Recognition of Facial Expressions with Autoencoders and Convolutional-Nets

Les humains communiquent via différents types de canaux: les mots, la voix, les gestes du corps, des émotions, etc. Pour cette raison, un ordinateur doit percevoir ces divers canaux de communication pour pouvoir interagir intelligemment avec les humains, par exemple en faisant usage de microphones et de webcams. Dans cette thèse, nous nous intéressons à déterminer les émotions humaines à partir d’images ou de vidéo de visages afin d’ensuite utiliser ces informations dans différents domaines d’applications. Ce mémoire débute par une brève introduction à l'apprentissage machine en s’attardant aux modèles et algorithmes que nous avons utilisés tels que les perceptrons multicouches, réseaux de neurones à convolution et autoencodeurs. Elle présente ensuite les résultats de l'application de ces modèles sur plusieurs ensembles de données d'expressions et émotions faciales. Nous nous concentrons sur l'étude des différents types d’autoencodeurs (autoencodeur débruitant, autoencodeur contractant, etc) afin de révéler certaines de leurs limitations, comme la possibilité d'obtenir de la coadaptation entre les filtres ou encore d’obtenir une courbe spectrale trop lisse, et étudions de nouvelles idées pour répondre à ces problèmes. Nous proposons également une nouvelle approche pour surmonter une limite des autoencodeurs traditionnellement entrainés de façon purement non-supervisée, c'est-à-dire sans utiliser aucune connaissance de la tâche que nous voulons finalement résoudre (comme la prévision des étiquettes de classe) en développant un nouveau critère d'apprentissage semi-supervisé qui exploite un faible nombre de données étiquetées en combinaison avec une grande quantité de données non-étiquetées afin d'apprendre une représentation adaptée à la tâche de classification, et d'obtenir une meilleure performance de classification. Finalement, nous décrivons le fonctionnement général de notre système de détection d'émotions et proposons de nouvelles idées pouvant mener à de futurs travaux.

Étude et caractérisation de détecteurs à liquide en surchauffe

Les preuves astronomiques stipulent qu'environ 4\% de la densité de masse-énergie de l'univers serait composé d'atomes. Le reste est séparé entre la matière sombre, qui représente 24\% de la densité de masse-énergie, et l'énergie sombre, qui s'accapare les 71\% restant. Le neutralino est une particule prédite par la théorie de la supersymétrie et est un candidat à la composition de la matière sombre. Le Projet d'Identification des Candidats Supersymétriques Sombres (PICASSO) vise à détecter le neutralino en utilisant des détecteurs à gouttelettes de C$_4$F$_{10}$ en surchauffe. Lors du passage d'une particule dans les gouttelettes de C$_4$F$_{10}$, une transition de phase aura lieu si l'énergie déposée est au-delà du seuil prédit par le critère de nucléation d'une transition de phase (théorie de Seitz). L'onde acoustique émise durant la transition de phase est ensuite transformée en impulsion électrique par des capteurs piézoélectriques placés sur le pourtour du détecteur. Le signal est amplifié, numérisé puis enregistré afin de pouvoir être analysé par des outils numériques. L'ouvrage qui suit présente les travaux effectués sur la compréhension des signaux des détecteurs à gouttelettes en surchauffe dans le but d'améliorer la discrimination du bruit de fond. Un détecteur à petites gouttelettes, r $\approx 15\mu m$ a été étudié et comparé à une simulation Monte Carlo. Il s'est avéré que les possibilités de discrimination du bruit de fond provenant des particules alpha étaient réduites pour un détecteur à petites gouttelettes, et ce en accord avec le modèle théorique. Différentes composantes du système d'acquisition ont été testées dont le couplage entre le capteur piézoélectrique et la paroi en acrylique, l'efficacité des capteurs piézoélectriques à gain intégré et les conséquences de la force du gain sur la qualité du signal. Une comparaison avec des résultats de l'expérience SIMPLE (Superheated Instrument for Massive ParticLe Experiments) a été effectuée en mesurant des signaux de détecteurs PICASSO à l'aide d'un microphone électrostatique à électret. Il a été conclu que les détecteurs PICASSO ne parviennent pas à reproduire la discrimination quasi parfaite présentée par SIMPLE.

Théâtralité, médialité et sociomédialité : fondements et enjeux de l'intermédialité théâtrale

Née de l’avènement des technologies numériques, l’intermédialité recouvre des phénomènes aussi vieux que les plus vieux médias. Qu’on considère le théâtre comme un média ou qu’on lui reconnaisse des propriétés médiatiques, l’approche intermédiale renouvelle considérablement la compréhension que nous avons de sa nature et de sa conjoncture : ses rapports avec d’autres médias et d’autres pratiques, en particulier ceux et celles nés des technologies électriques puis numériques. La dynamique intermédiale repose autant sur le principe de remédiation tel que l’ont défini Bolter et Grusin que sur celui de reste et résidu—les médias rési-duels d’Acland. Mais un média peut aussi bloquer la logique remédiante quand des agents considèrent que cette remédiation va à l’encontre des intérêts ou de l’identité du média. C’est ce que nous appelons la résistance médiatique dont l’histoire récente du théâtre offre de nombreux exemples. Parmi eux, celui de la très lente pénétration des technologies de reproduction sonore sur les scènes et dans les processus de création. L’article suggère différentes hypothèses qui expliqueraient à la fois les causes, les modalités et les effets de cette résistance au théâtre au cours du Long Siècle (1880 à aujourd’hui), c’est-à-dire depuis la révolution électrique qui a vu, en même temps, naître la lampe à incandescence et le micro. Si la première a envahi toutes les scènes du théâtre occidental en moins de vingt ans, le second a été tenu à l’écart de la représentation pendant trois quarts de siècle.

Background-noise reduction

Interference by siren background-noise with speech transmitted from radio equipment (3) of an emergency-service vehicle, is reduced by apparatus (1) that subtracts (43) an estimate nk of the correlated siren-noise component from the contaminated signal yk supplied by the cab-microphone (2). The estimate nk is computed by FIR (finite impulse response) filtering of a siren-reference signal xk supplied by a unit (4) from one or more microphones located on or near the siren, or from the electric waveform driving the siren. The filter-coefficients wk are adjusted according to an LMS (least mean square) adaptive algorithm that is applied to the correlated-noise component ek of the fed-back noise-reduced signal, so as to bring about iterative cancellation with close frequency-tracking, of the siren noise.

A sodar for profiling in a spatially inhomogeneous urban environment

The urban boundary layer, above the canopy, is still poorly understood. One of the challenges is obtaining data by sampling more than a few meters above the rooftops, given the spatial and temporal inhomogeneities in both horizontal and vertical. Sodars are generally useful tools for ground-based remote sensing of winds and turbulence, but rely on horizontal homogeneity (as do lidars) in building up 3-component wind vectors from sampling three or more spatially separated volumes. The time taken for sound to travel to a typical range of 200 m and back is also a limitation. A sodar of radically different design is investigated, aimed at addressing these problems. It has a single vertical transmitted sound pulse. Doppler shifted signals are received from a number of volumes around the periphery of the transmitted beam with microphones which each having tight angular sensitivity at zenith angles slightly off-vertical. The spatial spread of sampled volumes is therefore smaller. By having more receiver microphones than a conventional sodar, the effect of smaller zenith angle is offset. More rapid profiling is also possible with a single vertical transmitted beam, instead of the usual multiple beams.A prototype design is described, together with initial field measurements. It is found that the beam forming using a single dish antenna and the drift of the sound pulse downwind both give rise to reduced performance compared with expectations. It is concluded that, while the new sodar works in principle, the compromises arising in the design mean that the expected advantages have not been realized

Voices of Occupiers/Liberators: The BBC's Radio Propaganda in Italy between 1942 and 1945

The ambiguity of the role played by British propaganda in Italy during the Second World War is clearly reflected in the phenomenon of Radio London. While Radio London raised the morale of the Italian civilians living under the Fascist regime and provided them with alternative information on the conflict, the microphones of the BBC were also used by the British government to address a country they were planning to occupy. In this article, I will analyse the occupation/liberation operations that were run at the BBC Italian Service from two separate angles. On the one hand, the analysis of the programmes broadcast between the months preceding the Allies’ landing in Sicily and the actual occupation shows how the Allies built their image as liberators and guarantors of better living conditions. On the other, the analysis of the relationships between the Foreign Office and the anti-Fascist exiles reveals that the Italian BBC broadcasters were not always allowed to freely express their political opinion or to dispose of their own lives.

Automatically linking live experiences captured with a ubiquitous infrastructure

Ubiquitous computing aims at providing services to users in everyday environments such as the home. One research theme in this area is that of building capture and access applications which support information to be recorded ( captured) during a live experience toward automatically producing documents for review (accessed). The recording demands instrumented environments with devices such as microphones, cameras, sensors and electronic whiteboards. Since each experience is usually related to many others ( e. g. several meetings of a project), there is a demand for mechanisms supporting the automatic linking among documents relative to different experiences. In this paper we present original results relative to the integration of our previous efforts in the Infrastructure for Capturing, Accessing, Linking, Storing and Presenting information (CALiSP). Ubiquitous computing aims at providing services to users in everyday environments such as the home. One research theme in this area is that of building capture and access applications which support information to be recorded (captured) during a live experience toward automatically producing documents for review (accessed). The recording demands instrumented environments with devices such as microphones, cameras, sensors and electronic whiteboards. Since each experience is usually related to many others (e.g. several meetings of a project), there is a demand for mechanisms supporting the automatic linking among documents relative to different experiences. In this paper we present original results relative to the integration of our previous efforts in the Infrastructure for Capturing, Accessing, Linking, Storing and Presenting information (CALiSP).

Utilização de Novas Tecnologias de Telecomunicações em Plataformas Petrolíferas Offshore

This dissertation describes the use of new Technologies of the Areas of Telecommunications, Networks and Industrial Automation for increase of the Operational Safety and obtaining of Operational Improvements in the Platforms Petroliferous Offshore. The presented solution represents the junction of several modules of these areas, making possible the Supervision and Contrai of the Platforms Petroliferous Offshore starting from an Station Onshore, in way similar to a remote contral, by virtue of the visualization possibility and audition of the operational area through cameras and microphones, looking the operator of the system to be "present" in the platform. This way, it diminishes the embarked people's need, increasing the Operational Safety. As consequence, we have the obtaining of Operational Improvements, by virtue of the use of a digital link of large band it releases multi-service. In this link traffic simultaneously digital signs of data (Ethernet Network), telephony (Phone VoIP), image and sound

Autenticação biométrica de usuários em sistemas de E-learning baseada em reconhecimento de faces a partir de vídeo

Pós-graduação em Ciência da Computação - IBILCE

Evaluation of hearing protection used by police officers in the shooting range

Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)

Analysis of acoustic frequency and wave amplitude generated by the Oscillatory Thoracic Thixotropic Device (Diottix(r)) in human chest

Diottix(r) was calibrated at 25 Hz to achieve the frequency indicated in literature as being effective to mobilize the airways secretions. However, the amplitude and frequency of the waves generated by the equipment in different regions of the chest still need to be investigated. The objective of this study was to analyze the frequency and amplitude of waves generated by Diottix(r) in chests of healthy subjects. Diottix(r) was used in the anterior and posterior regions of the chest. The mechanical waves were captured using stethoscopes connected to electret microphones, which were connected to a digital oscilloscope. Frequency and amplitude data were recorded by the stethoscope, positioned in six points in the anterior region and six in the posterior region of the chest, following the positions commonly used in pulmonary auscultation. Signals were recorded and transferred to a computer with software for their analysis. The frequency of waves did not present a significant change (from 24.9 to 26.4 Hz). The wave amplitude in the anterior versus the posterior region in each area of the lung, the upper, middle and lower, had differences. Diottix(r) produces frequencies in the chest according to the calibrated; thus, it can be a complementary resource to bronchial hygiene maneuvers. The amplitudes of waves seem to be affected by other structures like bone parts and heart.

AcMus: an Open, Integrated Platform for Room Acoustics Research

This article describes the design, implementation, and experiences with AcMus, an open and integrated software platform for room acoustics research, which comprises tools for measurement, analysis, and simulation of rooms for music listening and production. Through use of affordable hardware, such as laptops, consumer audio interfaces and microphones, the software allows evaluation of relevant acoustical parameters with stable and consistent results, thus providing valuable information in the diagnosis of acoustical problems, as well as the possibility of simulating modifications in the room through analytical models. The system is open-source and based on a flexible and extensible Java plug-in framework, allowing for cross-platform portability, accessibility and experimentation, thus fostering collaboration of users, developers and researchers in the field of room acoustics.

The uncertainty in standardised sound power measurements: complying with ISO 17025

In the context of “testing laboratory” one of the most important aspect to deal with is the measurement result. Whenever decisions are based on measurement results, it is important to have some indication of the quality of the results. In every area concerning with noise measurement many standards are available but without an expression of uncertainty, it is impossible to judge whether two results are in compliance or not. ISO/IEC 17025 is an international standard related with the competence of calibration and testing laboratories. It contains the requirements that testing and calibration laboratories have to meet if they wish to demonstrate that they operate to a quality system, are technically competent and are able to generate technically valid results. ISO/IEC 17025 deals specifically with the requirements for the competence of laboratories performing testing and calibration and for the reporting of the results, which may or may not contain opinions and interpretations of the results. The standard requires appropriate methods of analysis to be used for estimating uncertainty of measurement. In this point of view, for a testing laboratory performing sound power measurement according to specific ISO standards and European Directives, the measurement of uncertainties is the most important factor to deal with. Sound power level measurement, according to ISO 3744:1994 , performed with a limited number of microphones distributed over a surface enveloping a source is affected by a certain systematic error and a related standard deviation. Making a comparison of measurement carried out with different microphone arrays is difficult because results are affected by systematic errors and standard deviation that are peculiarities of the number of microphones disposed on the surface, their spatial position and the complexity of the sound field. A statistical approach could give an overview of the difference between sound power level evaluated with different microphone arrays and an evaluation of errors that afflict this kind of measurement. Despite the classical approach that tend to follow the ISO GUM this thesis present a different point of view of the problem related to the comparison of result obtained from different microphone arrays.

Fluidic oscillator design for water removal enhancement in a PEM fuel cell

This research initiative was triggered by the problems of water management of Polymer Electrolyte Membrane Fuel Cell (PEMFC). In low temperature fuel cells such as PEMFC, some of the water produced after the chemical reaction remains in its liquid state. Excess water produced by the fuel cell must be removed from the system to avoid flooding of the gas diffusion layers (GDL). The GDL is responsible for the transport of reactant gas to the active sites and remove the water produced from the sites. If the GDL is flooded, the supply gas will not be able to reach the reactive sites and the fuel cell fails. The choice of water removal method in this research is to exert a variable asymmetrical force on a liquid droplet. As the drop of liquid is subjected to an external vibrational force in the form of periodic wave, it will begin to oscillate. A fluidic oscillator is capable to produce a pulsating flow using simple balance of momentum fluxes between three impinging jets. By connecting the outputs of the oscillator to the gas channels of a fuel cell, a flow pulsation can be imposed on a water droplet formed within the gas channel during fuel cell operation. The lowest frequency produced by this design is approximately 202 Hz when a 20 inches feed-back port length was used and a supply pressure of 5 psig was introduced. This information was found by setting up a fluidic network with appropriate data acquisition. The components include a fluidic amplifier, valves and fittings, flow meters, a pressure gage, NI-DAQ system, Siglab®, Matlab software and four PCB microphones. The operating environment of the water droplet was reviewed, speed of the sound pressure which travels down the square channel was precisely estimated, and measurement devices were carefully selected. Applicable alternative measurement devices and its application to pressure wave measurement was considered. Methods for experimental setup and possible approaches were recommended, with some discussion of potential problems with implementation of this technique. Some computational fluid dynamic was also performed as an approach to oscillator design.

Speaker Diarization Based on Intensity Channel Contribution

The time delay of arrival (TDOA) between multiple microphones has been used since 2006 as a source of information (localization) to complement the spectral features for speaker diarization. In this paper, we propose a new localization feature, the intensity channel contribution (ICC) based on the relative energy of the signal arriving at each channel compared to the sum of the energy of all the channels. We have demonstrated that by joining the ICC features and the TDOA features, the robustness of the localization features is improved and that the diarization error rate (DER) of the complete system (using localization and spectral features) has been reduced. By using this new localization feature, we have been able to achieve a 5.2% DER relative improvement in our development data, a 3.6% DER relative improvement in the RT07 evaluation data and a 7.9% DER relative improvement in the last year's RT09 evaluation data.