No description available.

In this work, we first analyze the behavior of the acoustic wave from the theoretical point of view using a simplified 1-dimensional model. The model was developed based on the conservation of mass, the conservation of momentum, and the state equation; in addition, the fluid medium obeys Stokes assumption and it is homogeneous, adiabatic and isentropic. Experiment and simulation results show consistency to theoretical calculation. The second part of this thesis focuses on the PD signal analysis from an on-site PD measurement of the in-house design fiber optic sensors (by Virginia Tech, Center for Photonics Technology). Several commercial piezoelectric transducers (PZTs) were also used to compare the measurement results. The signal analysis employs the application of wavelet-based denoising technique to remove the noises, which mainly came from vibration, EMI, and light sources, embedded in the PD signal. The denoising technique includes the discrete wavelet transform (DWT) decomposition, thresh-holding of wavelet coefficients, and signal recovery by inverse discrete wavelet transform. Several approaches were compared to determine the optimal mother wavelet. The threshold limits are selected to remove the maximum Gaussian noises for each level of wavelet coefficients. The results indicate that this method could extract the PD spike from the noisy measurement effectively. The frequency of the PD pulse is also analyzed; it is shown that the frequencies lie in the range of 70 kHz to 250 kHz. In addition, with the assumed acoustic wave propagation delay between PD source and sensors, it was found that all PD activities occur in the first and third quadrant in reference to the applied sinusoidal transformer voltage. / Master of Science

Cette thèse s'inscrit dans le contexte des méthodes non locales pour le traitement d'images et a pour application principale le débruitage, bien que les méthodes étudiées soient suffisamment génériques pour être applicables à d'autres problèmes inverses en imagerie. Les images naturelles sont constituées de structures redondantes, et cette redondance peut être exploitée à des fins de restauration. Une manière classique d’exploiter cette auto-similarité est de découper l'image en patchs. Ces derniers peuvent ensuite être regroupés, comparés et filtrés ensemble.Dans le premier chapitre, le principe du "global denoising" est reformulé avec le formalisme classique de l'estimation diagonale et son comportement asymptotique est étudié dans le cas oracle. Des conditions précises à la fois sur l'image et sur le filtre global sont introduites pour assurer et quantifier la convergence.Le deuxième chapitre est consacré à l'étude d’a priori gaussiens ou de type mélange de gaussiennes pour le débruitage d'images par patches. Ces a priori sont largement utilisés pour la restauration d'image. Nous proposons ici quelques indices pour répondre aux questions suivantes : Pourquoi ces a priori sont-ils si largement utilisés ? Quelles informations encodent-ils ?Le troisième chapitre propose un modèle probabiliste de mélange pour les patchs bruités, adapté à la grande dimension. Il en résulte un algorithme de débruitage qui atteint les performance de l'état-de-l'art.Le dernier chapitre explore des pistes d'agrégation différentes et propose une écriture de l’étape d'agrégation sous la forme d'un problème de moindre carrés. / This thesis studies non-local methods for image processing, and their application to various tasks such as denoising. Natural images contain redundant structures, and this property can be used for restoration purposes. A common way to consider this self-similarity is to separate the image into "patches". These patches can then be grouped, compared and filtered together.In the first chapter, "global denoising" is reframed in the classical formalism of diagonal estimation and its asymptotic behaviour is studied in the oracle case. Precise conditions on both the image and the global filter are introduced to ensure and quantify convergence.The second chapter is dedicated to the study of Gaussian priors for patch-based image denoising. Such priors are widely used for image restoration. We propose some ideas to answer the following questions: Why are Gaussian priors so widely used? What information do they encode about the image?The third chapter proposes a probabilistic high-dimensional mixture model on the noisy patches. This model adopts a sparse modeling which assumes that the data lie on group-specific subspaces of low dimensionalities. This yields a denoising algorithm that demonstrates state-of-the-art performance.The last chapter explores different way of aggregating the patches together. A framework that expresses the patch aggregation in the form of a least squares problem is proposed.

In the emerging field of audio generation using diffusion models, this project pioneers the adaptation of the AudioLDM model framework, initially designed for text-to-daily sounds generation, towards text-to-music audio generation. This shift addresses a gap in the current scope of audio diffusion models, predominantly focused on everyday sounds. The motivation for this thesis stems from AudioLDM’s remarkable generative capabilities in producing daily sounds from text descriptions. However, its application in music audio generation remains underexplored. The thesis aims to modify AudioLDM’s architecture and training objectives to cater to the unique nuances of musical audio. The re-engineering process involved two primary methods. First, a dataset was constructed by sourcing a variety of music audio samples from the A Dataset For Music Analysis (FMA) [1] and generating pseudo captions using a Large Language Model specified in music captioning. This dataset served as the foundation for training the adapted model. Second, the model’s diffusion backbone, a UNet architecture, was revised in its text conditioning approach by incorporating both the CLAP encoder and the T5 text encoder. This dualencoding method, coupled with a shift from the traditional noise prediction objective to the V-objective, aimed to enhance the model’s performance in generating coherent and musically relevant audio. The effectiveness of these adaptations was validated through both subjective and objective evaluations. Compared to the original AudioLDM model, the adapted version demonstrated superior quality in the audio output and a higher relevance between text prompts and generated music. This advancement not only proves the feasibility of transforming AudioLDM for music generation but also opens new avenues for research and application in text-to-music audio synthesis / Inom det framväxande området för ljudgenerering med användning av diffusionsmodeller, banar detta projekt för anpassningen av AudioLDMmodellramverket, som ursprungligen utformades för generering av text-tilldagliga ljud, mot ljudgenerering av text-till-musik. Denna förändring tar itu med en lucka i den nuvarande omfattningen av ljuddiffusionsmodeller, främst inriktade på vardagliga ljud. Motivationen för denna avhandling kommer från AudioLDM:s anmärkningsvärda generativa förmåga att producera dagliga ljud från textbeskrivningar. Dock är dess tillämpning i musikljudgenerering fortfarande underutforskad. Avhandlingen syftar till att modifiera AudioLDM:s arkitektur och utbildningsmål för att tillgodose de unika nyanserna av musikaliskt ljud. Omarbetningsprocessen involverade två primära metoder. Först konstruerades en datauppsättning genom att hämta en mängd olika musikljudprover från A Dataset For Music Analysis (FMA) [1] och generera pseudotexter med hjälp av en Large Language Model specificerad i musiktextning. Denna datauppsättning fungerade som grunden för att träna den anpassade modellen. För det andra reviderades modellens diffusionsryggrad, en UNet-arkitektur, i sin textkonditioneringsmetod genom att inkludera både CLAP-kodaren och T5-textkodaren. Denna dubbelkodningsmetod, i kombination med en övergång från det traditionella brusförutsägelsemålet till V-målet, syftade till att förbättra modellens prestanda för att generera sammanhängande och musikaliskt relevant ljud. Effektiviteten av dessa anpassningar validerades genom både subjektiva och objektiva utvärderingar. Jämfört med den ursprungliga AudioLDMmodellen visade den anpassade versionen överlägsen kvalitet i ljudutgången och en högre relevans mellan textmeddelanden och genererad musik. Detta framsteg bevisar inte bara möjligheten att transformera AudioLDM för musikgenerering utan öppnar också nya vägar för forskning och tillämpning inom text-till-musik ljudsyntes.

Les modèles génératifs servent à générer des échantillons d'une loi de probabilité (ex. : du texte, des images, de la musique, des vidéos, des molécules, et beaucoup plus) à partir d'un jeu de données (ex. : une banque d'images, de texte, ou autre). Entrainer des modèles génératifs est une tâche très difficile, mais ces outils ont un très grand potentiel en termes d'applications. Par exemple, dans le futur lointain, on pourrait envisager qu'un modèle puisse générer les épisodes d'une émission de télévision à partir d'un script et de voix générés par d'autres modèles génératifs. Il existe plusieurs types de modèles génératifs. Pour la génération d'images, l'approche la plus fructueuse est sans aucun doute la méthode de réseaux adverses génératifs (GANs). Les GANs apprennent à générer des images par un jeu compétitif entre deux joueurs, le Discriminateur et le Générateur. Le Discriminateur tente de prédire si une image est vraie ou fausse, tandis que le Générateur tente de générer des images plus réalistes en apprenant à faire croire au discriminateur que ces fausses images générées sont vraies. En complétant ce jeu, les GANs arrivent à générer des images presque photo-réalistes. Il est souvent possible pour des êtres humains de distinguer les fausses images (générés par les GANs) des vraies images (ceux venant du jeu de données), mais la tâche devient plus difficile au fur et à mesure que cette technologie s'améliore. Le plus gros défaut des GANs est que les données générées par les GANs manquent souvent de diversité (ex. : les chats au visage aplati sont rares dans la banque d'images, donc les GANs génèrent juste des races de chats plus fréquentes). Ces méthodes souvent aussi souvent très instables. Il y a donc encore beaucoup de chemin à faire avant l'obtention d'images parfaitement photo-réalistes et diverses. De nouvelles méthodes telles que les modèles de diffusion à la base de score semblent produire de meilleurs résultats que les GANs, donc tout n'est pas gagné pour les GANs. C'est pourquoi cette thèse n'est pas concentrée seulement sur les GANs, mais aussi sur les modèles de diffusion. Notez que cette thèse est exclusivement concentrée sur la génération de données continues (ex. : images, musique, vidéos) plutôt que discrètes (ex. : texte), car cette dernière fait usage de méthodes complètement différentes. Le premier objectif de cette thèse est d'étudier les modèles génératifs de façon théorique pour mieux les comprendre. Le deuxième objectif de cette thèse est d'inventer de nouvelles astuces (nouvelles fonctions objectives, régularisations, architectures, etc.) permettant d'améliorer les modèles génératifs. Le troisième objectif est de généraliser ces approches au-delà de leur formulation initiale, pour permettre la découverte de nouveaux liens entre différentes approches. Ma première contribution est de proposer un discriminateur relativiste qui estime la probabilité qu'une donnée réelle, soit plus réaliste qu'une donnée fausse (inventée par un modèle générateur). Les GANs relativistes forment une nouvelle classe de fonctions de perte qui apportent beaucoup de stabilité durant l'entrainement. Ma seconde contribution est de prouver que les GANs relativistes forment une mesure de dissimilarité. Ma troisième contribution est de concevoir une variante adverse au appariement de score pour produire des données de meilleure qualité avec les modèles de diffusion. Ma quatrième contribution est d'améliorer la vitesse de génération des modèles de diffusion par la création d'une méthode numérique de résolution pour équations différentielles stochastiques (SDEs). / Generative models are powerful tools to generate samples (e.g., images, music, text) from an unknown distribution given a finite set of examples. Generative models are hard to train successfully, but they have the potential to revolutionize arts, science, and business. These models can generate samples from various data types (e.g., text, images, audio, videos, 3d). In the future, we can envision generative models being used to create movies or episodes from a TV show given a script (possibly also generated by a generative model). One of the most successful methods for generating images is Generative Adversarial Networks (GANs). This approach consists of a game between two players, the Discriminator and the Generator. The goal of the Discriminator is to classify an image as real or fake, while the Generator attempts to fool the Discriminator into thinking that the fake images it generates are real. Through this game, GANs are able to generate very high-quality samples, such as photo-realistic images. Humans are still generally able to distinguish real images (from the training dataset) from fake images (generated by GANs), but the gap is lessening as GANs become better over time. The biggest weakness of GANs is that they have trouble generating diverse data representative of the full range of the data distribution. Thus, there is still much progress to be made before GANs reach their full potential. New methods performing better than GANs are also appearing. One prime example is score-based diffusion models. This thesis focuses on generative models that seemed promising at the time for continuous data generation: GANs and score-based diffusion models. I seek to improve generative models so that they reach their full potential (Objective 1: Improving) and to understand these approaches better on a theoretical level (Objective 2: Theoretical understanding). I also want to generalize these approaches beyond their original setting (Objective 3: Generalizing), allowing the discovery of new connections between different concepts/fields. My first contribution is to propose using a relativistic discriminator, which estimates the probability that a given real data is more realistic than a randomly sampled fake data. Relativistic GANs form a new class of GAN loss functions that are much more stable with respect to optimization hyperparameters. My second contribution is to take a more rigorous look at relativistic GANs and prove that they are proper statistical divergences. My third contribution is to devise an adversarial variant to denoising score matching, which leads to higher quality data with score-based diffusion models. My fourth contribution is to significantly improve the speed of score-based diffusion models through a carefully devised Stochastic Differential Equation (SDE) solver.

Nous étudions dans cette thèse une représentation particulière de signaux fondée sur une méthode d’apprentissage statistique, qui consiste à modéliser des données comme combinaisons linéaires de quelques éléments d’un dictionnaire appris. Ceci peut être vu comme une extension du cadre classique des ondelettes, dont le but est de construire de tels dictionnaires (souvent des bases orthonormales) qui sont adaptés aux signaux naturels. Un succès important de cette approche a été sa capacité à modéliser des imagettes, et la performance des méthodes de débruitage d’images fondées sur elle. Nous traitons plusieurs questions ouvertes, qui sont reliées à ce cadre : Comment apprendre efficacement un dictionnaire ? Comment enrichir ce modèle en ajoutant une structure sous-jacente au dictionnaire ? Est-il possible d’améliorer les méthodes actuelles de traitement d’image fondées sur cette approche ? Comment doit-on apprendre le dictionnaire lorsque celui-ci est utilisé pour une tâche autre que la reconstruction de signaux ? Y a-t-il des applications intéressantes de cette méthode en vision par ordinateur ? Nous répondons à ces questions, avec un point de vue multidisciplinaire, en empruntant des outils d’apprentissage statistique, d’optimisation convexe et stochastique, de traitement des signaux et des images, de vison par ordinateur, mais aussi d'optimisation sur des graphes. / We study in this thesis a particular machine learning approach to represent signals that that consists of modelling data as linear combinations of a few elements from a learned dictionary. It can be viewed as an extension of the classical wavelet framework, whose goal is to design such dictionaries (often orthonormal basis) that are adapted to natural signals. An important success of dictionary learning methods has been their ability to model natural image patches and the performance of image denoising algorithms that it has yielded. We address several open questions related to this framework: How to efficiently optimize the dictionary? How can the model be enriched by adding a structure to the dictionary? Can current image processing tools based on this method be further improved? How should one learn the dictionary when it is used for a different task than signal reconstruction? How can it be used for solving computer vision problems? We answer these questions with a multidisciplinarity approach, using tools from statistical machine learning, convex and stochastic optimization, image and signal processing, computer vision, but also optimization on graphs.

Since the birth of charge coupled devices (CCD) and the complementary metal-oxide-semiconductor (CMOS) active pixel sensors, pixel pitch of digital image sensors has been continuously shrinking to meet the resolution and size requirements of the cameras. However, shrinking pixels reduces the maximum number of photons a sensor can hold, a phenomenon broadly known as the full-well capacity limit. The drop in full-well capacity causes drop in signal-to-noise ratio and dynamic range.<div><br></div><div>The Quanta Image Sensor (QIS) is a class of solid-state image sensors proposed by Eric Fossum in 2005 as a potential solution for the limited full-well capacity problem. QIS is envisioned to be the next generation image sensor after CCD and CMOS since it enables sub-diffraction-limit pixels without the inherited problems of pixel shrinking. Equipped with a massive number of detectors that have single-photon sensitivity, the sensor counts the incoming photons and triggers a binary response “1” if the photon count exceeds a threshold, or “0” otherwise. To acquire an image, the sensor oversamples the space and time to generate a sequence of binary bit maps. Because of this binary sensing mechanism, the full-well capacity, signal-to-noise ratio and the dynamic range can all be improved using an appropriate image reconstruction algorithm. The contribution of this thesis is to address three image processing problems in QIS: 1) Image reconstruction, 2) Threshold design and 3) Color filter array design.</div><div><br></div><div>Part 1 of the thesis focuses on reconstructing the latent grayscale image from the QIS binary measurements. Image reconstruction is a necessary step for QIS because the raw binary measurements are not images. Previous methods in the literature use iterative algorithms which are computationally expensive. By modeling the QIS binary measurements as quantized Poisson random variables, a new non-iterative image reconstruction method based on the Transform-Denoise framework is proposed. Experimental results show that the new method produces better quality images while requiring less computing time.</div><div><br></div><div>Part 2 of the thesis considers the threshold design problem of a QIS. A spatially-varying threshold can significantly improve the reconstruction quality and the dynamic range. However, no known method of how to achieve this can be found in the literature. The theoretical analysis of this part shows that the optimal threshold should match with the underlying pixel intensity. In addition, the analysis proves the existence of a set of thresholds around the optimal threshold that give asymptotically unbiased reconstructions. The asymptotic unbiasedness has a phase transition behavior. A new threshold update scheme based on this idea is proposed. Experimentally, the new method can provide good estimates of the thresholds with less computing budget compared to existing methods.</div><div><br></div><div>Part 3 of the thesis extends QIS capabilities to color imaging by studying how a color filter array should be designed. Because of the small pixel pitch of QIS, crosstalk between neighboring pixels is inevitable and should be considered when designing the color filter arrays. However, optimizing the light efficiency while suppressing aliasing and crosstalk in a color filter array are conflicting tasks. A new optimization framework is proposed to solve the problem. The new framework unifies several mainstream design criteria while offering generality and flexibility. Extensive experimental comparisons demonstrate the effectiveness of the framework.</div>

La recherche en biométrie a connu une grande évolution durant les dernières annéessurtout avec le développement des méthodes de décomposition de visage. Cependant,ces méthodes ne sont pas robustes particulièrement dans les environnements incontrôlés.Pour faire face à ce problème, l'imagerie multispectrale s'est présentée comme une nouvelletechnologie qui peut être utilisée en biométrie basée sur la reconnaissance de visage.Dans tous ce processus, la qualité des images est un facteur majeur pour concevoirun système de reconnaissance fiable. Il est essentiel de se disposer d'images de hautequalité. Ainsi, il est indispensable de développer des algorithmes et des méthodes pourl'amélioration des données sensorielles. Cette amélioration inclut plusieurs tâches tellesque la déconvolution des images, le defloutage, la segmentation, le débruitage. . . Dansle cadre de cette thèse, nous étudions particulièrement la suppression de bruit ainsi quela segmentation de visage.En général, le bruit est inévitable dans toutes applications et son élimination doit sefaire tout en assurant l'intégrité de l'information confinée dans l'image. Cette exigenceest essentielle dans la conception d'un algorithme de débruitage. Le filtre Gaussienanisotropique est conçu spécifiquement pour répondre à cette caractéristique. Nous proposonsd'étendre ce filtre au cas vectoriel où les données en disposition ne sont plus desvaleurs de pixels mais un ensemble de vecteurs dont les attribues sont la réflectance dansune longueur d'onde spécifique. En outre, nous étendons aussi le filtre de la moyennenon-local (NLM) dans le cas vectoriel. La particularité de ce genre de filtre est la robustesseface au bruit Gaussien.La deuxième tâche dans le but d'amélioration de données sensorielles est la segmentation.Le clustering est l'une des techniques souvent utilisées pour la segmentation etclassification des images. L'analyse du clustering implique le développement de nouveauxalgorithmes particulièrement ceux qui sont basés sur la méthode partitionnelle.Avec cette approche, le nombre de clusters doit être connu d'avance, chose qui n'est pastoujours vraie surtout si nous disposons de données ayant des caractéristiques inconnues.Dans le cadre de cette thèse, nous proposons de nouveaux indices de validationde clusters qui sont capables de prévoir le vrai nombre de clusters même dans le cas dedonnées complexes.A travers ces deux tâches, des expériences sur des images couleurs et multispectrales sontréalisées. Nous avons utilisé des bases de données d'image très connues pour analyserl'approche proposée. / In this thesis, we focus on multispectral image for face recognition. With such application,the quality of the image is an important factor that affects the accuracy of therecognition. However, the sensory data are in general corrupted by noise. Thus, wepropose several denoising algorithms that are able to ensure a good tradeoff betweennoise removal and details preservation. Furthermore, characterizing regions and detailsof the face can improve recognition. We focus also in this thesis on multispectral imagesegmentation particularly clustering techniques and cluster analysis. The effectiveness ofthe proposed algorithms is illustrated by comparing them with state-of-the-art methodsusing both simulated and real multispectral data sets.

Neste trabalho são apresentadas, complementadas e melhoradas duas técnicas de restauração de imagens: uma abordando o problema de retoque digital/remoção de objetos enquanto a segunda é direcionada ao problema deneliminação de ruído. Em ambas as técnicas, a ideia é trabalhar com imagens contendo texturas e outras características de interesse para um observador humano como a preservação de padrões, bordas, estruturas e regiões de natureza oscilatória. A técnica descrita sobre retoque digital de imagens combina difusão anisotrópica, síntese de texturas, busca dinâmica e um novo termo empregado no mecanismo de atribuição da ordem de prioridade durante o processo de reconstrução. Assim, dada uma imagem com regiões a serem recompostas, uma técnica de difusão anisotrópica é aplicada à imagem afim de se obter um mapa de saliência contendo bordas, estruturas e demais informações de baixa frequência da imagem. Na sequência, um mecanismo de prioridade baseado em um novo termo de confiabilidade regularizado é calculado a partir da combinação do mapa anteriormente gerado com a equação do transporte. Tal mecanismo é utilizado para determinar a ordem de preenchimento das partes faltantes da imagem. Para essa tarefa, a abordagem apresentada utiliza uma nova medida de similaridade entre blocos de pixels(amostrados dinamicamente para acelerar o processo), afim de encontrar os melhores candidatos a serem alocados nas regiões danificadas. A técnica destinada à remoção de ruídos alia a teoria da difusão anisotrópica, técnicas de análise harmônica e modelos numéricos de discretização de EDPs não-lineares em uma equação diferencial parcial regularizada, a qual atua de forma incisiva em regiões mais homogêneas da imagem e de forma mais suave em regiões caracterizadas como textura e bordas, preservando, assim, essas regiões. Além da natureza anisotrópica, a EDP procura recompor partes texturizadas perdidas no processo de eliminação de ruído através da aplicação de técnicas robustas de análise harmônica. Uma validação teórica e experimental para esta EDP e um estudo do ajuste paramétrico do método de eliminação de ruído baseado nesta EDP foram realizados neste trabalho. A eficiência e a performance das técnicas propostas são atestadas por meio das análises experimentais quantitativas e qualitativas com outras abordagens clássicas da literatura. / In this work two techniques of image restoration are presented, complemented and improved: one approaching the problem of image inpainting/object removal problem while the second one dealing with the image denoising problem. In both cases, the core idea is to process images containing textures and other features perceptible to a human observer such as patterns, contours, structures and oscillatory information. The image inpainting technique combines anisotropic diffusion, texture synthesis, dynamic search and a mechanism to set the order of priority during the image completion process. More precisely, given an image and target region to be inpainted, an anisotropic diffusion technique is applied in order to generate a saliency map containing edges, structures and other low frequency parts of the image. Next, apriority mechanism based on a new biased confidence term is computed from the map previously generated with the transport equation to define the level of priority of the pixels during the filling procedure. To accomplish this task, the presented approach employs a novel measure of similarity wich measures the distance between blocks of pixels (sampled dynamically to speed up the process) in order to find the best candidates to be allocated in the damaged regions. The technique devoted to denoising an image combines the theory of anisotropic diffusion, harmonic analysis techniques and numerical models into a regularized partial differential equation, which diffuses the pixels more incisively on homogeneous regions of the image while still seeking to attenuate regions formed by textures and patterns, thus preserving those information. Moreover, the proposed PDE aims at recovering texturized regions which have been degraded during the denoising process by employing harmonic analysis tools. A theoretical and experimental validation for this EDP and a study of the parametric adjustment of the image denoising method based on this EDP were performed in this work. The effectivenss and performance of the proposed approaches are attested through a comprehensive set of comparisons against other representative techniques in the literature.

Imagens digitais são utilizadas para diversas finalidades, variando de uma simples foto com os amigos até a identificação de doenças em exames médicos. Por mais que as tecnologias de captura de imagens tenham evoluído, toda imagem adquirida digitalmente possui um ruído intrínseco a ela que normalmente é adquirido durante os processo de captura ou transmissão da imagem. O grande desafio neste tipo de problema consiste em recuperar a imagem perdendo o mínimo possível de características importantes da imagem, como cantos, bordas e texturas. Este trabalho propõe uma abordagem baseada em um Algoritmo Genético Híbrido (AGH) para lidar com este tipo de problema. O AGH combina um algoritmo genético com alguns dos melhores métodos de supressão de ruídos em imagens encontrados na literatura, utilizando-os como operadores de busca local. O AGH foi testado em imagens normalmente utilizadas como benchmark corrompidas com um ruído branco aditivo Gaussiano (N; 0), com diversos níveis de desvio padrão para o ruído. Seus resultados, medidos pelas métricas PSNR e SSIM, são comparados com os resultados obtidos por diferentes métodos. O AGH também foi testado para recuperar imagens SAR (Synthetic Aperture Radar), corrompidas com um ruído Speckle multiplicativo, e também teve seus resultados comparados com métodos especializados em recuperar imagens SAR. Através dessa abordagem híbrida, o AGH foi capaz de obter resultados competitivos em ambos os tipos de testes, chegando inclusive a obter melhores resultados em diversos casos em relação aos métodos da literatura. / Digital images are used for many purposes, ranging from a simple picture with friends to the identification of diseases in medical exams. Even though the technology for acquiring pictures has been evolving, every image digitally acquired has a noise intrinsic to it that is normally gotten during the processes of transmission or capture of the image. A big challenge in this kind of problem consists in recovering the image while losing the minimum amount of important features of the image, such as corners, borders and textures. This work proposes an approach based on a Hybrid Genetic Algorithm (HGA) to deal with this kind of problem. The HGA combines a genetic algorithm with some of the best image denoising methods found in literature, using them as local search operators. The HGA was tested on benchmark images corrupted with an additive white Gaussian noise (N;0) with many levels of standard deviation for the noise. The HGAs results, which were measured by the PSNR and SSIM metrics, were compared to the results obtained by different methods. The HGA was also tested to recover SAR (Synthetic Aperture Radar) images that were corrupted by a multiplicative Speckle noise and had its results compared against the results by other methods specialized in recovering with SAR images. Through this hybrid approach, the HGA was able to obtain results competitive in both types of tests, even being able to obtain the best results in many cases, when compared to the other methods found in the literature.