Software zur automatischen Tumordiagnose in der funktionalen MRT der weiblichen Brust

In den westlichen Industrieländern ist das Mammakarzinom der häufigste bösartige Tumor der Frau. Sein weltweiter Anteil an allen Krebserkrankungen der Frau beläuft sich auf etwa 21 %. Inzwischen ist jede neunte Frau bedroht, während ihres Lebens an Brustkrebs zu erkranken. Die alterstandardisierte Mortalitätrate liegt derzeit bei knapp 27 %.rnrnDas Mammakarzinom hat eine relative geringe Wachstumsrate. Die Existenz eines diagnostischen Verfahrens, mit dem alle Mammakarzinome unter 10 mm Durchmesser erkannt und entfernt werden, würden den Tod durch Brustkrebs praktisch beseitigen. Denn die 20-Jahres-Überlebungsrate bei Erkrankung durch initiale Karzinome der Größe 5 bis 10 mm liegt mit über 95 % sehr hoch.rnrnMit der Kontrastmittel gestützten Bildgebung durch die MRT steht eine relativ junge Untersuchungsmethode zur Verfügung, die sensitiv genug zur Erkennung von Karzinomen ab einer Größe von 3 mm Durchmesser ist. Die diagnostische Methodik ist jedoch komplex, fehleranfällig, erfordert eine lange Einarbeitungszeit und somit viel Erfahrung des Radiologen.rnrnEine Computer unterstützte Diagnosesoftware kann die Qualität einer solch komplexen Diagnose erhöhen oder zumindest den Prozess beschleunigen. Das Ziel dieser Arbeit ist die Entwicklung einer vollautomatischen Diagnose Software, die als Zweitmeinungssystem eingesetzt werden kann. Meines Wissens existiert eine solche komplette Software bis heute nicht.rnrnDie Software führt eine Kette von verschiedenen Bildverarbeitungsschritten aus, die dem Vorgehen des Radiologen nachgeahmt wurden. Als Ergebnis wird eine selbstständige Diagnose für jede gefundene Läsion erstellt: Zuerst eleminiert eine 3d Bildregistrierung Bewegungsartefakte als Vorverarbeitungsschritt, um die Bildqualität der nachfolgenden Verarbeitungsschritte zu verbessern. Jedes kontrastanreichernde Objekt wird durch eine regelbasierte Segmentierung mit adaptiven Schwellwerten detektiert. Durch die Berechnung kinetischer und morphologischer Merkmale werden die Eigenschaften der Kontrastmittelaufnahme, Form-, Rand- und Textureeigenschaften für jedes Objekt beschrieben. Abschließend werden basierend auf den erhobenen Featurevektor durch zwei trainierte neuronale Netze jedes Objekt in zusätzliche Funde oder in gut- oder bösartige Läsionen klassifiziert.rnrnDie Leistungsfähigkeit der Software wurde auf Bilddaten von 101 weiblichen Patientinnen getested, die 141 histologisch gesicherte Läsionen enthielten. Die Vorhersage der Gesundheit dieser Läsionen ergab eine Sensitivität von 88 % bei einer Spezifität von 72 %. Diese Werte sind den in der Literatur bekannten Vorhersagen von Expertenradiologen ähnlich. Die Vorhersagen enthielten durchschnittlich 2,5 zusätzliche bösartige Funde pro Patientin, die sich als falsch klassifizierte Artefakte herausstellten.rn

Classification of interstitial lung disease patterns using local DCT features and random forest.

Over the last decade, a plethora of computer-aided diagnosis (CAD) systems have been proposed aiming to improve the accuracy of the physicians in the diagnosis of interstitial lung diseases (ILD). In this study, we propose a scheme for the classification of HRCT image patches with ILD abnormalities as a basic component towards the quantification of the various ILD patterns in the lung. The feature extraction method relies on local spectral analysis using a DCT-based filter bank. After convolving the image with the filter bank, q-quantiles are computed for describing the distribution of local frequencies that characterize image texture. Then, the gray-level histogram values of the original image are added forming the final feature vector. The classification of the already described patches is done by a random forest (RF) classifier. The experimental results prove the superior performance and efficiency of the proposed approach compared against the state-of-the-art.

Técnicas de análisis, caracterización y detección de señales de voz en entornos acústicos adversos

Este trabajo de Tesis ha abordado el objetivo de dar robustez y mejorar la Detección de Actividad de Voz en entornos acústicos adversos con el fin de favorecer el comportamiento de muchas aplicaciones vocales, por ejemplo aplicaciones de telefonía basadas en reconocimiento automático de voz, aplicaciones en sistemas de transcripción automática, aplicaciones en sistemas multicanal, etc. En especial, aunque se han tenido en cuenta todos los tipos de ruido, se muestra especial interés en el estudio de las voces de fondo, principal fuente de error de la mayoría de los Detectores de Actividad en la actualidad. Las tareas llevadas a cabo poseen como punto de partida un Detector de Actividad basado en Modelos Ocultos de Markov, cuyo vector de características contiene dos componentes: la energía normalizada y la variación de la energía. Las aportaciones fundamentales de esta Tesis son las siguientes: 1) ampliación del vector de características de partida dotándole así de información espectral, 2) ajuste de los Modelos Ocultos de Markov al entorno y estudio de diferentes topologías y, finalmente, 3) estudio e inclusión de nuevas características, distintas de las del punto 1, para filtrar los pulsos de pronunciaciones que proceden de las voces de fondo. Los resultados de detección, teniendo en cuenta los tres puntos anteriores, muestran con creces los avances realizados y son significativamente mejores que los resultados obtenidos, bajo las mismas condiciones, con otros detectores de actividad de referencia. This work has been focused on improving the robustness at Voice Activity Detection in adverse acoustic environments in order to enhance the behavior of many vocal applications, for example telephony applications based on automatic speech recognition, automatic transcription applications, multichannel systems applications, and so on. In particular, though all types of noise have taken into account, this research has special interest in the study of pronunciations coming from far-field speakers, the main error source of most activity detectors today. The tasks carried out have, as starting point, a Hidden Markov Models Voice Activity Detector which a feature vector containing two components: normalized energy and delta energy. The key points of this Thesis are the following: 1) feature vector extension providing spectral information, 2) Hidden Markov Models adjustment to environment and study of different Hidden Markov Model topologies and, finally, 3) study and inclusion of new features, different from point 1, to reject the pronunciations coming from far-field speakers. Detection results, taking into account the above three points, show the advantages of using this method and are significantly better than the results obtained under the same conditions by other well-known voice activity detectors.

Contributions to the contextualization of human-machine spoken interaction systems

The design and development of spoken interaction systems has been a thoroughly studied research scope for the last decades. The aim is to obtain systems with the ability to interact with human agents with a high degree of naturalness and efficiency, allowing them to carry out the actions they desire using speech, as it is the most natural means of communication between humans. To achieve that degree of naturalness, it is not enough to endow systems with the ability to accurately understand the user’s utterances and to properly react to them, even considering the information provided by the user in his or her previous interactions. The system has also to be aware of the evolution of the conditions under which the interaction takes place, in order to act the most coherent way as possible at each moment. Consequently, one of the most important features of the system is that it has to be context-aware. This context awareness of the system can be reflected in the modification of the behaviour of the system taking into account the current situation of the interaction. For instance, the system should decide which action it has to carry out, or the way to perform it, depending on the user that requests it, on the way that the user addresses the system, on the characteristics of the environment in which the interaction takes place, and so on. In other words, the system has to adapt its behaviour to these evolving elements of the interaction. Moreover that adaptation has to be carried out, if possible, in such a way that the user: i) does not perceive that the system has to make any additional effort, or to devote interaction time to perform tasks other than carrying out the requested actions, and ii) does not have to provide the system with any additional information to carry out the adaptation, which could imply a lesser efficiency of the interaction, since users should devote several interactions only to allow the system to become adapted. In the state-of-the-art spoken dialogue systems, researchers have proposed several disparate strategies to adapt the elements of the system to different conditions of the interaction (such as the acoustic characteristics of a specific user’s speech, the actions previously requested, and so on). Nevertheless, to our knowledge there is not any consensus on the procedures to carry out these adaptation. The approaches are to an extent unrelated from one another, in the sense that each one considers different pieces of information, and the treatment of that information is different taking into account the adaptation carried out. In this regard, the main contributions of this Thesis are the following ones: Definition of a contextualization framework. We propose a unified approach that can cover any strategy to adapt the behaviour of a dialogue system to the conditions of the interaction (i.e. the context). In our theoretical definition of the contextualization framework we consider the system’s context as all the sources of variability present at any time of the interaction, either those ones related to the environment in which the interaction takes place, or to the human agent that addresses the system at each moment. Our proposal relies on three aspects that any contextualization approach should fulfill: plasticity (i.e. the system has to be able to modify its behaviour in the most proactive way taking into account the conditions under which the interaction takes place), adaptivity (i.e. the system has also to be able to consider the most appropriate sources of information at each moment, both environmental and user- and dialogue-dependent, to effectively adapt to the conditions aforementioned), and transparency (i.e. the system has to carry out the contextualizaton-related tasks in such a way that the user neither perceives them nor has to do any effort in providing the system with any information that it needs to perform that contextualization). Additionally, we could include a generality aspect to our proposed framework: the main features of the framework should be easy to adopt in any dialogue system, regardless of the solution proposed to manage the dialogue. Once we define the theoretical basis of our contextualization framework, we propose two cases of study on its application in a spoken dialogue system. We focus on two aspects of the interaction: the contextualization of the speech recognition models, and the incorporation of user-specific information into the dialogue flow. One of the modules of a dialogue system that is more prone to be contextualized is the speech recognition system. This module makes use of several models to emit a recognition hypothesis from the user’s speech signal. Generally speaking, a recognition system considers two types of models: an acoustic one (that models each of the phonemes that the recognition system has to consider) and a linguistic one (that models the sequences of words that make sense for the system). In this work we contextualize the language model of the recognition system in such a way that it takes into account the information provided by the user in both his or her current utterance and in the previous ones. These utterances convey information useful to help the system in the recognition of the next utterance. The contextualization approach that we propose consists of a dynamic adaptation of the language model that is used by the recognition system. We carry out this adaptation by means of a linear interpolation between several models. Instead of training the best interpolation weights, we make them dependent on the conditions of the dialogue. In our approach, the system itself will obtain these weights as a function of the reliability of the different elements of information available, such as the semantic concepts extracted from the user’s utterance, the actions that he or she wants to carry out, the information provided in the previous interactions, and so on. One of the aspects more frequently addressed in Human-Computer Interaction research is the inclusion of user specific characteristics into the information structures managed by the system. The idea is to take into account the features that make each user different from the others in order to offer to each particular user different services (or the same service, but in a different way). We could consider this approach as a user-dependent contextualization of the system. In our work we propose the definition of a user model that contains all the information of each user that could be potentially useful to the system at a given moment of the interaction. In particular we will analyze the actions that each user carries out throughout his or her interaction. The objective is to determine which of these actions become the preferences of that user. We represent the specific information of each user as a feature vector. Each of the characteristics that the system will take into account has a confidence score associated. With these elements, we propose a probabilistic definition of a user preference, as the action whose likelihood of being addressed by the user is greater than the one for the rest of actions. To include the user dependent information into the dialogue flow, we modify the information structures on which the dialogue manager relies to retrieve information that could be needed to solve the actions addressed by the user. Usage preferences become another source of contextual information that will be considered by the system towards a more efficient interaction (since the new information source will help to decrease the need of the system to ask users for additional information, thus reducing the number of turns needed to carry out a specific action). To test the benefits of the contextualization framework that we propose, we carry out an evaluation of the two strategies aforementioned. We gather several performance metrics, both objective and subjective, that allow us to compare the improvements of a contextualized system against the baseline one. We will also gather the user’s opinions as regards their perceptions on the behaviour of the system, and its degree of adaptation to the specific features of each interaction. Resumen El diseño y el desarrollo de sistemas de interacción hablada ha sido objeto de profundo estudio durante las pasadas décadas. El propósito es la consecución de sistemas con la capacidad de interactuar con agentes humanos con un alto grado de eficiencia y naturalidad. De esta manera, los usuarios pueden desempeñar las tareas que deseen empleando la voz, que es el medio de comunicación más natural para los humanos. A fin de alcanzar el grado de naturalidad deseado, no basta con dotar a los sistemas de la abilidad de comprender las intervenciones de los usuarios y reaccionar a ellas de manera apropiada (teniendo en consideración, incluso, la información proporcionada en previas interacciones). Adicionalmente, el sistema ha de ser consciente de las condiciones bajo las cuales transcurre la interacción, así como de la evolución de las mismas, de tal manera que pueda actuar de la manera más coherente en cada instante de la interacción. En consecuencia, una de las características primordiales del sistema es que debe ser sensible al contexto. Esta capacidad del sistema de conocer y emplear el contexto de la interacción puede verse reflejada en la modificación de su comportamiento debida a las características actuales de la interacción. Por ejemplo, el sistema debería decidir cuál es la acción más apropiada, o la mejor manera de llevarla a término, dependiendo del usuario que la solicita, del modo en el que lo hace, etcétera. En otras palabras, el sistema ha de adaptar su comportamiento a tales elementos mutables (o dinámicos) de la interacción. Dos características adicionales son requeridas a dicha adaptación: i) el usuario no ha de percibir que el sistema dedica recursos (temporales o computacionales) a realizar tareas distintas a las que aquél le solicita, y ii) el usuario no ha de dedicar esfuerzo alguno a proporcionar al sistema información adicional para llevar a cabo la interacción. Esto último implicaría una menor eficiencia de la interacción, puesto que los usuarios deberían dedicar parte de la misma a proporcionar información al sistema para su adaptación, sin ningún beneficio inmediato. En los sistemas de diálogo hablado propuestos en la literatura, se han propuesto diferentes estrategias para llevar a cabo la adaptación de los elementos del sistema a las diferentes condiciones de la interacción (tales como las características acústicas del habla de un usuario particular, o a las acciones a las que se ha referido con anterioridad). Sin embargo, no existe una estrategia fija para proceder a dicha adaptación, sino que las mismas no suelen guardar una relación entre sí. En este sentido, cada una de ellas tiene en cuenta distintas fuentes de información, la cual es tratada de manera diferente en función de las características de la adaptación buscada. Teniendo en cuenta lo anterior, las contribuciones principales de esta Tesis son las siguientes: Definición de un marco de contextualización. Proponemos un criterio unificador que pueda cubrir cualquier estrategia de adaptación del comportamiento de un sistema de diálogo a las condiciones de la interacción (esto es, el contexto de la misma). En nuestra definición teórica del marco de contextualización consideramos el contexto del sistema como todas aquellas fuentes de variabilidad presentes en cualquier instante de la interacción, ya estén relacionadas con el entorno en el que tiene lugar la interacción, ya dependan del agente humano que se dirige al sistema en cada momento. Nuestra propuesta se basa en tres aspectos que cualquier estrategia de contextualización debería cumplir: plasticidad (es decir, el sistema ha de ser capaz de modificar su comportamiento de la manera más proactiva posible, teniendo en cuenta las condiciones en las que tiene lugar la interacción), adaptabilidad (esto es, el sistema ha de ser capaz de considerar la información oportuna en cada instante, ya dependa del entorno o del usuario, de tal manera que adecúe su comportamiento de manera eficaz a las condiciones mencionadas), y transparencia (que implica que el sistema ha de desarrollar las tareas relacionadas con la contextualización de tal manera que el usuario no perciba la manera en que dichas tareas se llevan a cabo, ni tampoco deba proporcionar al sistema con información adicional alguna). De manera adicional, incluiremos en el marco propuesto el aspecto de la generalidad: las características del marco de contextualización han de ser portables a cualquier sistema de diálogo, con independencia de la solución propuesta en los mismos para gestionar el diálogo. Una vez hemos definido las características de alto nivel de nuestro marco de contextualización, proponemos dos estrategias de aplicación del mismo a un sistema de diálogo hablado. Nos centraremos en dos aspectos de la interacción a adaptar: los modelos empleados en el reconocimiento de habla, y la incorporación de información específica de cada usuario en el flujo de diálogo. Uno de los módulos de un sistema de diálogo más susceptible de ser contextualizado es el sistema de reconocimiento de habla. Este módulo hace uso de varios modelos para generar una hipótesis de reconocimiento a partir de la señal de habla. En general, un sistema de reconocimiento emplea dos tipos de modelos: uno acústico (que modela cada uno de los fonemas considerados por el reconocedor) y uno lingüístico (que modela las secuencias de palabras que tienen sentido desde el punto de vista de la interacción). En este trabajo contextualizamos el modelo lingüístico del reconocedor de habla, de tal manera que tenga en cuenta la información proporcionada por el usuario, tanto en su intervención actual como en las previas. Estas intervenciones contienen información (semántica y/o discursiva) que puede contribuir a un mejor reconocimiento de las subsiguientes intervenciones del usuario. La estrategia de contextualización propuesta consiste en una adaptación dinámica del modelo de lenguaje empleado en el reconocedor de habla. Dicha adaptación se lleva a cabo mediante una interpolación lineal entre diferentes modelos. En lugar de entrenar los mejores pesos de interpolación, proponemos hacer los mismos dependientes de las condiciones actuales de cada diálogo. El propio sistema obtendrá estos pesos como función de la disponibilidad y relevancia de las diferentes fuentes de información disponibles, tales como los conceptos semánticos extraídos a partir de la intervención del usuario, o las acciones que el mismo desea ejecutar. Uno de los aspectos más comúnmente analizados en la investigación de la Interacción Persona-Máquina es la inclusión de las características específicas de cada usuario en las estructuras de información empleadas por el sistema. El objetivo es tener en cuenta los aspectos que diferencian a cada usuario, de tal manera que el sistema pueda ofrecer a cada uno de ellos el servicio más apropiado (o un mismo servicio, pero de la manera más adecuada a cada usuario). Podemos considerar esta estrategia como una contextualización dependiente del usuario. En este trabajo proponemos la definición de un modelo de usuario que contenga toda la información relativa a cada usuario, que pueda ser potencialmente utilizada por el sistema en un momento determinado de la interacción. En particular, analizaremos aquellas acciones que cada usuario decide ejecutar a lo largo de sus diálogos con el sistema. Nuestro objetivo es determinar cuáles de dichas acciones se convierten en las preferencias de cada usuario. La información de cada usuario quedará representada mediante un vector de características, cada una de las cuales tendrá asociado un valor de confianza. Con ambos elementos proponemos una definición probabilística de una preferencia de uso, como aquella acción cuya verosimilitud es mayor que la del resto de acciones solicitadas por el usuario. A fin de incluir la información dependiente de usuario en el flujo de diálogo, llevamos a cabo una modificación de las estructuras de información en las que se apoya el gestor de diálogo para recuperar información necesaria para resolver ciertos diálogos. En dicha modificación las preferencias de cada usuario pasarán a ser una fuente adicional de información contextual, que será tenida en cuenta por el sistema en aras de una interacción más eficiente (puesto que la nueva fuente de información contribuirá a reducir la necesidad del sistema de solicitar al usuario información adicional, dando lugar en consecuencia a una reducción del número de intervenciones necesarias para llevar a cabo una acción determinada). Para determinar los beneficios de las aplicaciones del marco de contextualización propuesto, llevamos a cabo una evaluación de un sistema de diálogo que incluye las estrategias mencionadas. Hemos recogido diversas métricas, tanto objetivas como subjetivas, que nos permiten determinar las mejoras aportadas por un sistema contextualizado en comparación con el sistema sin contextualizar. De igual manera, hemos recogido las opiniones de los participantes en la evaluación acerca de su percepción del comportamiento del sistema, y de su capacidad de adaptación a las condiciones concretas de cada interacción.

Histograms of oriented gradients for fast on-board vehicle verification

Histograms of Oriented Gradients (HoGs) provide excellent results in object detection and verification. However, their demanding processing requirements bound their applicability in some critical real-time scenarios, such as for video-based on-board vehicle detection systems. In this work, an efficient HOG configuration for pose-based on-board vehicle verification is proposed, which alleviates both the processing requirements and required feature vector length without reducing classification performance. The impact on classification of some critical configuration and processing parameters is in depth analyzed to propose a baseline efficient descriptor. Based on the analysis of its cells contribution to classification, new view-dependent cell-configuration patterns are proposed, resulting in reduced descriptors which provide an excellent balance between performance and computational requirements, rendering higher verification rates than other works in the literature.

Estudo do espectro Laplaciano na categorização de imagens

Uma imagem engloba informação que precisa ser organizada para interpretar e compreender seu conteúdo. Existem diversas técnicas computacionais para extrair a principal informação de uma imagem e podem ser divididas em três áreas: análise de cor, textura e forma. Uma das principais delas é a análise de forma, por descrever características de objetos baseadas em seus pontos fronteira. Propomos um método de caracterização de imagens, por meio da análise de forma, baseada nas propriedades espectrais do laplaciano em grafos. O procedimento construiu grafos G baseados nos pontos fronteira do objeto, cujas conexões entre vértices são determinadas por limiares T_l. A partir dos grafos obtêm-se a matriz de adjacência A e a matriz de graus D, as quais definem a matriz Laplaciana L=D -A. A decomposição espectral da matriz Laplaciana (autovalores) é investigada para descrever características das imagens. Duas abordagens são consideradas: a) Análise do vetor característico baseado em limiares e a histogramas, considera dois parâmetros o intervalo de classes IC_l e o limiar T_l; b) Análise do vetor característico baseado em vários limiares para autovalores fixos; os quais representam o segundo e último autovalor da matriz L. As técnicas foram testada em três coleções de imagens: sintéticas (Genéricas), parasitas intestinais (SADPI) e folhas de plantas (CNShape), cada uma destas com suas próprias características e desafios. Na avaliação dos resultados, empregamos o modelo de classificação support vector machine (SVM), o qual avalia nossas abordagens, determinando o índice de separação das categorias. A primeira abordagem obteve um acerto de 90 % com a coleção de imagens Genéricas, 88 % na coleção SADPI, e 72 % na coleção CNShape. Na segunda abordagem, obtém-se uma taxa de acerto de 97 % com a coleção de imagens Genéricas; 83 % para SADPI e 86 % no CNShape. Os resultados mostram que a classificação de imagens a partir do espectro do Laplaciano, consegue categorizá-las satisfatoriamente.

A connectionist computational method for face recognition

In this work, a modified version of the elastic bunch graph matching (EBGM) algorithm for face recognition is introduced. First, faces are detected by using a fuzzy skin detector based on the RGB color space. Then, the fiducial points for the facial graph are extracted automatically by adjusting a grid of points to the result of an edge detector. After that, the position of the nodes, their relation with their neighbors and their Gabor jets are calculated in order to obtain the feature vector defining each face. A self-organizing map (SOM) framework is shown afterwards. Thus, the calculation of the winning neuron and the recognition process are performed by using a similarity function that takes into account both the geometric and texture information of the facial graph. The set of experiments carried out for our SOM-EBGM method shows the accuracy of our proposal when compared with other state-of the-art methods.

Towards effective indexing for very large video sequence database

With rapid advances in video processing technologies and ever fast increments in network bandwidth, the popularity of video content publishing and sharing has made similarity search an indispensable operation to retrieve videos of user interests. The video similarity is usually measured by the percentage of similar frames shared by two video sequences, and each frame is typically represented as a high-dimensional feature vector. Unfortunately, high complexity of video content has posed the following major challenges for fast retrieval: (a) effective and compact video representations, (b) efficient similarity measurements, and (c) efficient indexing on the compact representations. In this paper, we propose a number of methods to achieve fast similarity search for very large video database. First, each video sequence is summarized into a small number of clusters, each of which contains similar frames and is represented by a novel compact model called Video Triplet (ViTri). ViTri models a cluster as a tightly bounded hypersphere described by its position, radius, and density. The ViTri similarity is measured by the volume of intersection between two hyperspheres multiplying the minimal density, i.e., the estimated number of similar frames shared by two clusters. The total number of similar frames is then estimated to derive the overall similarity between two video sequences. Hence the time complexity of video similarity measure can be reduced greatly. To further reduce the number of similarity computations on ViTris, we introduce a new one dimensional transformation technique which rotates and shifts the original axis system using PCA in such a way that the original inter-distance between two high-dimensional vectors can be maximally retained after mapping. An efficient B+-tree is then built on the transformed one dimensional values of ViTris' positions. Such a transformation enables B+-tree to achieve its optimal performance by quickly filtering a large portion of non-similar ViTris. Our extensive experiments on real large video datasets prove the effectiveness of our proposals that outperform existing methods significantly.

Processing of Byzantine Neume Notation in Ancient Historical Manuscripts

This article presents the principal results of the doctoral thesis “Recognition of neume notation in historical documents” by Lasko Laskov (Institute of Mathematics and Informatics at Bulgarian Academy of Sciences), successfully defended before the Specialized Academic Council for Informatics and Mathematical Modelling on 07 June 2010.

Field programmable gate array based target detection and gesture recognition

The move from Standard Definition (SD) to High Definition (HD) represents a six times increases in data, which needs to be processed. With expanding resolutions and evolving compression, there is a need for high performance with flexible architectures to allow for quick upgrade ability. The technology advances in image display resolutions, advanced compression techniques, and video intelligence. Software implementation of these systems can attain accuracy with tradeoffs among processing performance (to achieve specified frame rates, working on large image data sets), power and cost constraints. There is a need for new architectures to be in pace with the fast innovations in video and imaging. It contains dedicated hardware implementation of the pixel and frame rate processes on Field Programmable Gate Array (FPGA) to achieve the real-time performance. ^ The following outlines the contributions of the dissertation. (1) We develop a target detection system by applying a novel running average mean threshold (RAMT) approach to globalize the threshold required for background subtraction. This approach adapts the threshold automatically to different environments (indoor and outdoor) and different targets (humans and vehicles). For low power consumption and better performance, we design the complete system on FPGA. (2) We introduce a safe distance factor and develop an algorithm for occlusion occurrence detection during target tracking. A novel mean-threshold is calculated by motion-position analysis. (3) A new strategy for gesture recognition is developed using Combinational Neural Networks (CNN) based on a tree structure. Analysis of the method is done on American Sign Language (ASL) gestures. We introduce novel point of interests approach to reduce the feature vector size and gradient threshold approach for accurate classification. (4) We design a gesture recognition system using a hardware/ software co-simulation neural network for high speed and low memory storage requirements provided by the FPGA. We develop an innovative maximum distant algorithm which uses only 0.39% of the image as the feature vector to train and test the system design. Database set gestures involved in different applications may vary. Therefore, it is highly essential to keep the feature vector as low as possible while maintaining the same accuracy and performance^

Dynamic classifiers for neonatal brain monitoring

Brain injury due to lack of oxygen or impaired blood flow around the time of birth, may cause long term neurological dysfunction or death in severe cases. The treatments need to be initiated as soon as possible and tailored according to the nature of the injury to achieve best outcomes. The Electroencephalogram (EEG) currently provides the best insight into neurological activities. However, its interpretation presents formidable challenge for the neurophsiologists. Moreover, such expertise is not widely available particularly around the clock in a typical busy Neonatal Intensive Care Unit (NICU). Therefore, an automated computerized system for detecting and grading the severity of brain injuries could be of great help for medical staff to diagnose and then initiate on-time treatments. In this study, automated systems for detection of neonatal seizures and grading the severity of Hypoxic-Ischemic Encephalopathy (HIE) using EEG and Heart Rate (HR) signals are presented. It is well known that there is a lot of contextual and temporal information present in the EEG and HR signals if examined at longer time scale. The systems developed in the past, exploited this information either at very early stage of the system without any intelligent block or at very later stage where presence of such information is much reduced. This work has particularly focused on the development of a system that can incorporate the contextual information at the middle (classifier) level. This is achieved by using dynamic classifiers that are able to process the sequences of feature vectors rather than only one feature vector at a time.

Person Re-Identification using Deep Convnets with Multi-task Learning

Person re-identification involves recognizing a person across non-overlapping camera views, with different pose, illumination, and camera characteristics. We propose to tackle this problem by training a deep convolutional network to represent a person’s appearance as a low-dimensional feature vector that is invariant to common appearance variations encountered in the re-identification problem. Specifically, a Siamese-network architecture is used to train a feature extraction network using pairs of similar and dissimilar images. We show that use of a novel multi-task learning objective is crucial for regularizing the network parameters in order to prevent over-fitting due to the small size the training dataset. We complement the verification task, which is at the heart of re-identification, by training the network to jointly perform verification, identification, and to recognise attributes related to the clothing and pose of the person in each image. Additionally, we show that our proposed approach performs well even in the challenging cross-dataset scenario, which may better reflect real-world expected performance. 

Automatically Evolving Rotation-invariant Texture Image Descriptors by Genetic Programming

In computer vision, training a model that performs classification effectively is highly dependent on the extracted features, and the number of training instances. Conventionally, feature detection and extraction are performed by a domain-expert who, in many cases, is expensive to employ and hard to find. Therefore, image descriptors have emerged to automate these tasks. However, designing an image descriptor still requires domain-expert intervention. Moreover, the majority of machine learning algorithms require a large number of training examples to perform well. However, labelled data is not always available or easy to acquire, and dealing with a large dataset can dramatically slow down the training process. In this paper, we propose a novel Genetic Programming based method that automatically synthesises a descriptor using only two training instances per class. The proposed method combines arithmetic operators to evolve a model that takes an image and generates a feature vector. The performance of the proposed method is assessed using six datasets for texture classification with different degrees of rotation, and is compared with seven domain-expert designed descriptors. The results show that the proposed method is robust to rotation, and has significantly outperformed, or achieved a comparable performance to, the baseline methods.

Implementación de un detector de coral utilizando filtros Gabor Wavelets y máquinas de aprendizaje

Este trabajo se enfoca en la implementación de un detector de arrecife de coral de desempeño rápido que se utiliza para un vehículo autónomo submarino (Autonomous Underwater Vehicle, AUV, por sus siglas en inglés). Una detección rápida de la presencia de coral asegura la estabilización del AUV frente al arrecife en el menor tiempo posible, evitando colisiones con el coral. La detección de coral se hace en una imagen que captura la escena que percibe la cámara del AUV. Se realiza una clasificación píxel por píxel entre dos clases: arrecife de coral y el plano de fondo que no es coral. A cada píxel de la imagen se le asigna un vector característico, el mismo que se genera mediante el uso de filtros Gabor Wavelets. Éstos son implementados en C++ y la librería OpenCV. Los vectores característicos son clasificados a través de nueve algoritmos de máquinas de aprendizaje. El desempeño de cada algoritmo se compara mediante la precisión y el tiempo de ejecución. El algoritmo de Árboles de Decisión resultó ser el más rápido y preciso de entre todos los algoritmos. Se creó una base de datos de 621 imágenes de corales de Belice (110 imágenes de entrenamiento y 511 imágenes de prueba).

Modelling human behaviour and event detection from videos

Automatic analysis of human behaviour in large collections of videos is gaining interest, even more so with the advent of file sharing sites such as YouTube. However, challenges still exist owing to several factors such as inter- and intra-class variations, cluttered backgrounds, occlusion, camera motion, scale, view and illumination changes. This research focuses on modelling human behaviour for action recognition in videos. The developed techniques are validated on large scale benchmark datasets and applied on real-world scenarios such as soccer videos. Three major contributions are made. The first contribution is in the area of proper choice of a feature representation for videos. This involved a study of state-of-the-art techniques for action recognition, feature extraction processing and dimensional reduction techniques so as to yield the best performance with optimal computational requirements. Secondly, temporal modelling of human behaviour is performed. This involved frequency analysis and temporal integration of local information in the video frames to yield a temporal feature vector. Current practices mostly average the frame information over an entire video and neglect the temporal order. Lastly, the proposed framework is applied and further adapted to real-world scenario such as soccer videos. A dataset consisting of video sequences depicting events of players falling is created from actual match data to this end and used to experimentally evaluate the proposed framework.