O sinal de entrada na retina é decomposto em termos de frequência espacial (FE), variações periódicas de luminância ao longo do espaço. Existe vasta literatura sobre o processamento de FE no córtex visual primário. No entanto, não se sabe ao certo como esta informação sensorial básica é processada e integrada numa visão de alto nível. Esta tese aborda este tema ao investigar lateralidade cerebral, tempo de processamento e contexto cognitivo em três diferentes seções com objetivos específicos. Estas seções investigaram comportamentalmente visão de alto nível tendo a face humana como estímulo, dado sua relevância biológica e social. Na primeira seção (Theoretical Review), uma revisão apresenta estudos clínicos e neuropsicológicos que mostram áreas cerebrais envolvidas na percepção de faces e como os hemisférios esquerdo e direito realizam um processamento holístico e analítico baseado em informações de FEs. A especialização hemisférica de FE no reconhecimento de faces é então revisada e discutida. Concluiu-se que assimetrias sensoriais podem ser a base para assimetrias cognitivas de alta ordem. Ademais, foi destacado a influência do tempo de processamento. Na segunda seção (Study 1), foi investigado por método psicofísico a lateralidade de baixas e altas FEs no reconhecimento de faces em diferentes tempos de exposição. Faces com filtragem de FE foram apresentadas em campo visual dividido em alta e baixa restrição temporal em duas tarefas: reconhecimento facial (Experimento 1) e reconhecimento do sexo facial (Experimento 2). No Experimento 1, informações faciais de baixas e altas FEs foram mais eficientemente processadas no hemisfério direito e esquerdo, respectivamente, sem efeito do tempo de exposição das faces. Os resultados do Experimento 2 mostraram uma assimetria do hemisfério direito para baixas FEs em baixa restrição temporal. Conclui-se que o processamento de altas e baixas FEs é lateralizado nos hemisférios cerebrais no reconhecimento de faces. No entanto, a contribuição de altas e baixas FEs é dependente da tarefa e do tempo de exposição. Na terceira seção (Study 2) foi investigado qual estratégia temporal, coarse-to-fine (de baixas para altas FEs) ou fine-to-coarse, cada hemisfério cerebral utiliza para integrar informação de FE de faces humanas numa tarefa de categorização facial homem-mulher. Sequências dinâmicas breves coarse-to-fine e fine-to-coarse de faces foram apresentadas no campo visual esquerdo, direito e central. Os resultados do tempo de resposta e do score de eficiência invertida mostraram uma prevalência geral de um processamento coarse-to-fine, independente do campo visual de apresentação. Ainda, os dados da taxa de erro ressaltam o processamento coarse-to-fine realizado pelo hemisfério direito. No geral, esta tese fornece insights sobre assimetria cerebral funcional, integração de alto nível e curso temporal do processamento de FEs, principalmente para aqueles interessados na percepção de faces. Também foi mostrado que operações lateralizadas, tarefa-dependente e coarse-to-fine podem coexistir e interagir no cérebro para processar informação de FE. / Retinal input is decomposed in terms of spatial frequency (SF), i.e., periodic variations of luminance through space. There is extensive literature on the processing of SF in the primary visual cortex. However, it is still unclear how SF information is processed and integrated in high-level vision. This thesis addressed this issue in terms of laterality effects, processing time-course, and the cognitive context in three different sections with specific purposes. These sections behaviorally tackle high-level vision using human faces as stimuli due to their biological and social relevance. In the first section (Theoretical Review) a literature review presented clinical and neurophysiological studies that show brain areas that are involved in face perception and how the right and left hemispheres perform holistic and analytic processing, depending on SF information. The SF hemispheric specialization in face recognition is then reviewed and discussed. Our conclusion is that functional sensorial asymmetries may be the basis for high-level cognitive asymmetries. In addition, we highlighted the role of the processing time. In the second section (Study 1), we psychophysically investigated laterality of low and high SF in face recognition at different exposure times. The SF filtered faces were presented in a divided visual field at high and low temporal constraint in two tasks: face recognition (Experiment 1) and face gender recognition (Experiment 2). In Experiment 1, low and high SF facial information were more efficiently processed in the right and in the left hemisphere, respectively, with no effect of exposure time. In Experiment 2, results showed a right hemisphere asymmetry for low SF faces at low temporal constraint. We concluded that the processing of low and high SF is lateralized in the brain hemispheres for face recognition. However, low and high SF contribution is dependent on the task and the exposure time. In the third section (Study 2), we aimed to investigate which temporal strategy, i.e., coarse-to-fine (from low to high SF) or fine-to-course, each brain hemisphere performs to integrate SF information of human faces in a male-female categorization task. Coarse-to-fine and fine-to-course brief dynamic sequences of faces were presented in the left, right and central visual field. Results of the correct response time and the inverse efficiency score showed an overall advantage of coarse-to-fine processing, irrespective of the visual field of presentation. Data of the error rate also highlights the role of the right hemisphere in the coarse-to-fine processing. All in all, this thesis provided some insights on functional brain asymmetry, high-level integration, and processing time-course of SF information, mainly for those interested in face perception. It was also shown that lateralized, diagnostic-oriented, and coarse-to-fine operations may coexist and interact in the human brain to process SF information.

O sinal de entrada na retina é decomposto em termos de frequência espacial (FE), variações periódicas de luminância ao longo do espaço. Existe vasta literatura sobre o processamento de FE no córtex visual primário. No entanto, não se sabe ao certo como esta informação sensorial básica é processada e integrada numa visão de alto nível. Esta tese aborda este tema ao investigar lateralidade cerebral, tempo de processamento e contexto cognitivo em três diferentes seções com objetivos específicos. Estas seções investigaram comportamentalmente visão de alto nível tendo a face humana como estímulo, dado sua relevância biológica e social. Na primeira seção (Theoretical Review), uma revisão apresenta estudos clínicos e neuropsicológicos que mostram áreas cerebrais envolvidas na percepção de faces e como os hemisférios esquerdo e direito realizam um processamento holístico e analítico baseado em informações de FEs. A especialização hemisférica de FE no reconhecimento de faces é então revisada e discutida. Concluiu-se que assimetrias sensoriais podem ser a base para assimetrias cognitivas de alta ordem. Ademais, foi destacado a influência do tempo de processamento. Na segunda seção (Study 1), foi investigado por método psicofísico a lateralidade de baixas e altas FEs no reconhecimento de faces em diferentes tempos de exposição. Faces com filtragem de FE foram apresentadas em campo visual dividido em alta e baixa restrição temporal em duas tarefas: reconhecimento facial (Experimento 1) e reconhecimento do sexo facial (Experimento 2). No Experimento 1, informações faciais de baixas e altas FEs foram mais eficientemente processadas no hemisfério direito e esquerdo, respectivamente, sem efeito do tempo de exposição das faces. Os resultados do Experimento 2 mostraram uma assimetria do hemisfério direito para baixas FEs em baixa restrição temporal. Conclui-se que o processamento de altas e baixas FEs é lateralizado nos hemisférios cerebrais no reconhecimento de faces. No entanto, a contribuição de altas e baixas FEs é dependente da tarefa e do tempo de exposição. Na terceira seção (Study 2) foi investigado qual estratégia temporal, coarse-to-fine (de baixas para altas FEs) ou fine-to-coarse, cada hemisfério cerebral utiliza para integrar informação de FE de faces humanas numa tarefa de categorização facial homem-mulher. Sequências dinâmicas breves coarse-to-fine e fine-to-coarse de faces foram apresentadas no campo visual esquerdo, direito e central. Os resultados do tempo de resposta e do score de eficiência invertida mostraram uma prevalência geral de um processamento coarse-to-fine, independente do campo visual de apresentação. Ainda, os dados da taxa de erro ressaltam o processamento coarse-to-fine realizado pelo hemisfério direito. No geral, esta tese fornece insights sobre assimetria cerebral funcional, integração de alto nível e curso temporal do processamento de FEs, principalmente para aqueles interessados na percepção de faces. Também foi mostrado que operações lateralizadas, tarefa-dependente e coarse-to-fine podem coexistir e interagir no cérebro para processar informação de FE. / Retinal input is decomposed in terms of spatial frequency (SF), i.e., periodic variations of luminance through space. There is extensive literature on the processing of SF in the primary visual cortex. However, it is still unclear how SF information is processed and integrated in high-level vision. This thesis addressed this issue in terms of laterality effects, processing time-course, and the cognitive context in three different sections with specific purposes. These sections behaviorally tackle high-level vision using human faces as stimuli due to their biological and social relevance. In the first section (Theoretical Review) a literature review presented clinical and neurophysiological studies that show brain areas that are involved in face perception and how the right and left hemispheres perform holistic and analytic processing, depending on SF information. The SF hemispheric specialization in face recognition is then reviewed and discussed. Our conclusion is that functional sensorial asymmetries may be the basis for high-level cognitive asymmetries. In addition, we highlighted the role of the processing time. In the second section (Study 1), we psychophysically investigated laterality of low and high SF in face recognition at different exposure times. The SF filtered faces were presented in a divided visual field at high and low temporal constraint in two tasks: face recognition (Experiment 1) and face gender recognition (Experiment 2). In Experiment 1, low and high SF facial information were more efficiently processed in the right and in the left hemisphere, respectively, with no effect of exposure time. In Experiment 2, results showed a right hemisphere asymmetry for low SF faces at low temporal constraint. We concluded that the processing of low and high SF is lateralized in the brain hemispheres for face recognition. However, low and high SF contribution is dependent on the task and the exposure time. In the third section (Study 2), we aimed to investigate which temporal strategy, i.e., coarse-to-fine (from low to high SF) or fine-to-course, each brain hemisphere performs to integrate SF information of human faces in a male-female categorization task. Coarse-to-fine and fine-to-course brief dynamic sequences of faces were presented in the left, right and central visual field. Results of the correct response time and the inverse efficiency score showed an overall advantage of coarse-to-fine processing, irrespective of the visual field of presentation. Data of the error rate also highlights the role of the right hemisphere in the coarse-to-fine processing. All in all, this thesis provided some insights on functional brain asymmetry, high-level integration, and processing time-course of SF information, mainly for those interested in face perception. It was also shown that lateralized, diagnostic-oriented, and coarse-to-fine operations may coexist and interact in the human brain to process SF information.

L'analyse visuelle de scènes débute par l'extraction en parallèle de l'information à différentes fréquences spatiales, en suivant un mode de traitement par défaut « coarse-to-fine ». L'analyse rapide de l'information grossière (« coarse ») en basses fréquences spatiales fournirait un aperçu global de la scène, qui serait ensuite affiné par l'analyse des détails de la scène (« fine ») en hautes fréquences spatiales. L'objectif de cette thèse a été de préciser les bases cérébrales du traitement des fréquences spatiales lors de la catégorisation de scènes. A travers deux études comportementales, nous avons tout d'abord montré qu'une analyse « coarse-to-fine » est plus avantageuse pour la catégorisation rapide de scènes, et ce, indépendamment de la valeur de contraste de luminance associée aux différentes fréquences spatiales (Expériences 1 et 2). Des études en IRMf nous ont par la suite permis de mettre en évidence l'implication d'un large réseau cérébral lors de l'analyse « coarse-to-fine » des scènes, incluant les aires visuelles primaires et occipito-temporales, mais également le cortex frontal inférieur (Expérience 3). Une analyse de la connectivité a révélé que lors de cette analyse, le cortex frontal inférieur exercerait une influence « top-down » sur le cortex visuel primaire et les gyri fusiforme et parahippocampique au sein du cortex occipito-temporal. Ces résultats soulignent le rôle du cortex visuel primaire comme région intégrative, codant à la fois les afférences rétino-thalamiques et les influences « top-down » de régions supérieures. Nous avons également observé que le gyrus frontal inférieur et le gyrus fusiforme participaient activement à l'intégration de l'information sémantique contenue dans les basses et hautes fréquences spatiales d'une scène (Expérience 4). Enfin, nous avons spécifiquement étudié le traitement des fréquences spatiales au sein de régions occipito-temporales sélectives aux scènes : la « parahippocampal place area » (PPA), le cortex retrosplenial et l'« occipital place area ». Nous avons montré que ces trois régions participent de façon distincte au traitement des fréquences spatiales dans les scènes (Expérience 5) et qu'une stratégie d'analyse « coarse-to-fine » serait privilégiée par la PPA (Expérience 6). Les résultats de ces travaux nous permettent de conforter et de préciser les modèles actuels de la catégorisation visuelle de scènes basés sur un traitement fréquentiel de l'information visuelle. / Visual analysis begins with the parallel extraction of different attributes at different spatial frequencies following a predominantly coarse-to-fine default processing sequence. Rapid processing of low spatial frequency information would permit a coarse parsing of the visual input, prior to the detailed analysis of fine information in high spatial frequencies. Our aim was to further address the neural bases of spatial frequency processing during scene categorization. We first demonstrated in two behavioral studies that a coarse-to-fine processing is indeed an advantageous strategy for rapid scene categorization, and is independent of the luminance contrast values associated with the different spatial frequencies (Experiments 1 and 2). In two fMRI studies, we showed first the involvement of a large cerebral network during coarse-to-fine processing of scenes, including early visual and occipito-temporal areas, but also the inferior frontal cortex. Effective connectivity analysis revealed that the inferior frontal gyrus exerts top-down influence on the early visual cortex as well as on the parahippocampal and fusiform gyri in the occipito-temporal cortex (Experiment 3). These results highlight the role of the primary visual cortex in integrating top-down influences from frontal areas to retino-thalamic incoming signals. We also evidenced that the inferior frontal and fusiform gyri actively participate to the integration of the semantic information contained in low and high spatial frequency (Experiment 4). Finally, we specifically investigated the spatial frequency processing of scenes within scene-selective areas of the occipito-temporal cortex: the parahippocampal place area (PPA), the retrosplenial cortex, and the occipital place area. We demonstrated that these regions participate differently in the spatial frequency processing of scenes (Experiment 5) and that a coarse-to-fine processing is favored within the PPA (Experiment 6). Overall, results allow us to refine current model of visual scene categorization based on a spatial frequency analysis.

In Statistical Machine Translation (SMT), inference needs to be performed over a high-complexity discrete distribution de ned by the intersection between a translation hypergraph and a target language model. This distribution is too complex to be represented exactly and one typically resorts to approximation techniques either to perform optimisation { the task of searching for the optimum translation { or sampling { the task of nding a subset of translations that is statistically representative of the goal distribution. Beam-search is an example of an approximate optimisation technique, where maximisation is performed over a heuristically pruned representation of the goal distribution. For inference tasks other than optimisation, rather than nding a single optimum, one is really interested in obtaining a set of probabilistic samples from the distribution. This is the case in training where one wishes to obtain unbiased estimates of expectations in order to t the parameters of a model. Samples are also necessary in consensus decoding where one chooses from a sample of likely translations the one that minimises a loss function. Due to the additional computational challenges posed by sampling, n-best lists, a by-product of optimisation, are typically used as a biased approximation to true probabilistic samples. A more direct procedure is to attempt to directly draw samples from the underlying distribution rather than rely on n-best list approximations. Markov Chain Monte Carlo (MCMC) methods, such as Gibbs sampling, o er a way to overcome the tractability issues in sampling, however their convergence properties are hard to assess. That is, it is di cult to know when, if ever, an MCMC sampler is producing samples that are compatible iii with the goal distribution. Rejection sampling, a Monte Carlo (MC) method, is more fundamental and natural, it o ers strong guarantees, such as unbiased samples, but is typically hard to design for distributions of the kind addressed in SMT, rendering an intractable method. A recent technique that stresses a uni ed view between the two types of inference tasks discussed here | optimisation and sampling | is the OS approach. OS can be seen as a cross between Adaptive Rejection Sampling (an MC method) and A optimisation. In this view the intractable goal distribution is upperbounded by a simpler (thus tractable) proxy distribution, which is then incrementally re ned to be closer to the goal until the maximum is found, or until the sampling performance exceeds a certain level. This thesis introduces an approach to exact optimisation and exact sampling in SMT by addressing the tractability issues associated with the intersection between the translation hypergraph and the language model. The two forms of inference are handled in a uni ed framework based on the OS approach. In short, an intractable goal distribution, over which one wishes to perform inference, is upperbounded by tractable proposal distributions. A proposal represents a relaxed version of the complete space of weighted translation derivations, where relaxation happens with respect to the incorporation of the language model. These proposals give an optimistic view on the true model and allow for easier and faster search using standard dynamic programming techniques. In the OS approach, such proposals are used to perform a form of adaptive rejection sampling. In rejection sampling, samples are drawn from a proposal distribution and accepted or rejected as a function of the mismatch between the proposal and the goal. The technique is adaptive in that rejected samples are used to motivate a re nement of the upperbound proposal that brings it closer to the goal, improving the rate of acceptance. Optimisation can be connected to an extreme form of sampling, thus the framework introduced here suits both exact optimisation and exact iv sampling. Exact optimisation means that the global maximum is found with a certi cate of optimality. Exact sampling means that unbiased samples are independently drawn from the goal distribution. We show that by using this approach exact inference is feasible using only a fraction of the time and space that would be required by a full intersection, without recourse to pruning techniques that only provide approximate solutions. We also show that the vast majority of the entries (n-grams) in a language model can be summarised by shorter and optimistic entries. This means that the computational complexity of our approach is less sensitive to the order of the language model distribution than a full intersection would be. Particularly in the case of sampling, we show that it is possible to draw exact samples compatible with distributions which incorporate a high-order language model component from proxy distributions that are much simpler. In this thesis, exact inference is performed in the context of both hierarchical and phrase-based models of translation, the latter characterising a problem that is NP-complete in nature.

Thesis (MScEng)--Stellenbosch University, 2013. / ENGLISH ABSTRACT: There is a need for a coarse-to-fine positioning system as per a case study presented by the project collaboration partner, the Technical University of Chemnitz. The case study involves the picking and placing of piezo-ceramic micro parts into milled micro cavities. The focus of the project is the creation and development of a systematic approach for the design and the implementation of a coarse-to-fine positioning system for micro material handling. A second focus is to determine the applicability of the system for highly accurate and repeatable micro drilling and micro-milling. A systematic approach entails combining innovation management (assists in overall project structure), systems engineering (assists in specific design steps and tools) and research questions. Micro-milling was achieved, however the system proved unsuitable for highly accurate and repeatable micro drilling. The coarse-to-fine positioning system was successfully designed, built, and tested for accurate micro material handling. / AFRIKAANSE OPSOMMING: Daar bestaan ’n behoefte aan ’n grof-tot-fyn-posisioneringstelsel - soos blyk uit die gevallestudie uiteengesit deur die samewerkende projekvennoot, die Tegniese Universiteit van Chemnitz. Die gevallestudie behels die uitsoek en plasing van piezo-keramiek partikels in gefreesde mikroholtes. Hierdie projek het gefokus op die skepping en ontwikkeling van ’n stelselmatige benadering tot die ontwerp en implementering van ’n grof-tot-fyn-posisioneringstelsel vir mikromateriaalhantering en mikromasjienering. ’n Stelselmatige benadering behels dat innovasiebestuur (hulp met die algehele projekstruktuur), stelselingenieurswese (hulp met spesifieke ontwerpstappe en -hulpmiddels) en navorsingsdoelwitte gekombineer word. Die geïmplementeerde stelsel is eksperimenteel getoets en daar is bevind dat dit aan die spesifikasies en vereistes voldoen.

No description available.

The objective of this master’s thesis work is to evaluate the potential benefit of a superpixel preprocessing step for general object detection in a traffic environment. The various effects of different superpixel parameters on object detection performance, as well as the benefit of including depth information when generating the superpixels are investigated. In this work, three superpixel algorithms are implemented and compared, including a proposal for an improved version of the popular Spectral Linear Iterative Clustering superpixel algorithm (SLIC). The proposed improved algorithm utilises a coarse-to-fine approach which outperforms the original SLIC for high-resolution images. An object detection algorithm is also implemented and evaluated. The algorithm makes use of depth information obtained by a stereo camera to extract superpixels corresponding to foreground objects in the image. Hierarchical clustering is then applied, with the segments formed by the clustered superpixels indicating potential objects in the input image. The object detection algorithm managed to detect on average 58% of the objects present in the chosen dataset. It performed especially well for detecting pedestrians or other objects close to the car. Altering the density distribution of the superpixels in the image yielded an increase in detection rate, and could be achieved both with or without utilising depth information. It was also shown that the use of superpixels greatly reduces the amount of computations needed for the algorithm, indicating that a real-time implementation is feasible.

Le but des systèmes de recherche d'images est de diriger rapidement l'utilisateur vers des contenus qui sont pertinents par rapport à la requête qu'il a formulée. Après une présentation de la problématique et un état d'art du domaine, cette thèse présente nos contributions dans le cadre de l'apprentissage avec très peu d'exemples qui est propre à l'imagerie satellitaire. Ces contributions se situent principalement autour de l'utilisation de méthodes semi-supervisées pour exploiter l'information contenue dans les données non-labellisées et pallier en quelque sorte la faiblesse et la non-exhaustivité des bases d'apprentissage. Nous présentons deux scénarios d'utilisation de méthodes semi-supervisées. Le premier se place dans le cadre d'un système d'annotation automatique d'images. Le but est alors de détecter les structures inconnues, c'est à dire les ensembles cohérents de données qui ne sont pas représentées dans la base d'apprentissage et ainsi de guider l'utilisateur dans son exploration de la base. Le second scénario concerne les systèmes de recherche interactive d'images. L'idée est d'exploiter une structuration des données, sous la forme d'un clustering par exemple, pour accélérer l'apprentissage (i.e. minimiser le nombre d'itérations de feedback) dans le cadre d'un système avec boucle de pertinence. La nouveauté de nos contributions se situe autour du fait que la plupart des méthodes semi-supervisées ne permettent pas de travailler avec de gros volumes de données comme on en rencontre en imagerie satellitaire ou alors ne sont pas temps-réel ce qui est problématique dans un système avec retour de pertinence où la fluidité des interactions avec l'utilisateur est à privilégier. Un autre problème qui justifie nos contributions est le fait que la plupart des méthodes semi-supervisées font l'hypothèse que la distribution des données labellisées suit la distribution des données non labellisées, hypothèse qui n'est pas vérifiée dans notre cas du fait de la non-exhaustivité des bases d'apprentissage et donc de l'existence de structures inconnues au niveau des données non labellisées. La dernière partie de cette thèse concerne un système de recherche d'objets à l'intérieur d'un schéma de type apprentissage actif. Une stratégie de type "coarse-to-fine" est introduite pour autoriser l'analyse de la base d'images à une taille de patch beaucoup plus "fine" tout en maintenant un nombre raisonnable d'évaluations de la fonction de décision du classificateur utilisé à chaque itération de la boucle d'apprentissage actif. L'idée est d' élaguer de grandes parties de la base de données à une échelle d'analyse dite "grossière'', afin de réserver un traitement plus complexe et plus coûteux sur des zones restreintes et plus prometteuses des images.

No description available.