Development of a visual object tracking system using a compressive sensing framework for modeling the appearance

En el presente trabajo se aborda el problema del seguimiento de objetos, cuyo objetivo es encontrar la trayectoria de un objeto en una secuencia de video. Para ello, se ha desarrollado un método de seguimiento-por-detección que construye un modelo de apariencia en un dominio comprimido usando una nueva e innovadora técnica: “compressive sensing”. La única información necesaria es la situación del objeto a seguir en la primera imagen de la secuencia. El seguimiento de objetos es una aplicación típica del área de visión artificial con un desarrollo de bastantes años. Aun así, sigue siendo una tarea desafiante debido a varios factores: cambios de iluminación, oclusión parcial o total de los objetos y complejidad del fondo de la escena, los cuales deben ser considerados para conseguir un seguimiento robusto. Para lidiar lo más eficazmente posible con estos factores, hemos propuesto un algoritmo de tracking que entrena un clasificador Máquina Vector Soporte (“Support Vector Machine” o SVM en sus siglas en inglés) en modo online para separar los objetos del fondo de la escena. Con este fin, hemos generado nuestro modelo de apariencia por medio de un descriptor de características muy robusto que describe los objetos y el fondo devolviendo un vector de dimensiones muy altas. Por ello, se ha implementado seguidamente un paso para reducir la dimensionalidad de dichos vectores y así poder entrenar nuestro clasificador en un dominio mucho menor, al que denominamos domino comprimido. La reducción de la dimensionalidad de los vectores de características se basa en la teoría de “compressive sensing”, que dice que una señal con poca dispersión (pocos componentes distintos de cero) puede estar bien representada, e incluso puede ser reconstruida, a partir de un conjunto muy pequeño de muestras. La teoría de “compressive sensing” se ha aplicado satisfactoriamente en este trabajo y diferentes técnicas de medida y reconstrucción han sido probadas para evaluar nuestros vectores reducidos, de tal forma que se ha verificado que son capaces de preservar la información de los vectores originales. También incluimos una actualización del modelo de apariencia del objeto a seguir, mediante el reentrenamiento de nuestro clasificador en cada cuadro de la secuencia con muestras positivas y negativas, las cuales han sido obtenidas a partir de la posición predicha por el algoritmo de seguimiento en cada instante temporal. El algoritmo propuesto ha sido evaluado en distintas secuencias y comparado con otros algoritmos del estado del arte de seguimiento, para así demostrar el éxito de nuestro método.

Estudio de verificación biométrica de voz

En este proyecto estudia la posibilidad de realizar una verificación de locutor por medio de la biometría de voz. En primer lugar se obtendrán las características principales de la voz, que serían los coeficientes MFCC, partiendo de una base de datos de diferentes locutores con 10 muestras por cada locutor. Con estos resultados se procederá a la creación de los clasificadores con los que luego testearemos y haremos la verificación. Como resultado final obtendremos un sistema capaz de identificar si el locutor es el que buscamos o no. Para la verificación se utilizan clasificadores Support Vector Machine (SVM), especializado en resolver problemas biclase. Los resultados demuestran que el sistema es capaz de verificar que un locutor es quien dice ser comparándolo con el resto de locutores disponibles en la base de datos. ABSTRACT. Verification based on voice features is an important task for a wide variety of applications concerning biometric verification systems. In this work, we propose a human verification though the use of their voice features focused on supervised training classification algorithms. To this aim we have developed a voice feature extraction system based on MFCC features. For classification purposed we have focused our work in using a Support Vector Machine classificator due to it’s optimization for biclass problems. We test our system in a dataset composed of various individuals of di↵erent gender to evaluate our system’s performance. Experimental results reveal that the proposed system is capable of verificating one individual against the rest of the dataset.

Sistemas de ayuda al diagnóstico y reconocimiento de microcalcificaciones en mamografía mediante descriptores de escala y redes jerárquicas

La presente Tesis investiga el campo del reconocimiento automático de imágenes mediante ordenador aplicado al análisis de imágenes médicas en mamografía digital. Hay un interés por desarrollar sistemas de aprendizaje que asistan a los radiólogos en el reconocimiento de las microcalcificaciones para apoyarles en los programas de cribado y prevención del cáncer de mama. Para ello el análisis de las microcalcificaciones se ha revelado como técnica clave de diagnóstico precoz, pero sin embargo el diseño de sistemas automáticos para reconocerlas es complejo por la variabilidad y condiciones de las imágenes mamográficas. En este trabajo se analizan los planteamientos teóricos de diseño de sistemas de reconocimiento de imágenes, con énfasis en los problemas específicos de detección y clasificación de microcalcificaciones. Se ha realizado un estudio que incluye desde las técnicas de operadores morfológicos, redes neuronales, máquinas de vectores soporte, hasta las más recientes de aprendizaje profundo mediante redes neuronales convolucionales, contemplando la importancia de los conceptos de escala y jerarquía a la hora del diseño y sus implicaciones en la búsqueda de la arquitectura de conexiones y capas de la red. Con estos fundamentos teóricos y elementos de diseño procedentes de otros trabajos en este área realizados por el autor, se implementan tres sistemas de reconocimiento de mamografías que reflejan una evolución tecnológica, culminando en un sistema basado en Redes Neuronales Convolucionales (CNN) cuya arquitectura se diseña gracias al análisis teórico anterior y a los resultados prácticos de análisis de escalas llevados a cabo en nuestra base de datos de imágenes. Los tres sistemas se entrenan y validan con la base de datos de mamografías DDSM, con un total de 100 muestras de entrenamiento y 100 de prueba escogidas para evitar sesgos y reflejar fielmente un programa de cribado. La validez de las CNN para el problema que nos ocupa queda demostrada y se propone un camino de investigación para el diseño de su arquitectura. ABSTRACT This Dissertation investigates the field of computer image recognition applied to medical imaging in mammography. There is an interest in developing learning systems to assist radiologists in recognition of microcalcifications to help them in screening programs for prevention of breast cancer. Analysis of microcalcifications has emerged as a key technique for early diagnosis of breast cancer, but the design of automatic systems to recognize them is complicated by the variability and conditions of mammographic images. In this Thesis the theoretical approaches to design image recognition systems are discussed, with emphasis on the specific problems of detection and classification of microcalcifications. Our study includes techniques ranging from morphological operators, neural networks and support vector machines, to the most recent deep convolutional neural networks. We deal with learning theory by analyzing the importance of the concepts of scale and hierarchy at the design stage and its implications in the search for the architecture of connections and network layers. With these theoretical facts and design elements coming from other works in this area done by the author, three mammogram recognition systems which reflect technological developments are implemented, culminating in a system based on Convolutional Neural Networks (CNN), whose architecture is designed thanks to the previously mentioned theoretical study and practical results of analysis conducted on scales in our image database. All three systems are trained and validated against the DDSM mammographic database, with a total of 100 training samples and 100 test samples chosen to avoid bias and stand for a real screening program. The validity of the CNN approach to the problem is demonstrated and a research way to help in designing the architecture of these networks is proposed.

Sistemas de ayuda al diagnóstico y a la terapia funcional en enfermedades neurodegenerativas

El incremento de la esperanza de vida en los países desarrollados (más de 80 años en 2013), está suponiendo un crecimiento considerable en la incidencia y prevalencia de enfermedades discapacitantes, que si bien pueden aparecer a edades tempranas, son más frecuentes en la tercera edad, o en sus inmediaciones. Enfermedades neuro-degenerativas que suponen un gran hándicap funcional, pues algunas de ellas están asociadas a movimientos involuntarios de determinadas partes del cuerpo, sobre todo de las extremidades. Tareas cotidianas como la ingesta de alimento, vestirse, escribir, interactuar con el ordenador, etc… pueden llegar a ser grandes retos para las personas que las padecen. El diagnóstico precoz y certero resulta fundamental para la prescripción de la terapia o tratamiento óptimo. Teniendo en cuenta incluso que en muchos casos, por desgracia la mayoría, sólo se puede actuar para mitigar los síntomas, y no para sanarlos, al menos de momento. Aun así, acertar de manera temprana en el diagnóstico supone proporcionar al enfermo una mayor calidad de vida durante mucho más tiempo, por lo cual el esfuerzo merece, y mucho, la pena. Los enfermos de Párkinson y de temblor esencial suponen un porcentaje importante de la casuística clínica en los trastornos del movimiento que impiden llevar una vida normal, que producen una discapacidad física y una no menos importante exclusión social. Las vías de tratamiento son dispares de ahí que sea crítico acertar en el diagnóstico lo antes posible. Hasta la actualidad, los profesionales y expertos en medicina, utilizan unas escalas cualitativas para diferenciar la patología y su grado de afectación. Dichas escalas también se utilizan para efectuar un seguimiento clínico y registrar la historia del paciente. En esta tesis se propone una serie de métodos de análisis y de identificación/clasificación de los tipos de temblor asociados a la enfermedad de Párkinson y el temblor esencial. Empleando técnicas de inteligencia artificial basadas en clasificadores inteligentes: redes neuronales (MLP y LVQ) y máquinas de soporte vectorial (SVM), a partir del desarrollo e implantación de un sistema para la medida y análisis objetiva del temblor: DIMETER. Dicho sistema además de ser una herramienta eficaz para la ayuda al diagnóstico, presenta también las capacidades necesarias para proporcionar un seguimiento riguroso y fiable de la evolución de cada paciente. ABSTRACT The increase in life expectancy in developed countries in more than 80 years (data belongs to 2013), is assuming considerable growth in the incidence and prevalence of disabling diseases. Although they may appear at an early age, they are more common in the elderly ages or in its vicinity. Nuero-degenerative diseases that are a major functional handicap, as some of them are associated with involuntary movements of certain body parts, especially of the limbs. Everyday tasks such as food intake, dressing, writing, interact with the computer, etc ... can become large debris for people who suffer. Early and accurate diagnosis is crucial for prescribing optimal therapy or treatment. Even taking into account that in many cases, unfortunately the majority, can only act to mitigate the symptoms, not to cure them, at least for now. Nevertheless, early diagnosis may provide the patient a better quality of life for much longer time, so the effort is worth, and much, grief. Sufferers of Parkinson's and essential tremor represent a significant percentage of clinical casuistry in movement disorders that prevent a normal life, leading to physical disability and not least social exclusion. There are various treatment methods, which makes it necessary the immediate diagnosis. Up to date, professionals and medical experts, use a qualitative scale to differentiate the disease and degree of involvement. Therefore, those scales are used in clinical follow-up. In this thesis, several methods of analysis and identification / classification of types of tremor associated with Parkinson's disease and essential tremor are proposed. Using artificial intelligence techniques based on intelligent classification: neural networks (MLP and LVQ) and support vector machines (SVM), starting from the development and implementation of a system for measuring and objective analysis of the tremor: DIMETER. This system besides being an effective tool to aid diagnosis, it also has the necessary capabilities to provide a rigorous and reliable monitoring of the evolution of each patient.

Real-Time Gene-Expression Extraction with Fast Classification (FC) Networks

Fast Classification (FC) networks were inspired by a biologically plausible mechanism for short term memory where learning occurs instantaneously. Both weights and the topology for an FC network are mapped directly from the training samples by using a prescriptive training scheme. Only two presentations of the training data are required to train an FC network. Compared with iterative learning algorithms such as Back-propagation (which may require many hundreds of presentations of the training data), the training of FC networks is extremely fast and learning convergence is always guaranteed. Thus FC networks may be suitable for applications where real-time classification is needed. In this paper, the FC networks are applied for the real-time extraction of gene expressions for Chlamydia microarray data. Both the classification performance and learning time of the FC networks are compared with the Multi-Layer Proceptron (MLP) networks and support-vector-machines (SVM) in the same classification task. The FC networks are shown to have extremely fast learning time and comparable classification accuracy.

Gaussian processes for classification: mean field algorithms:mean-field algorithms

We derive a mean field algorithm for binary classification with Gaussian processes which is based on the TAP approach originally proposed in Statistical Physics of disordered systems. The theory also yields an approximate leave-one-out estimator for the generalization error which is computed with no extra computational cost. We show that from the TAP approach, it is possible to derive both a simpler 'naive' mean field theory and support vector machines (SVM) as limiting cases. For both mean field algorithms and support vectors machines, simulation results for three small benchmark data sets are presented. They show 1. that one may get state of the art performance by using the leave-one-out estimator for model selection and 2. the built-in leave-one-out estimators are extremely precise when compared to the exact leave-one-out estimate. The latter result is a taken as a strong support for the internal consistency of the mean field approach.

Novel speech signal processing algorithms for high-accuracy classification of Parkinson's disease

There has been considerable recent research into the connection between Parkinson's disease (PD) and speech impairment. Recently, a wide range of speech signal processing algorithms (dysphonia measures) aiming to predict PD symptom severity using speech signals have been introduced. In this paper, we test how accurately these novel algorithms can be used to discriminate PD subjects from healthy controls. In total, we compute 132 dysphonia measures from sustained vowels. Then, we select four parsimonious subsets of these dysphonia measures using four feature selection algorithms, and map these feature subsets to a binary classification response using two statistical classifiers: random forests and support vector machines. We use an existing database consisting of 263 samples from 43 subjects, and demonstrate that these new dysphonia measures can outperform state-of-the-art results, reaching almost 99% overall classification accuracy using only ten dysphonia features. We find that some of the recently proposed dysphonia measures complement existing algorithms in maximizing the ability of the classifiers to discriminate healthy controls from PD subjects. We see these results as an important step toward noninvasive diagnostic decision support in PD.

Predicting delay reductions from freeway traffic diversion for incident management

Traffic incidents are a major source of traffic congestion on freeways. Freeway traffic diversion using pre-planned alternate routes has been used as a strategy to reduce traffic delays due to major traffic incidents. However, it is not always beneficial to divert traffic when an incident occurs. Route diversion may adversely impact traffic on the alternate routes and may not result in an overall benefit. This dissertation research attempts to apply Artificial Neural Network (ANN) and Support Vector Regression (SVR) techniques to predict the percent of delay reduction from route diversion to help determine whether traffic should be diverted under given conditions. The DYNASMART-P mesoscopic traffic simulation model was applied to generate simulated data that were used to develop the ANN and SVR models. A sample network that comes with the DYNASMART-P package was used as the base simulation network. A combination of different levels of incident duration, capacity lost, percent of drivers diverted, VMS (variable message sign) messaging duration, and network congestion was simulated to represent different incident scenarios. The resulting percent of delay reduction, average speed, and queue length from each scenario were extracted from the simulation output. The ANN and SVR models were then calibrated for percent of delay reduction as a function of all of the simulated input and output variables. The results show that both the calibrated ANN and SVR models, when applied to the same location used to generate the calibration data, were able to predict delay reduction with a relatively high accuracy in terms of mean square error (MSE) and regression correlation. It was also found that the performance of the ANN model was superior to that of the SVR model. Likewise, when the models were applied to a new location, only the ANN model could produce comparatively good delay reduction predictions under high network congestion level.

Predicting delay reductions from freeway traffic diversion for incident management

Traffic incidents are a major source of traffic congestion on freeways. Freeway traffic diversion using pre-planned alternate routes has been used as a strategy to reduce traffic delays due to major traffic incidents. However, it is not always beneficial to divert traffic when an incident occurs. Route diversion may adversely impact traffic on the alternate routes and may not result in an overall benefit. This dissertation research attempts to apply Artificial Neural Network (ANN) and Support Vector Regression (SVR) techniques to predict the percent of delay reduction from route diversion to help determine whether traffic should be diverted under given conditions. The DYNASMART-P mesoscopic traffic simulation model was applied to generate simulated data that were used to develop the ANN and SVR models. A sample network that comes with the DYNASMART-P package was used as the base simulation network. A combination of different levels of incident duration, capacity lost, percent of drivers diverted, VMS (variable message sign) messaging duration, and network congestion was simulated to represent different incident scenarios. The resulting percent of delay reduction, average speed, and queue length from each scenario were extracted from the simulation output. The ANN and SVR models were then calibrated for percent of delay reduction as a function of all of the simulated input and output variables. The results show that both the calibrated ANN and SVR models, when applied to the same location used to generate the calibration data, were able to predict delay reduction with a relatively high accuracy in terms of mean square error (MSE) and regression correlation. It was also found that the performance of the ANN model was superior to that of the SVR model. Likewise, when the models were applied to a new location, only the ANN model could produce comparatively good delay reduction predictions under high network congestion level.

Identificación de imágenes por análisis de texturas utilizando aprendizaje estadístico

Este documento explica la creación, implementación y uso del proyecto de fin de grado, desarrollado dentro del grupo de investigación ISCAR (Ingeniería de Sistemas, Control, Automática y Robótica) de la Facultad de Informática de la Universidad Complutense. El proyecto consiste en la implementación de una aplicación capaz de clasificar texturas extraídas de distintas imágenes mediante técnicas de visión por computador. Dicha aplicación se divide en tres pilares fundamentales: interfaz gráfica de usuario, algoritmos de extracción de características texturales y aprendizaje supervisado mediante una máquina “SVM” (Support Vector Machine). Interfaz gráfica: proporciona al usuario una forma fácil de uso de la aplicación por medio de la visualización gráfica de una imagen con una serie de elementos de configuración para su posterior análisis. Una vez analizada, el usuario si así lo desea, podrá visualizar los resultados de manera intuitiva, así como guardar dichos resultados después de la ejecución de los algoritmos pertinentes. Algoritmos de análisis de texturas: Procede al cálculo de las configuraciones y las muestras provistas por el usuario en la interfaz gráfica como el cálculo de la matriz de co-ocurrencia y el cálculo de los vectores de características (homogeneidad, media, varianza, Entropía, etc…). SVM: Utiliza los vectores de características obtenidos en los cálculos estadísticos de texturas para realizar el proceso de aprendizaje de un clasificador SVM. La aplicación ha sido construida en JAVA haciendo uso de librerías como JNI_SVM-light-6.01, commons-math3-3.0 y WindowsBuilder, para la construcción de la ventana gráfica, cálculo de los métodos estadísticos y máquina de aprendizaje automático. Dicha aplicación se ha utilizado con el objetivo de identificar y clasificar el quiste de Baker mediante imágenes obtenidas por Resonancias Magnéticas de la rodilla.

Improving drug discovery using hybrid softcomputing methods

Virtual screening (VS) methods can considerably aid clinical research, predicting how ligands interact with drug targets. Most VS methods suppose a unique binding site for the target, but it has been demonstrated that diverse ligands interact with unrelated parts of the target and many VS methods do not take into account this relevant fact. This problem is circumvented by a novel VS methodology named BINDSURF that scans the whole protein surface in order to find new hotspots, where ligands might potentially interact with, and which is implemented in last generation massively parallel GPU hardware, allowing fast processing of large ligand databases. BINDSURF can thus be used in drug discovery, drug design, drug repurposing and therefore helps considerably in clinical research. However, the accuracy of most VS methods and concretely BINDSURF is constrained by limitations in the scoring function that describes biomolecular interactions, and even nowadays these uncertainties are not completely understood. In order to improve accuracy of the scoring functions used in BINDSURF we propose a hybrid novel approach where neural networks (NNET) and support vector machines (SVM) methods are trained with databases of known active (drugs) and inactive compounds, being this information exploited afterwards to improve BINDSURF VS predictions.

Improvement of Virtual Screening Predictions using Computational Intelligence Methods

Virtual Screening (VS) methods can considerably aid clinical research, predicting how ligands interact with drug targets. However, the accuracy of most VS methods is constrained by limitations in the scoring function that describes biomolecular interactions, and even nowadays these uncertainties are not completely understood. In order to improve accuracy of scoring functions used in most VS methods we propose a hybrid novel approach where neural networks (NNET) and support vector machines (SVM) methods are trained with databases of known active (drugs) and inactive compounds, this information being exploited afterwards to improve VS predictions.

Progetto e implementazione di software di segmentazione assistita di immagini mediche e predizione di biomarkers

La segmentazione prevede la partizione di un'immagine in aree strutturalmente o semanticamente coerenti. Nell'imaging medico, è utilizzata per identificare, contornandole, Regioni di Interesse (ROI) clinico, quali lesioni tumorali, oggetto di approfondimento tramite analisi semiautomatiche e automatiche, o bersaglio di trattamenti localizzati. La segmentazione di lesioni tumorali, assistita o automatica, consiste nell’individuazione di pixel o voxel, in immagini o volumi, appartenenti al tumore. La tecnica assistita prevede che il medico disegni la ROI, mentre quella automatica è svolta da software addestrati, tra cui i sistemi Computer Aided Detection (CAD). Mediante tecniche di visione artificiale, dalle ROI si estraggono caratteristiche numeriche, feature, con valore diagnostico, predittivo, o prognostico. L’obiettivo di questa Tesi è progettare e sviluppare un software di segmentazione assistita che permetta al medico di disegnare in modo semplice ed efficace una o più ROI in maniera organizzata e strutturata per futura elaborazione ed analisi, nonché visualizzazione. Partendo da Aliza, applicativo open-source, visualizzatore di esami radiologici in formato DICOM, è stata estesa l’interfaccia grafica per gestire disegno, organizzazione e memorizzazione automatica delle ROI. Inoltre, è stata implementata una procedura automatica di elaborazione ed analisi di ROI disegnate su lesioni tumorali prostatiche, per predire, di ognuna, la probabilità di cancro clinicamente non-significativo e significativo (con prognosi peggiore). Per tale scopo, è stato addestrato un classificatore lineare basato su Support Vector Machine, su una popolazione di 89 pazienti con 117 lesioni (56 clinicamente significative), ottenendo, in test, accuratezza = 77%, sensibilità = 86% e specificità = 69%. Il sistema sviluppato assiste il radiologo, fornendo una seconda opinione, non vincolante, adiuvante nella definizione del quadro clinico e della prognosi, nonché delle scelte terapeutiche.

Machine learning for spatial environmental data. Theory, applications and software

The book presents the state of the art in machine learning algorithms (artificial neural networks of different architectures, support vector machines, etc.) as applied to the classification and mapping of spatially distributed environmental data. Basic geostatistical algorithms are presented as well. New trends in machine learning and their application to spatial data are given, and real case studies based on environmental and pollution data are carried out. The book provides a CD-ROM with the Machine Learning Office software, including sample sets of data, that will allow both students and researchers to put the concepts rapidly to practice.

Road and Traffic Signs Recognition using Vector Machines

Intelligent Transportation System (ITS) is a system that builds a safe, effective and integrated transportation environment based on advanced technologies. Road signs detection and recognition is an important part of ITS, which offer ways to collect the real time traffic data for processing at a central facility.This project is to implement a road sign recognition model based on AI and image analysis technologies, which applies a machine learning method, Support Vector Machines, to recognize road signs. We focus on recognizing seven categories of road sign shapes and five categories of speed limit signs. Two kinds of features, binary image and Zernike moments, are used for representing the data to the SVM for training and test. We compared and analyzed the performances of SVM recognition model using different features and different kernels. Moreover, the performances using different recognition models, SVM and Fuzzy ARTMAP, are observed.