FundaÃÃo Cearense de Apoio ao Desenvolvimento Cientifico e TecnolÃgico / Uma alternativa econÃmica importante para a regiÃo semi-Ãrida do Brasil à a criaÃÃo de ovinos e caprinos. AlÃm de leite e carne de caprinos/ovinos, as peles sÃo muito apreciadas na fabricaÃÃo de artefatos finos (por exemplo, sapatos, bolsas, carteiras e casacos). No entanto, devido ao modo extensivo de criaÃÃo/reproduÃÃo e informalidade do abate, as peles de ovinos/caprinos sÃo entregues ao curtume com diferentes tipos e nÃveis de defeitos. Na indÃstria, trabalhadores especializados tÃm a tarefa de classificar/discriminar as peles de acordo com a qualidade das mesmas. Este trabalho à artesanal, demorado e extremamente dependente da experiÃncia do funcionÃrio responsÃvel pela discriminaÃÃo da qualidade da pele. O mesmo funcionÃrio pode produzir diferentes classificaÃÃes se ele/ela tiver que reclassificar o lote de pele. Assim, a fim de lidar com esses problemas, neste trabalho, apresentam-se os primeiros resultados de um sistema baseado em visÃo computacional cujo objetivo à classificar automaticamente a qualidade da pele de caprinos/ovinos. Para isso, comparamos os desempenhos de classificadores estatÃsticos e neurais utilizando diversas tÃcnicas de extraÃÃo de caracterÃsticas, tais como a VariÃncia das colunas (VAR), Transformada Wavelet de Haar (HAAR), FatoraÃÃo em Matrizes NÃo-Negativas (NMF), AnÃlise de Componentes Principais (PCA) e Matrizes de Co-ocorrÃncia de nÃveis de cinza (GLCM). TambÃm foram implementados mecanismos de opÃÃo de rejeiÃÃo nos classificadores avaliados. OpÃÃo de rejeiÃÃo à uma tÃcnica usada para aumentar a confiabilidade do classificador em sistemas de apoio à tomada de decisÃo, que consiste em reter a classificaÃÃo automÃtica de um item, caso a decisÃo nÃo seja considerada suficientemente confiÃvel. Jà com a utilizaÃÃo da opÃÃo de rejeiÃÃo, de uma forma geral, foi possÃvel observar uma considerÃvel melhora nas taxas de acerto dos classificadores avaliados, Ãs expensas de uma taxa de rejeiÃÃo relativamente alta. TambÃm foi possÃvel observar que, para os classificadores analisados, os extratores HAAR e GLCM foram menos sensÃveis à aplicaÃÃo da opÃÃo de rejeiÃÃo, em comparaÃÃo com os resultados obtidos para o caso sem opÃÃo de rejeiÃÃo.

This thesis investigates the computer vision application of live fish recognition, which is needed in application scenarios where manual annotation is too expensive, when there are too many underwater videos. This system can assist ecological surveillance research, e.g. computing fish population statistics in the open sea. Some pre-processing procedures are employed to improve the recognition accuracy, and then 69 types of features are extracted. These features are a combination of colour, shape and texture properties in different parts of the fish such as tail/head/top/bottom, as well as the whole fish. Then, we present a novel Balance-Guaranteed Optimized Tree with Reject option (BGOTR) for live fish recognition. It improves the normal hierarchical method by arranging more accurate classifications at a higher level and keeping the hierarchical tree balanced. BGOTR is automatically constructed based on inter-class similarities. We apply a Gaussian Mixture Model (GMM) and Bayes rule as a reject option after the hierarchical classification to evaluate the posterior probability of being a certain species to filter less confident decisions. This novel classification-rejection method cleans up decisions and rejects unknown classes. After constructing the tree architecture, a novel trajectory voting method is used to eliminate accumulated errors during hierarchical classification and, therefore, achieves better performance. The proposed BGOTR-based hierarchical classification method is applied to recognize the 15 major species of 24150 manually labelled fish images and to detect new species in an unrestricted natural environment recorded by underwater cameras in south Taiwan sea. It achieves significant improvements compared to the state-of-the-art techniques. Furthermore, the sequence of feature selection and constructing a multi-class SVM is investigated. We propose that an Individual Feature Selection (IFS) procedure can be directly exploited to the binary One-versus-One SVMs before assembling the full multiclass SVM. The IFS method selects different subsets of features for each Oneversus- One SVM inside the multiclass classifier so that each vote is optimized to discriminate the two specific classes. The proposed IFS method is tested on four different datasets comparing the performance and time cost. Experimental results demonstrate significant improvements compared to the normal Multiclass Feature Selection (MFS) method on all datasets.

Die Auswahl von Aktien ist ein Gebiet der Finanzanalyse, die von speziellem Interesse sowohl für viele professionelle Investoren als auch für Wissenschaftler ist. Empirische Untersuchungen belegen, dass Aktienerträge vorhergesagt werden können. Während verschiedene Modellierungstechniken zur Aktienselektion eingesetzt werden könnten, analysiert diese Arbeit die meist verbreiteten Methoden, darunter allgemeine Gleichgewichtsmodelle und Asset Pricing Modelle; parametrische, nichtparametrische und semiparametrische Regressionsmodelle; sowie beliebte Black-Box Klassifikationsmethoden. Aufgrund vorteilhafter Eigenschaften binärer Klassifikationsbäume, wie zum Beispiel einer herausragenden Interpretationsmöglichkeit von Entscheidungsregeln, wird der Kern des Handelsalgorithmus unter Verwendung dieser modernen, nichtparametrischen Methode konstruiert. Die optimale Größe des Baumes wird als der entscheidende Faktor für die Vorhersageperformance von Klassifikationsbäumen angesehen. Während eine Vielfalt alternativer populärer Bauminduktions- und Pruningtechniken existiert, die in dieser Studie kritisch gewürdigt werden, besteht eines der Hauptanliegen dieser Arbeit in einer neuartigen Methode asymmetrischen Baumprunings mit Abweisungsoption. Diese Methode wird als Best Node Selection (BNS) bezeichnet. Eine wichtige inverse Fortpflanzungseigenschaft der BNS wird bewiesen. Diese eröffnet eine einfache Möglichkeit, um die Suche der optimalen Baumgröße in der Praxis zu implementieren. Das traditionelle costcomplexity Pruning zeigt eine ähnliche Performance hinsichtlich der Baumgenauigkeit verglichen mit beliebten alternativen Techniken, und es stellt die Standard Pruningmethode für viele Anwendungen dar. Die BNS wird mit cost-complexity Pruning empirisch verglichen, indem zwei rekursive Portfolios aus DAX-Aktien zusammengestellt werden. Vorhersagen über die Performance für jede einzelne Aktie werden von Entscheidungsbäumen gemacht, die aktualisiert werden, sobald neue Marktinformationen erhältlich sind. Es wird gezeigt, dass die BNS der traditionellen Methode deutlich überlegen ist, und zwar sowohl gemäß den Backtesting Ergebnissen als auch nach dem Diebold-Marianto Test für statistische Signifikanz des Performanceunterschieds zwischen zwei Vorhersagemethoden. Ein weiteres neuartiges Charakteristikum dieser Arbeit liegt in der Verwendung individueller Entscheidungsregeln für jede einzelne Aktie im Unterschied zum traditionellen Zusammenfassen lernender Muster. Empirische Daten in Form individueller Entscheidungsregeln für einen zufällig ausgesuchten Zeitpunkt in der Überprüfungsreihe rechtfertigen diese Methode. / Stock picking is the field of financial analysis that is of particular interest for many professional investors and researchers. There is a lot of research evidence supporting the fact that stock returns can effectively be forecasted. While various modeling techniques could be employed for stock price prediction, a critical analysis of popular methods including general equilibrium and asset pricing models; parametric, non- and semiparametric regression models; and popular black box classification approaches is provided. Due to advantageous properties of binary classification trees including excellent level of interpretability of decision rules, the trading algorithm core is built employing this modern nonparametric method. Optimal tree size is believed to be the crucial factor of forecasting performance of classification trees. While there exists a set of widely adopted alternative tree induction and pruning techniques, which are critically examined in the study, one of the main contributions of this work is a novel methodology of nonsymmetrical tree pruning with reject option called Best Node Selection (BNS). An important inverse propagation property of BNS is proven that provides an easy way to implement the search for the optimal tree size in practice. Traditional cost-complexity pruning shows similar performance in terms of tree accuracy when assessed against popular alternative techniques, and it is the default pruning method for many applications. BNS is compared with costcomplexity pruning empirically by composing two recursive portfolios out of DAX30 stocks. Performance forecasts for each of the stocks are provided by constructed decision trees that are updated when new market information becomes available. It is shown that BNS clearly outperforms the traditional approach according to the backtesting results and the Diebold-Mariano test for statistical significance of the performance difference between two forecasting methods. Another novel feature of this work is the use of individual decision rules for each stock as opposed to pooling of learning samples, which is done traditionally. Empirical data in the form of provided individual decision rules for a randomly selected time point in the backtesting set justify this approach.