The Cultural Capital of Europe project aims to highlight and promote culture of a specific region in Europe during a calendar year. Umeå was chosen to become the Capital of Culture in 2014 and highlight the northern corner of Europe. A new “Open source” approach with the objective to increase participation and co-creation of different actors in the cultural capital year was introduced. This study examines media descriptions of the Samis and alternative cultural groups’ participation in the Cultural Capital year. Both quantitative and qualitative methodological tools were used to analyse descriptions of these groups’ in digital articles from local newspapers. The findings indicate of widespread discontent with how these groups were represented in the Cultural Capital year. The Samis were described as being excluded from planning but included in the implementation of the programme. The alternative cultural groups were instead described as being included in the planning but excluded from participation in the programme year. The relations between these groups and the organisers were analysed in terms of hegemony. The results indicate the relations being in balance at the beginning of the year but started to weaken at an early stage.

Le tri automatique des déchets est un sujetcomplexe en raison de la diversité des objets et desmatériaux présents. Il nécessite un apport de donnéesvariées et hétérogènes. Cette thèse traite du problème defusion de données découlant d’un dispositif de troiscapteurs dont une caméra hyperspectrale dans ledomaine NIR. Nous avons étudié l’avantage d’utiliser lecadre des fonctions de croyance (BFT) tout au long de ladémarche de fusion en utilisant notamment la mesure deconflit comme un critère clé de notre approche. Dans unepremière partie, nous avons étudié l'intérêt de la BFTpour la classification multiclasse des donnéeshyperspectrales à partir d’Error Correcting OutputCodes (ECOC) qui consistent à séparer le problèmemulticlasse en un ensemble de sous-problèmes binairesplus simples à résoudre. Les questions de commentidéalement séparer le problème multiclasse (codage)ainsi que celle de la combinaison des réponses de cesproblèmes binaires (décodage) sont encore aujourd’huides questions ouvertes. Le cadre des fonctions decroyance permet de proposer une étape de décodage quimodélise chaque classifieur binaire comme une sourceindividuelle d'information grâce notamment à lamanipulation des hypothèses composées. Par ailleurs laBFT fournit des indices pour détecter les décisions peufiables ce qui permet une auto-évaluation de la méthoderéalisée sans vérité terrain. Dans une deuxième partietraitant de la fusion de données, nous proposons unedémarche ‘orientée-objet’ composée d’un module desegmentation et d’un module de classification afin defaire face aux problèmes d’échelle, de différences derésolutions et de recalage des capteurs. L’objectif estalors d’estimer une segmentation où les segmentscoïncident avec les objets individuels et sont labellisés entermes de matériau. Nous proposons une interactionentre les modules à base de validation mutuelle. Ainsi,d’une part la fiabilité de la labellisation est évaluée auniveau des segments, d’autre part l’information declassification interagit sur les segments initiaux pour serapprocher d’une segmentation au niveau « objet » : leconsensus (ou l’absence de consensus) parmi lesinformations de classification au sein d’un segment ouentre segments connexes permet de faire évoluer lesupport spatial vers le niveau objet. / Automatic waste sorting is a complex matterbecause of the diversity of the objects and of the presentmaterials. It requires input from various andheterogeneous data. This PhD work deals with the datafusion problem derived from an acquisition devicecomposed of three sensors, including an hyperspectralsensor in the NIR field. We first studied the benefit ofusing the belief function theory framework (BFT)throughout the fusion approach, using in particularconflict measures to drive the process. We first studiedthe BFT in the multiclass classification problem createdby hyperspectral data. We used the Error CorrectingOutput Codes (ECOC) framework which consists inseparating the multiclass problem into several binaryones, simpler to solve. The questions of the idealdecomposition of the multiclass problem (coding) and ofthe answer combination coming from the binaryclassifiers (decoding) are still open-ended questions. Thebelief function framework allows us to propose adecoding step modelling each binary classifier as anindividual source of information, thanks to the possibilityof handling compound hypotheses. Besides, the BFTprovides indices to detect non reliable decisions whichallow for an auto-evaluation of the method performedwithout using any ground truth. In a second part dealingwith the data fusion,we propose an evidential version ofan object-based approach composed with a segmentationmodule and a classification module in order to tackle theproblems of the differences in scale, resolutions orregistrations of the sensors. The objective is then toestimate a relevant spatial support corresponding to theobjects while labelling them in terms of material. Weproposed an interactive approach with cooperationbetween the two modules in a cross-validation kind ofway. This way, the reliability of the labelling isevaluated at the segment level, while the classificationinformation acts on the initial segments in order toevolve towards an object level segmentation: consensusamong the classification information within a segment orbetween adjacent regions allow the spatial support toprogressively reach object level

<p>The computer vision task of classifying natural images is a primary driving force behind modern AI algorithms. Deep Convolutional Neural Networks (CNNs) demonstrate state of the art performance in large scale multi-class image classification tasks. However, due to the many layers and millions of parameters these models are considered to be black box algorithms. The decisions of these models are further obscured due to a cumbersome multi-class decision process. There exists another approach called class binarization in the literature which determines the multi-class prediction outcome through a sequence of binary decisions.The focus of this dissertation is on the integration of the class-binarization approach to multi-class classification with deep learning models, such as CNNs, for addressing large scale image classification problems. Three works are presented to address the integration.</p> <p>In the first work, Error Correcting Output Codes (ECOCs) are integrated into CNNs by inserting a latent-binarization layer prior to the CNNs final classification layer.  This approach encapsulates both encoding and decoding steps of ECOC into a single CNN architecture. EM and Gibbs sampling algorithms are combined with back-propagation to train CNN models with Latent Binarization (LB-CNN). The training process of LB-CNN guides the model to discover hidden relationships similar to the semantic relationships known apriori between the categories. The proposed models and algorithms are applied to several image recognition tasks, producing excellent results.</p> <p>In the second work, Hierarchically Decodeable Output Codes (HD-OCs) are proposedto compactly describe a hierarchical probabilistic binary decision process model over the features of a CNN. HD-OCs enforce more homogeneous assignments of the categories to the dichotomy labels. A novel concept called average decision depth is presented to quantify the average number of binary questions needed to classify an input. An HD-OC is trained using a hierarchical log-likelihood loss that is empirically shown to orient the output of the latent feature space to resemble the hierarchical structure described by the HD-OC. Experiments are conducted at several different scales of category labels. The experiments demonstrate strong performance and powerful insights into the decision process of the model.</p> <p>In the final work, the literature of enumerative combinatorics and partially ordered sets isused to establish a unifying framework of class-binarization methods under the Multivariate Bernoulli family of models. The unifying framework theoretically establishes simple relationships for transitioning between the different binarization approaches. Such relationships provide useful investigative tools for the discovery of statistical dependencies between large groups of categories. They are additionally useful for incorporating taxonomic information as well as enforcing structural model constraints. The unifying framework lays the groundwork for future theoretical and methodological work in addressing the fundamental issues of large scale multi-class classification.</p> <p><br></p>