Dans cette thèse, nous nous sommes proposé d'explorer les contributions relatives des caractéristiques de haut et de bas niveau des stimuli dans la catégorisation d'objet. Ce travail comporte trois études, chez l'homme et le singe. L'originalité de cette thèse réside donc dans la construction des stimuli. Notre première étude a visé à caractériser les corrélats neuraux de la reconnaissance d'images chez le singe en ECoG. Pour cela nous avons développé un protocole de catégorisation où les stimuli étaient des séquences visuelles dans lesquelles les contours des objets (information sémantique, caractéristique de haut niveau) étaient modulés cycliquement grâce à la technique SWIFT (créée par Roger Koenig et Rufin VanRullen) alors que la luminance, les contrastes et les fréquences spatiales (caractéristiques de bas niveau) étaient conservées. Grâce à une analyse en potentiels évoqués, nous avons pu mettre en évidence une activité électrophysiologique tardive en " tout ou rien " spécifique de la reconnaissance de la cible de la tâche par le singe. Mais parce que les objets sont rarement isolés en conditions réelles, nous nous sommes penchés dans une deuxième étude sur l'effet de congruence contextuelle lors de la catégorisation d'objets chez l'homme et le singe. Nous avons comparé la contribution du spectre d'amplitude d'une transformée de Fourier à cet effet de congruence chez ces deux espèces. Nous avons révélé une divergence de stratégie, le singe semblant davantage sensible à ces caractéristiques de bas niveau que l'homme. Enfin dans une dernière étude nous avons tenté de quantifier l'effet de congruence sémantique multisensorielle dans une tâche de catégorisation audiovisuelle chez l'homme. Dans cette étude nous avons égalisé un maximum de paramètres de bas niveau dans les deux modalités sensorielles, que nous avons toujours stimulées conjointement. Dans le domaine visuel, nous avons réutilisé la technique SWIFT, et dans le domaine auditif nous avons utilisé une technique de randomisation de snippets. Nous avons pu alors constater un gain multisensoriel important pour les essais congruents (l'image et le son désignant le même objet), s'expliquant spécifiquement par le contenu sémantique des stimuli. Cette thèse ouvre donc de nouvelles perspectives, tant sur la cognition comparée entre homme et primate non humain que sur la nécessité de contrôler les caractéristiques physiques de stimuli utilisés dans les tâches de reconnaissance d'objets. / In this thesis, we explored the relative contributions of high level and low level features of stimuli used in object categorization tasks. This work consists of three studies in human and monkey. The originality of this thesis lies in stimuli construction. Our first study aimed to characterize neural correlates of image recognition in monkey, using ECoG recordings. For that purpose we developped a categorization task using SWIFT technique (technique created by Roger Koenig and Rufin VanRullen). Stimuli were visual sequences in which object contours (semantic content, high level feature) were cyclically modulated while luminance, contrasts and spatial frequencies (low level features) remained stable. By analyzing evoked potentials, we brought to light a late electrophysiological activity, in an " all or none " fashion, specifically related to the target recognition in monkey. But because in real condition objects are never isolated, we explored in a second study contextual congruency effect in visual categorization task in humans and monkeys. We compared the contribution of Fourier transform amplitude spectrum to this congruency effect in the both species. We found a strategy divergence showing that monkeys were more sensitive to the low level features of stimuli than humans. Finally, in the last study, we tried to quantify multisensory semantic congruency effect, during a audiovisual categorization task in humans. In that experiment, we equalized a maximum of low level features, in both sensory modalities which were always jointly stimulated. In the visual domain, we used again the SWIFT technique, whereas in auditory domain we used a snippets randomization technique. We highlighted a large multisensory gain in congruent trials (i.e. image and sound related to the same object), specifically linked to the semantic content of stimuli. This thesis offers new perspectives both for comparative cognition between human and non human primates and for the importance of controlling the physical features of stimuli used in object recognition tasks.

Face recognition that is invariant to pose and illumination is a problem solved effortlessly by the human brain, but the computational details that underlie such efficient recognition are still far from clear. This thesis draws on research from psychology and neuroscience about face and object recognition and the visual system in order to develop a novel computational method for face detection, feature selection and representation, and memory structure for recall. A biologically plausible framework for developing a face recognition system will be presented. This framework can be divided into four parts: 1) A face detection system. This is an improved version of a biologically inspired feedforward neural network that has modifiable connections and reflects the hierarchical and elastic structure of the visual system. The face detection system can detect if a face is present in an input image, and determine the region which contains that face. The system is also capable of detecting the pose of the face. 2) A face region selection mechanism. This mechanism is used to determine the Gabor-style features corresponding to the detected face, i.e., the features from the region of interest. This region of interest is selected using a feedback mechanism that connects the higher level layer of the feedforward neural network where ultimately the face is detected to an intermediate level where the Gabor style features are detected. 3) A face recognition system which is based on the binary encoding of the Gabor style features selected to represent a face. Two alternative coding schemes are presented, using 2 and 4 bits to represent a winning orientation at each location. The effectiveness of the Gabor-style features and the different coding schemes in discriminating faces from different classes is evaluated using the Yale B Face Database. The results from this evaluation show that this representation is close to other results on the same database. 4) A theoretical approach for a memory system capable of memorising sequences of poses. A basic network for memorisation and recall of sequences of labels have been implemented, and from this it is extrapolated a memory model that could use the ability of this model to memorise and recall sequences, to assist in the recognition of faces by memorising sequences of poses. Finally, the capabilities of the detection and recognition parts of the system are demonstrated using a demo application that can learn and recognise faces from a webcam.

Fundação de Amparo a Pesquisa do Estado de Minas Gerais / Image representation in Content Based Image Retrieval systems is a fundamental task. The results obtained by these systems strongly depend on the choice of features selected to represent an image. Works in the literature show that intelligent techniques are used to minimize the semantic gap between the limited power of machine interpretation and human subjectivity. In this work the use of artificial neural networks to characterize images in a high-level space from an initial characterization based on low-level features (color, shape and texture) is proposed. Experiments on 3 databases of various kinds, one with general images (BD-12750 ), one with texture images (Vistex-167 ) and other with buildings (ZuBuD) are performed to exemplify the application of the method and to show the effectiveness of the model. Furthermore, the application of the proposed method in the high-level characterization of complex motions patterns is presented. / Em sistemas de Recuperação de Imagens Baseada em Conteúdo a representação das imagens desempenham um papel fundamental. Os resultados obtidos por esses sistemas dependem fortemente da escolha das características selecionadas para representar uma imagem. Trabalhos existentes na literatura evidenciam que técnicas inteligentes conseguem minimizar o gap- semântico existente entre o poder de interpretação limitado das máquinas e a subjetividade humana. Neste trabalho é proposto o uso das redes neurais artificiais para caracterizar imagens neurosemânticamente à partir de uma caracterização inicial baseada em características de baixo nível (cor, forma e textura). Testes em 3 bases de dados de naturezas diferentes, um de imagens mais gerais (BD-12750 ), um de texturas (Vistex-167 ) e outro de prédios (ZuBuD) exemplificam a aplicação do método como também mostram a eficácia do modelo. Ainda é apresentada a aplicação do método proposto na caracterização neurosemântica de movimentos complexos em vídeos. / Mestre em Ciência da Computação

This thesis is focused on knowledge discovery from databases, especially on methods of classification and prediction. These methods are described in detail.  Furthermore, this work deals with multimedia databases and the way these databases store data. In particular, the method for processing low-level image and video data is described.  The practical part of the thesis focuses on the implementation of this GMM method used for extracting low-level features of video data and images. In other parts, input data and tools, which the implemented method was compared with, are described.  The last section focuses on experiments comparing extraction efficiency features of high-level attributes of low-level data and the methods implemented in selected classification tools LibSVM.

This master's thesis deals with knowledge discovery in databases, especially basic methods of classification and prediction used for data mining are described here. The next chapter contains introduction to multimedia databases and knowledge discovery in multimedia databases. The main goal of this chapter was to focus on extraction of low level features from video data and images. In the next parts of this work, there is described data set and results of experiments in applications RapidMiner, LibSVM and own developed application. The last chapter summarises results of used methods for high level feature extraction from low level description of data.

This master's thesis is dedicated to theme of knowledge discovery in Multimedia Databases, especially basic methods of classification and prediction used for data mining. The other part described about extraction of low level features from video data and images and summarizes information about content-based search in multimedia content and indexing this type of data. Final part is dedicated to implementation Gaussian mixtures model for classification and compare the final result with other method SVM.