AdaBoost has been considered one of the best classifiers ever developed, but two important problems have not yet been addressed. The first is the dependency on the ``weak" learner, and the second is the failure to maintain the performance of learners with small error rates (i.e. ``strong" learners). To solve the first problem, this work proposes using a different learner in each iteration - known as AdaBoost Dynamic (AD) - thereby ensuring that the performance of the algorithm is almost equal to that of the best ``weak" learner executed with AdaBoost.M1. The work then further modifies the procedure to vary the learner in each iteration, in order to locate the learner with the smallest error rate in its training data. This is done using the same weight calculation as in the original AdaBoost; this version is known as AdaBoost Dynamic with Exponential Loss (AB-EL). The results were poor, because AdaBoost does not perform well with strong learners, so, in this sense, the work confirmed previous works' results. To determine how to improve the performance, the weight calculation is modified to use the sigmoid function with algorithm output being the derivative of the same sigmoid function, rather than the logistic regression weight calculation originally used by AdaBoost; this version is known as AdaBoost Dynamic with Logistic Loss (AB-DL). This work presents the convergence proof that binomial weight calculation works, and that this approach improves the results for the strong learner, both theoretically and empirically. AB-DL also has some disadvantages, like the search for the ``best" classifier and that this search reduces the diversity among the classifiers. In order to attack these issues, another algorithm is proposed that combines AD ``weak" leaner execution policy with a small modification of AB-DL's weight calculation, called AdaBoost Dynamic with Added Cost (AD-AC). AD-AC also has a theoretical upper bound error, and the algorithm offers a small accuracy improvement when compared with AB-DL, and traditional AdaBoost approaches. Lastly, this work also adapts AD-AC's weight calculation approach to deal with data stream problem, where classifiers must deal with very large data sets (in the order of millions of instances), and limited memory availability.

AdaBoost has been considered one of the best classifiers ever developed, but two important problems have not yet been addressed. The first is the dependency on the ``weak" learner, and the second is the failure to maintain the performance of learners with small error rates (i.e. ``strong" learners). To solve the first problem, this work proposes using a different learner in each iteration - known as AdaBoost Dynamic (AD) - thereby ensuring that the performance of the algorithm is almost equal to that of the best ``weak" learner executed with AdaBoost.M1. The work then further modifies the procedure to vary the learner in each iteration, in order to locate the learner with the smallest error rate in its training data. This is done using the same weight calculation as in the original AdaBoost; this version is known as AdaBoost Dynamic with Exponential Loss (AB-EL). The results were poor, because AdaBoost does not perform well with strong learners, so, in this sense, the work confirmed previous works' results. To determine how to improve the performance, the weight calculation is modified to use the sigmoid function with algorithm output being the derivative of the same sigmoid function, rather than the logistic regression weight calculation originally used by AdaBoost; this version is known as AdaBoost Dynamic with Logistic Loss (AB-DL). This work presents the convergence proof that binomial weight calculation works, and that this approach improves the results for the strong learner, both theoretically and empirically. AB-DL also has some disadvantages, like the search for the ``best" classifier and that this search reduces the diversity among the classifiers. In order to attack these issues, another algorithm is proposed that combines AD ``weak" leaner execution policy with a small modification of AB-DL's weight calculation, called AdaBoost Dynamic with Added Cost (AD-AC). AD-AC also has a theoretical upper bound error, and the algorithm offers a small accuracy improvement when compared with AB-DL, and traditional AdaBoost approaches. Lastly, this work also adapts AD-AC's weight calculation approach to deal with data stream problem, where classifiers must deal with very large data sets (in the order of millions of instances), and limited memory availability.

The purpose of this paper is to analyze the performance of the Histograms of Oriented Gradients (HOG) as descriptors for traffic signs recognition. The test dataset consists of speed limit traffic signs because of their high inter-class similarities.   HOG features of speed limit signs, which were extracted from different traffic scenes, were computed and a Gentle AdaBoost classifier was invoked to evaluate the different features. The performance of HOG was tested with a dataset consisting of 1727 Swedish speed signs images. Different numbers of HOG features per descriptor, ranging from 36 features up 396 features, were computed for each traffic sign in the benchmark testing. The results show that HOG features perform high classification rate as the Gentle AdaBoost classification rate was 99.42%, and they are suitable to real time traffic sign recognition. However, it is found that changing the number of orientation bins has insignificant effect on the classification rate. In addition to this, HOG descriptors are not robust with respect to sign orientation.

A mesura que la tecnologia evoluciona, apareixen noves aplicacions en el mon de la classificació facial. En el reconeixement de patrons, normalment veiem les cares com a punts en un espai de alta dimensionalitat definit pels valors dels seus pixels. Aquesta aproximació pateix diversos problemes: el fenomen de la "la maledicció de la dimensionalitat", la presència d'oclusions parcials o canvis locals en la il·luminació. Tradicionalment, només les característiques internes de les imatges facials s'han utilitzat per a classificar, on normalment es fa una extracció de característiques. Les tècniques d'extracció de característiques permeten reduir la influencia dels problemes mencionats, reduint també el soroll inherent de les imatges naturals alhora que es poden aprendre característiques invariants de les imatges facials. En la primera part d'aquesta tesi presentem alguns mètodes d'extracció de característiques clàssics: Anàlisi de Components Principals (PCA), Anàlisi de Components Independents (ICA), Factorització No Negativa de Matrius (NMF), i l'Anàlisi Discriminant de Fisher (FLD), totes elles fent alguna mena d'assumpció en les dades a classificar. La principal contribució d'aquest treball es una nova família de tècniques d'extracció de característiques usant el algorisme del Adaboost. El nostre mètode no fa cap assumpció en les dades a classificar, i construeix de forma incremental la matriu de projecció tenint en compte els exemples mes difícilsPer altra banda, en la segon apart de la tesi explorem el rol de les característiques externes en el procés de classificació facial, i presentem un nou mètode per extreure un conjunt alineat de característiques a partir de la informació externa que poden ser combinades amb les tècniques clàssiques millorant els resultats globals de classificació. / As technology evolves, new applications dealing with face classification appear. In pattern recognition, faces are usually seen as points in a high dimensional spaces defined by their pixel values. This approach must deal with several problems such as: the curse of dimensionality, the presence of partial occlusions or local changes in the illumination. Traditionally, only the internal features of face images have been used for classification purposes, where usually a feature extraction step is performed. Feature extraction techniques allow to reduce the influence of the problems mentioned, reducing also the noise inherent from natural images and learning invariant characteristics from face images. In the first part of this thesis some internal feature extraction methods are presented: Principal Component Analysis (PCA), Independent Component Analysis (ICA), Non Negative Matrix Factorization (NMF), and Fisher Linear Discriminant Analysis (FLD), all of them making some kind of the assumption on the data to classify. The main contribution of our work is a non parametric feature extraction family of techniques using the Adaboost algorithm. Our method makes no assumptions on the data to classify, and incrementally builds the projection matrix taking into account the most difficult samples.On the other hand, in the second part of this thesis we also explore the role of external features in face classification purposes, and present a method for extracting an aligned feature set from external face information that can be combined with the classic internal features improving the global performance of the face classification task.

Submitted by Luiz Felipe Barbosa (luiz.fbabreu2@ufpe.br) on 2015-03-10T19:18:52Z No. of bitstreams: 2 DISSERTAÇÃO Thiago Lessa Prata.pdf: 10140720 bytes, checksum: 87a2cd6d796dcf0da7561f1225b82303 (MD5) license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) / Approved for entry into archive by Daniella Sodre (daniella.sodre@ufpe.br) on 2015-03-10T19:42:35Z (GMT) No. of bitstreams: 2 DISSERTAÇÃO Thiago Lessa Prata.pdf: 10140720 bytes, checksum: 87a2cd6d796dcf0da7561f1225b82303 (MD5) license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) / Made available in DSpace on 2015-03-10T19:42:35Z (GMT). No. of bitstreams: 2 DISSERTAÇÃO Thiago Lessa Prata.pdf: 10140720 bytes, checksum: 87a2cd6d796dcf0da7561f1225b82303 (MD5) license_rdf: 1232 bytes, checksum: 66e71c371cc565284e70f40736c94386 (MD5) Previous issue date: 2014-02-14 / Em visão computacional, a área de rastreamento de objetos tem crescido enormemente. O aumento do poder computacional na última década tem permitido que aplicações em tempo real sejam agora possíveis. Em particular, o ramo de rastreamento de objetos tem se beneficiado com essa evolução e agora é utilizado em diversas aplicações desde a área de segurança até a de entretenimento. As primeiras técnicas se baseiam principalmente no vetor de movimento de sub-regiões da imagem e comparação entre as sub-regiões de um quadro do vídeo com o seguinte. Com isso, uma pontuação é computada para cada posição no quadro seguinte no qual o objeto alvo tem maior probabilidade de estar e a posição com maior valor é escolhida como a sua nova posição. Esses rastreadores normalmente são chamados de rastreadores de curto prazo, isso porque uma vez que o objeto é perdido de vista não é possível que ele volte a ser rastreado. Em contrapartida, visando continuar o rastreamento mesmo quando ele é perdido por algum tempo, nos últimos anos uma nova classe de rastreadores foi criada: os rastreadores por detecção. Nestes métodos, uma fase de rastreamento define a posição do objeto em um quadro a partir da sua posição no quadro anterior. Além da fase de rastreamento, uma fase de detecção visa encontrar o objeto sem que haja qualquer dependência com o seu histórico de posicionamento. A resposta de cada uma das duas técnicas é combinada de forma que a nova posição seja determinada. Quando o rastreamento é perdido por causa de alguma condição de ruído (como oclusão ou algum movimento rápido), a detecção é utilizada para reinicializar o rastreamento, o que possibilita a criação de um rastreador de longo prazo. Visando construir tal tipo de rastreador, o presente trabalho elabora um método de rastreamento por detecção. Mais especificamente, o principal objetivo da técnica elaborada é rastrear um objeto em um cenário complexo onde existam outros objetos semelhantes com problemas de difícil tratamento como oclusão, mudança de escala e mudança de pose. Para que isso seja possível, foi utilizado um esquema baseado em detecção, rastreamento e aprendizagem. Na fase de rastreamento, um rastreador de curto prazo comum e consolidado é utilizado. A fase de aprendizagem tem a função de selecionar amostras para o treinamento do módulo de detecção. A fase de detecção é constituída por quatro classificadores em cascata. Dentre eles, o classificador online cascade boosted classifier (OCBC) é utilizado, uma das principais contribuições deste trabalho. O OCBC é um detector do tipo cascata que possui um treinamento em tempo de execução. O método criado foi testado utilizando várias bases de rastreamento de faces com diversos níveis de dificuldade e os resultados mostraram um avanço em relação ao estado da arte.

Protein structure prediction is one of the most important and difficult problems in computational molecular biology. Unlike sequence-only comparison, protein fold recognition based on machine learning algorithms attempts to detect similarities between protein structures which might not be accompanied with any significant sequence similarity. It takes advantage of the information from structural and physic properties beyond sequence information. In this thesis, we present a novel classifier on protein fold recognition, using AdaBoost algorithm that hybrids to k Nearest Neighbor classifier. The experiment framework consists of two tasks: (i) carry out cross validation within the training dataset, and (ii) test on unseen validation dataset, in which 90% of the proteins have less than 25% sequence identity in training samples. Our result yields 64.7% successful rate in classifying independent validation dataset into 27 types of protein folds. Our experiments on the task of protein folding recognition prove the merit of this approach, as it shows that AdaBoost strategy coupling with weak learning classifiers lead to improved and robust performance of 64.7% accuracy versus 61.2% accuracy in published literatures using identical sample sets, feature representation, and class labels.

This paper deals about AdaBoost algorithm, which is used to create a strong classification function using a number of weak classifiers. We familiarize ourselves with modifications of AdaBoost, namely Real AdaBoost, WaldBoost, FloatBoost and TCAcu. These modifications improve some of the properties of algorithm AdaBoost. We discuss some properties of feature and weak classifiers. We show a class of tasks for which AdaBoost algorithm is applicable. We indicate implementation of the library containing that method and we present some tests performed on the implemented library.

Submitted by Geyciane Santos (geyciane_thamires@hotmail.com) on 2015-06-18T15:41:57Z No. of bitstreams: 1 Dissertação - Walter Charles Sousa Seiffert Simões.pdf: 3539559 bytes, checksum: 459e3601771a481f4ba3f7912afe1c26 (MD5) / Approved for entry into archive by Divisão de Documentação/BC Biblioteca Central (ddbc@ufam.edu.br) on 2015-06-19T20:46:30Z (GMT) No. of bitstreams: 1 Dissertação - Walter Charles Sousa Seiffert Simões.pdf: 3539559 bytes, checksum: 459e3601771a481f4ba3f7912afe1c26 (MD5) / Approved for entry into archive by Divisão de Documentação/BC Biblioteca Central (ddbc@ufam.edu.br) on 2015-06-19T20:57:40Z (GMT) No. of bitstreams: 1 Dissertação - Walter Charles Sousa Seiffert Simões.pdf: 3539559 bytes, checksum: 459e3601771a481f4ba3f7912afe1c26 (MD5) / Made available in DSpace on 2015-06-19T20:57:40Z (GMT). No. of bitstreams: 1 Dissertação - Walter Charles Sousa Seiffert Simões.pdf: 3539559 bytes, checksum: 459e3601771a481f4ba3f7912afe1c26 (MD5) Previous issue date: 2014-03-19 / FAPEAM - Fundação de Amparo à Pesquisa do Estado do Amazonas / New Interactive Digital TV (iDTV) users, particularly the Smart TV (Smart TV), have experienced new ways of relating to the TV via remote controls that incorporate more functionality, motion sensing controls and controls through gesture recognition. The layouts operated by these controls had to adapt these new features and functionalities have options either by direct access (a single button), sometimes by indirect access (button combination), and in this second case, make the activity interaction with iDTV quite difficult because they require a level of familiarity with the very high electronic interaction devices. The remote control is the standard device used to allow user interaction with the TV, but sometimes its use becomes dull and difficult. In this scenario, the use of gestures emerges as a more natural and less invasive to assist or replace the remote control as a possible interaction mode. This paper proposed the construction of a prototype for Interactive Digital TV that could be controlled in its operations adjusting sound volume and exchanging programming channel through remote and through a set of gestures, defined from a set rules of Usability Engineering. The defining the layout and set of gestures process was conducted with the participation of users, and from these definitions, the features of the TV so constructed as to have the on-screen information about the volume and channel programming in addition to the gesture recognition process. The prototype built in this paper differs from commercial products because it took into account the relationship between the cost and performance of the devices used, seeking to offer an affordable and flexible option as to their use in different equipment. The approach described in this paper addresses the challenges that were faced in the areas of Engineering, Usability, iDTV and Computer Vision with final prototype compared to other methods and products, showing a performance of approximately 95% in the accuracy of the displayed gesture and a rate of speed 26 frames per second. / Usuários da nova TV Digital Interativa (TVDi), particularmente da TV Inteligente (Smart TV), têm experimentado novas formas de se relacionar com a TV através controles remotos que incorporaram mais funcionalidades, controles de detecção de movimento e os controles através de reconhecimento de gestos. Os layouts operados por estes controles tiveram que se adequar as estas novas funcionalidades e apresentam as opções de funcionalidades ora por acesso direto (um único botão), ora pelo acesso indireto (combinação de botões), e, neste segundo caso, tornam a atividade de interação com a TVDi bastante difícil pois exigem um nível de vivência com os dispositivos eletrônicos de interação muito elevado. O controle remoto é o dispositivo padrão utilizado para permitir a interação do usuário com a TV, mas em alguns momentos a sua utilização se torna maçante e difícil. Neste cenário, o uso dos gestos surge como um modo mais natural e menos invasivo para auxiliar ou substituir o controle remoto como possibilidade de interação. Este trabalho propôs a construção de um protótipo para TV Digital Interativa que pudesse ser controlado em suas operações de ajuste de volume do som e de troca de canal de programação através do controle remoto e através de um conjunto de gestos, definidos a partir de um conjunto de regras da Engenharia da Usabilidade. O processo de definição do layout e do conjunto de gestos foi realizado com a participação de usuários, e, a partir dessas definições, as funcionalidades da TV construídas de modo a se ter na tela as informações sobre o volume do som e o canal de programação, além do processo de reconhecimento de gestos. O protótipo construído neste trabalho se diferencia de produtos comerciais, pois levou-se em consideração a relação entre o custo e o desempenho dos dispositivos utilizados, buscando oferecer uma opção acessível e mais flexível quanto ao seu uso em equipamentos diversos. A abordagem descrita neste trabalho trata dos desafios que foram enfrentados nas áreas de Engenharia da Usabilidade, TVDi e Visão Computacional tendo seu protótipo final comparado a outros métodos e produtos, mostrando um desempenho de aproximadamente 95% no acerto do gesto exibido e uma taxa de velocidade de 26 frames por segundo.

<p>Face detection is a classical application of object detection. There are many practical applications in which face detection is the first step; face recognition, video surveillance, image database management, video coding. </p><p>This report presents the results of an implementation of the AdaBoost algorithm to train a Strong Classifier to be used for face detection. The AdaBoost algorithm is fast and shows a low false detection rate, two characteristics which are important for face detection algorithms. </p><p>The application is an implementation of the AdaBoost algorithm with several command-line executables that support testing of the algorithm. The training and detection algorithms are separated from the rest of the application by a well defined interface to allow reuse as a software library. </p><p>The source code is documented using the JavaDoc-standard, and CppDoc is then used to produce detailed information on classes and relationships in html format. </p><p>The implemented algorithm is found to produce relatively high detection rate and low false alarm rate, considering the badly suited training data used.</p>

Face detection is a classical application of object detection. There are many practical applications in which face detection is the first step; face recognition, video surveillance, image database management, video coding. This report presents the results of an implementation of the AdaBoost algorithm to train a Strong Classifier to be used for face detection. The AdaBoost algorithm is fast and shows a low false detection rate, two characteristics which are important for face detection algorithms. The application is an implementation of the AdaBoost algorithm with several command-line executables that support testing of the algorithm. The training and detection algorithms are separated from the rest of the application by a well defined interface to allow reuse as a software library. The source code is documented using the JavaDoc-standard, and CppDoc is then used to produce detailed information on classes and relationships in html format. The implemented algorithm is found to produce relatively high detection rate and low false alarm rate, considering the badly suited training data used.