on utilise plus de paramètres qu’il n’est nécessaire pour modéliser les données. Finalement, des modèles de mélanges ont été estimés sur des données de comportements et d’habiletés cognitives. Nous avons estimé les mélanges en supposant d’abord une structure de covariance puis l’indépendance. On se rend compte que dans la plupart des cas l’ajout d’une structure de covariance a pour conséquence d’estimer moins de groupes et les résultats sont plus simples et plus clairs à interpréter.

A Note on the Generalization Performance of Kernel Classifiers with Margin

We present distribution independent bounds on the generalization misclassification performance of a family of kernel classifiers with margin. Support Vector Machine classifiers (SVM) stem out of this class of machines. The bounds are derived through computations of the $V_gamma$ dimension of a family of loss functions where the SVM one belongs to. Bounds that use functions of margin distributions (i.e. functions of the slack variables of SVM) are derived.

On Convergence Properties of the EM Algorithm for Gaussian Mixtures

"Expectation-Maximization'' (EM) algorithm and gradient-based approaches for maximum likelihood learning of finite Gaussian mixtures. We show that the EM step in parameter space is obtained from the gradient via a projection matrix $P$, and we provide an explicit expression for the matrix. We then analyze the convergence of EM in terms of special properties of $P$ and provide new results analyzing the effect that $P$ has on the likelihood surface. Based on these mathematical results, we present a comparative discussion of the advantages and disadvantages of EM and other algorithms for the learning of Gaussian mixture models.

Learning from Incomplete Data

Real-world learning tasks often involve high-dimensional data sets with complex patterns of missing features. In this paper we review the problem of learning from incomplete data from two statistical perspectives---the likelihood-based and the Bayesian. The goal is two-fold: to place current neural network approaches to missing data within a statistical framework, and to describe a set of algorithms, derived from the likelihood-based framework, that handle clustering, classification, and function approximation from incomplete data in a principled and efficient manner. These algorithms are based on mixture modeling and make two distinct appeals to the Expectation-Maximization (EM) principle (Dempster, Laird, and Rubin 1977)---both for the estimation of mixture components and for coping with the missing data.

Multivariate Density Estimation: An SVM Approach

We formulate density estimation as an inverse operator problem. We then use convergence results of empirical distribution functions to true distribution functions to develop an algorithm for multivariate density estimation. The algorithm is based upon a Support Vector Machine (SVM) approach to solving inverse operator problems. The algorithm is implemented and tested on simulated data from different distributions and different dimensionalities, gaussians and laplacians in $R^2$ and $R^{12}$. A comparison in performance is made with Gaussian Mixture Models (GMMs). Our algorithm does as well or better than the GMMs for the simulations tested and has the added advantage of being automated with respect to parameters.