Apprentissage de représentations sur-complètes par entraînement d’auto-encodeurs

Les avancés dans le domaine de l’intelligence artificielle, permettent à des systèmes informatiques de résoudre des tâches de plus en plus complexes liées par exemple à la vision, à la compréhension de signaux sonores ou au traitement de la langue. Parmi les modèles existants, on retrouve les Réseaux de Neurones Artificiels (RNA), dont la popularité a fait un grand bond en avant avec la découverte de Hinton et al. [22], soit l’utilisation de Machines de Boltzmann Restreintes (RBM) pour un pré-entraînement non-supervisé couche après couche, facilitant grandement l’entraînement supervisé du réseau à plusieurs couches cachées (DBN), entraînement qui s’avérait jusqu’alors très difficile à réussir. Depuis cette découverte, des chercheurs ont étudié l’efficacité de nouvelles stratégies de pré-entraînement, telles que l’empilement d’auto-encodeurs traditionnels(SAE) [5, 38], et l’empilement d’auto-encodeur débruiteur (SDAE) [44]. C’est dans ce contexte qu’a débuté la présente étude. Après un bref passage en revue des notions de base du domaine de l’apprentissage machine et des méthodes de pré-entraînement employées jusqu’à présent avec les modules RBM, AE et DAE, nous avons approfondi notre compréhension du pré-entraînement de type SDAE, exploré ses différentes propriétés et étudié des variantes de SDAE comme stratégie d’initialisation d’architecture profonde. Nous avons ainsi pu, entre autres choses, mettre en lumière l’influence du niveau de bruit, du nombre de couches et du nombre d’unités cachées sur l’erreur de généralisation du SDAE. Nous avons constaté une amélioration de la performance sur la tâche supervisée avec l’utilisation des bruits poivre et sel (PS) et gaussien (GS), bruits s’avérant mieux justifiés que celui utilisé jusqu’à présent, soit le masque à zéro (MN). De plus, nous avons démontré que la performance profitait d’une emphase imposée sur la reconstruction des données corrompues durant l’entraînement des différents DAE. Nos travaux ont aussi permis de révéler que le DAE était en mesure d’apprendre, sur des images naturelles, des filtres semblables à ceux retrouvés dans les cellules V1 du cortex visuel, soit des filtres détecteurs de bordures. Nous aurons par ailleurs pu montrer que les représentations apprises du SDAE, composées des caractéristiques ainsi extraites, s’avéraient fort utiles à l’apprentissage d’une machine à vecteurs de support (SVM) linéaire ou à noyau gaussien, améliorant grandement sa performance de généralisation. Aussi, nous aurons observé que similairement au DBN, et contrairement au SAE, le SDAE possédait une bonne capacité en tant que modèle générateur. Nous avons également ouvert la porte à de nouvelles stratégies de pré-entraînement et découvert le potentiel de l’une d’entre elles, soit l’empilement d’auto-encodeurs rebruiteurs (SRAE).

Réseaux de neurones à relaxation entraînés par critère d'autoencodeur débruitant

L’apprentissage machine est un vaste domaine où l’on cherche à apprendre les paramètres de modèles à partir de données concrètes. Ce sera pour effectuer des tâches demandant des aptitudes attribuées à l’intelligence humaine, comme la capacité à traiter des don- nées de haute dimensionnalité présentant beaucoup de variations. Les réseaux de neu- rones artificiels sont un exemple de tels modèles. Dans certains réseaux de neurones dits profonds, des concepts "abstraits" sont appris automatiquement. Les travaux présentés ici prennent leur inspiration de réseaux de neurones profonds, de réseaux récurrents et de neuroscience du système visuel. Nos tâches de test sont la classification et le débruitement d’images quasi binaires. On permettra une rétroac- tion où des représentations de haut niveau (plus "abstraites") influencent des représentations à bas niveau. Cette influence s’effectuera au cours de ce qu’on nomme relaxation, des itérations où les différents niveaux (ou couches) du modèle s’interinfluencent. Nous présentons deux familles d’architectures, l’une, l’architecture complètement connectée, pouvant en principe traiter des données générales et une autre, l’architecture convolutionnelle, plus spécifiquement adaptée aux images. Dans tous les cas, les données utilisées sont des images, principalement des images de chiffres manuscrits. Dans un type d’expérience, nous cherchons à reconstruire des données qui ont été corrompues. On a pu y observer le phénomène d’influence décrit précédemment en comparant le résultat avec et sans la relaxation. On note aussi certains gains numériques et visuels en terme de performance de reconstruction en ajoutant l’influence des couches supérieures. Dans un autre type de tâche, la classification, peu de gains ont été observés. On a tout de même pu constater que dans certains cas la relaxation aiderait à apprendre des représentations utiles pour classifier des images corrompues. L’architecture convolutionnelle développée, plus incertaine au départ, permet malgré tout d’obtenir des reconstructions numériquement et visuellement semblables à celles obtenues avec l’autre architecture, même si sa connectivité est contrainte.

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.

Denoising swallowing sound to improve the evaluator`s qualitative analysis

Swallowing dynamics involves the coordination and interaction of several muscles and nerves which allow correct food transport from mouth to stomach without laryngotracheal penetration or aspiration. Clinical swallowing assessment depends on the evaluator`s knowledge of anatomic structures and of neurophysiological processes involved in swallowing. Any alteration in those steps is denominated oropharyngeal dysphagia, which may have many causes, such as neurological or mechanical disorders. Videofluoroscopy of swallowing is presently considered to be the best exam to objectively assess the dynamics of swallowing, but the exam needs to be conducted under certain restrictions, due to patient`s exposure to radiation, which limits periodical repetition for monitoring swallowing therapy. Another method, called cervical auscultation, is a promising new diagnostic tool for the assessment of swallowing disorders. The potential to diagnose dysphagia in a noninvasive manner by assessing the sounds of swallowing is a highly attractive option for the dysphagia clinician. Even so, the captured sound has an amount of noise, which can hamper the evaluator`s decision. In that way, the present paper proposes the use of a filter to improve the quality of audible sound and facilitate the perception of examination. The wavelet denoising approach is used to decompose the noisy signal. The signal to noise ratio was evaluated to demonstrate the quantitative results of the proposed methodology. (C) 2007 Elsevier Ltd. All rights reserved.

Fluorescence Microscopy Imaging Denoising with Log-Euclidean Priors and Photobleaching Compensation

Fluorescent protein microscopy imaging is nowadays one of the most important tools in biomedical research. However, the resulting images present a low signal to noise ratio and a time intensity decay due to the photobleaching effect. This phenomenon is a consequence of the decreasing on the radiation emission efficiency of the tagging protein. This occurs because the fluorophore permanently loses its ability to fluoresce, due to photochemical reactions induced by the incident light. The Poisson multiplicative noise that corrupts these images, in addition with its quality degradation due to photobleaching, make long time biological observation processes very difficult. In this paper a denoising algorithm for Poisson data, where the photobleaching effect is explicitly taken into account, is described. The algorithm is designed in a Bayesian framework where the data fidelity term models the Poisson noise generation process as well as the exponential intensity decay caused by the photobleaching. The prior term is conceived with Gibbs priors and log-Euclidean potential functions, suitable to cope with the positivity constrained nature of the parameters to be estimated. Monte Carlo tests with synthetic data are presented to characterize the performance of the algorithm. One example with real data is included to illustrate its application.

Convex Total Variation Denoising of Poisson Fluorescence Confocal Images With Anisotropic Filtering

Fluorescence confocal microscopy (FCM) is now one of the most important tools in biomedicine research. In fact, it makes it possible to accurately study the dynamic processes occurring inside the cell and its nucleus by following the motion of fluorescent molecules over time. Due to the small amount of acquired radiation and the huge optical and electronics amplification, the FCM images are usually corrupted by a severe type of Poisson noise. This noise may be even more damaging when very low intensity incident radiation is used to avoid phototoxicity. In this paper, a Bayesian algorithm is proposed to remove the Poisson intensity dependent noise corrupting the FCM image sequences. The observations are organized in a 3-D tensor where each plane is one of the images acquired along the time of a cell nucleus using the fluorescence loss in photobleaching (FLIP) technique. The method removes simultaneously the noise by considering different spatial and temporal correlations. This is accomplished by using an anisotropic 3-D filter that may be separately tuned in space and in time dimensions. Tests using synthetic and real data are described and presented to illustrate the application of the algorithm. A comparison with several state-of-the-art algorithms is also presented.

A denoising approach for iterative side information creation in distributed video coding

In distributed video coding, motion estimation is typically performed at the decoder to generate the side information, increasing the decoder complexity while providing low complexity encoding in comparison with predictive video coding. Motion estimation can be performed once to create the side information or several times to refine the side information quality along the decoding process. In this paper, motion estimation is performed at the decoder side to generate multiple side information hypotheses which are adaptively and dynamically combined, whenever additional decoded information is available. The proposed iterative side information creation algorithm is inspired in video denoising filters and requires some statistics of the virtual channel between each side information hypothesis and the original data. With the proposed denoising algorithm for side information creation, a RD performance gain up to 1.2 dB is obtained for the same bitrate.

ECG denoising based on adaptive signal processing technique

An Electrocardiogram (ECG) monitoring system deals with several challenges related with noise sources. The main goal of this text was the study of Adaptive Signal Processing Algorithms for ECG noise reduction when applied to real signals. This document presents an adaptive ltering technique based on Least Mean Square (LMS) algorithm to remove the artefacts caused by electromyography (EMG) and power line noise into ECG signal. For this experiments it was used real noise signals, mainly to observe the di erence between real noise and simulated noise sources. It was obtained very good results due to the ability of noise removing that can be reached with this technique. A recolha de sinais electrocardiogr a cos (ECG) sofre de diversos problemas relacionados com ru dos. O objectivo deste trabalho foi o estudo de algoritmos adaptativos para processamento digital de sinal, para redu c~ao de ru do em sinais ECG reais. Este texto apresenta uma t ecnica de redu c~ao de ru do baseada no algoritmo Least Mean Square (LMS) para remo c~ao de ru dos causados quer pela actividade muscular (EMG) quer por ru dos causados pela rede de energia el ectrica. Para as experiencias foram utilizados ru dos reais, principalmente para aferir a diferen ca de performance do algoritmo entre os sinais reais e os simulados. Foram conseguidos bons resultados, essencialmente devido as excelentes caracter sticas que esta t ecnica tem para remover ru dos.

Denoising of LSFCM images with compensation for the photoblinking/photobleaching effects

Fluorescence confocal microscopy images present a low signal to noise ratio and a time intensity decay due to the so called photoblinking and photobleaching effects. These effects, together with the Poisson multiplicative noise that corrupts the images, make long time biological observation processes very difficult.

Irregular to regular sampling, denoising and deconvolution

We propose a restoration algorithm for band limited images that considers irregular(perturbed) sampling, denoising, and deconvolution. We explore the application of a family ofregularizers that allow to control the spectral behavior of the solution combined with the irregular toregular sampling algorithms proposed by H.G. Feichtinger, K. Gr¨ochenig, M. Rauth and T. Strohmer.Moreover, the constraints given by the image acquisition model are incorporated as a set of localconstraints. And the analysis of such constraints leads to an early stopping rule meant to improvethe speed of the algorithm. Finally we present experiments focused on the restoration of satellite images, where the micro-vibrations are responsible of the type of distortions we are considering here. We will compare results of the proposed method with previous methods and show an extension tozoom.

Image Denoising Using Sure-Based Adaptive Thresholding In Directionlet Domain

The standard separable two dimensional wavelet transform has achieved a great success in image denoising applications due to its sparse representation of images. However it fails to capture efficiently the anisotropic geometric structures like edges and contours in images as they intersect too many wavelet basis functions and lead to a non-sparse representation. In this paper a novel de-noising scheme based on multi directional and anisotropic wavelet transform called directionlet is presented. The image denoising in wavelet domain has been extended to the directionlet domain to make the image features to concentrate on fewer coefficients so that more effective thresholding is possible. The image is first segmented and the dominant direction of each segment is identified to make a directional map. Then according to the directional map, the directionlet transform is taken along the dominant direction of the selected segment. The decomposed images with directional energy are used for scale dependent subband adaptive optimal threshold computation based on SURE risk. This threshold is then applied to the sub-bands except the LLL subband. The threshold corrected sub-bands with the unprocessed first sub-band (LLL) are given as input to the inverse directionlet algorithm for getting the de-noised image. Experimental results show that the proposed method outperforms the standard wavelet-based denoising methods in terms of numeric and visual quality

Spatially Adaptive Image Denoising Techniques Using Directionlets

The aim of the thesis was to design and develop spatially adaptive denoising techniques with edge and feature preservation, for images corrupted with additive white Gaussian noise and SAR images affected with speckle noise. Image denoising is a well researched topic. It has found multifaceted applications in our day to day life. Image denoising based on multi resolution analysis using wavelet transform has received considerable attention in recent years. The directionlet based denoising schemes presented in this thesis are effective in preserving the image specific features like edges and contours in denoising. Scope of this research is still open in areas like further optimization in terms of speed and extension of the techniques to other related areas like colour and video image denoising. Such studies would further augment the practical use of these techniques.

Apodisation, denoising and system identification techniques for THz transients in the wavelet domain

This work describes the use of a quadratic programming optimization procedure for designing asymmetric apodization windows to de-noise THz transient interferograms and compares these results to those obtained when wavelet signal processing algorithms are adopted. A systems identification technique in the wavelet domain is also proposed for the estimation of the complex insertion loss function. The proposed techniques can enhance the frequency dependent dynamic range of an experiment and should be of particular interest to the THz imaging and tomography community. Future advances in THz sources and detectors are likely to increase the signal-to-noise ratio of the recorded THz transients and high quality apodization techniques will become more important, and may set the limit on the achievable accuracy of the deduced spectrum.

Gaussian processes autoencoder for dimensionality reduction

Learning low dimensional manifold from highly nonlinear data of high dimensionality has become increasingly important for discovering intrinsic representation that can be utilized for data visualization and preprocessing. The autoencoder is a powerful dimensionality reduction technique based on minimizing reconstruction error, and it has regained popularity because it has been efficiently used for greedy pretraining of deep neural networks. Compared to Neural Network (NN), the superiority of Gaussian Process (GP) has been shown in model inference, optimization and performance. GP has been successfully applied in nonlinear Dimensionality Reduction (DR) algorithms, such as Gaussian Process Latent Variable Model (GPLVM). In this paper we propose the Gaussian Processes Autoencoder Model (GPAM) for dimensionality reduction by extending the classic NN based autoencoder to GP based autoencoder. More interestingly, the novel model can also be viewed as back constrained GPLVM (BC-GPLVM) where the back constraint smooth function is represented by a GP. Experiments verify the performance of the newly proposed model.

Denoising swallowing sound to improve the evaluator's qualitative analysis

Swallowing dynamics involves the coordination and interaction of several muscles and nerves which allow correct food transport from mouth to stomach without laryngotracheal penetration or aspiration. Clinical swallowing assessment depends on the evaluator's knowledge of anatomic structures and of neurophysiological processes involved in swallowing. Any alteration in those steps is denominated oropharyngeal dysphagia, which may have many causes, such as neurological or mechanical disorders. Videofluoroscopy of swallowing is presently considered to be the best exam to objectively assess the dynamics of swallowing, but the exam needs to be conducted under certain restrictions, due to patient's exposure to radiation, which limits periodical repetition for monitoring swallowing therapy. Another method, called cervical auscultation, is a promising new diagnostic tool for the assessment of swallowing disorders. The potential to diagnose dysphagia in a noninvasive manner by assessing the sounds of swallowing is a highly attractive option for the dysphagia clinician. Even so, the captured sound has an amount of noise, which can hamper the evaluator's decision. In that way, the present paper proposes the use of a filter to improve the quality of audible sound and facilitate the perception of examination. The wavelet denoising approach is used to decompose the noisy signal. The signal to noise ratio was evaluated to demonstrate the quantitative results of the proposed methodology. (C) 2007 Elsevier Ltd. All rights reserved.