Submitted by Milena Rubi ( ri.bso@ufscar.br) on 2017-10-17T16:07:35Z No. of bitstreams: 1 CUNHA_Bruno_2017.pdf: 1281922 bytes, checksum: 605ecd45f46a3b67332ed6bd13043af5 (MD5) / Approved for entry into archive by Milena Rubi ( ri.bso@ufscar.br) on 2017-10-17T16:07:44Z (GMT) No. of bitstreams: 1 CUNHA_Bruno_2017.pdf: 1281922 bytes, checksum: 605ecd45f46a3b67332ed6bd13043af5 (MD5) / Approved for entry into archive by Milena Rubi ( ri.bso@ufscar.br) on 2017-10-17T16:07:53Z (GMT) No. of bitstreams: 1 CUNHA_Bruno_2017.pdf: 1281922 bytes, checksum: 605ecd45f46a3b67332ed6bd13043af5 (MD5) / Made available in DSpace on 2017-10-17T16:08:04Z (GMT). No. of bitstreams: 1 CUNHA_Bruno_2017.pdf: 1281922 bytes, checksum: 605ecd45f46a3b67332ed6bd13043af5 (MD5) Previous issue date: 2016-11-10 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / This work proposes both a study and a development of an algorithm for super-resolution of digital images using projections onto convex sets. The method is based on a classic algorithm for spatial super-resolution which considering the subpixel information present in a set of lower resolution images, generate an image of higher resolution and better visual quality. We propose the incorporation of a new restriction based on the Richardson-Lucy algorithm in order to restore and recover part of the spatial frequencies lost during the degradation and decimation process of the high resolution images. In this way the algorithm provides a hybrid approach based on projections onto convex sets which is capable of promoting both the spatial and spectral image super-resolution. The proposed approach was compared with the original algorithm from Sezan and Tekalp and later with a method based on a robust framework that is considered nowadays one of the most effective methods for super-resolution. The results, considering both the visual and the mean square error analysis, demonstrate that the proposed method has great potential promoting increased visual quality over the images studied. / Este trabalho visa o estudo e o desenvolvimento de um algoritmo para super-resolução de imagens digitais baseado na teoria de projeções sobre conjuntos convexos. O método é baseado em um algoritmo clássico de projeções sobre restrições convexas para super- resolução espacial onde se busca, considerando as informações subpixel presentes em um conjunto de imagens de menor resolução, gerar uma imagem de maior resolução e com melhor qualidade visual. Propomos a incorporação de uma nova restrição baseada no algoritmo de Richardson-Lucy para restaurar e recuperar parte das frequências espaciais perdidas durante o processo de degradação e decimação das imagens de alta resolução. Nesse sentido o algoritmo provê uma abordagem híbrida baseada em projeções sobre conjuntos convexos que é capaz de promover simultaneamente a super-resolução espacial e a espectral. A abordagem proposta foi comparada com o algoritmo original de Sezan e Tekalp e posteriormente com um método baseado em um framework de super-resolução robusta, considerado um dos métodos mais eficazes na atualidade. Os resultados obtidos, considerando as análises visuais e também através do erro médio quadrático, demonstram que o método proposto possui grande potencialidade promovendo o aumento da qualidade visual das imagens estudadas.

This thesis deals with image colorization and image super-resolution using neural networks. It briefly explains neural networks principles and summarizes current approaches in this domain. It also describes the design, implementation and training of various neural network architectures. The best implemented architecture can colorize images, in particular, works well with outdoor areas. The architecture for image super-resolution with residual blocks that was trained with a perceptual loss function performs a double increase in image resolution (4x more pixels in total). Part of this thesis is also an implementation of a web application that uses trained models for image modification.

This diploma thesis deals with image super-resolution with conservation of good quality. Firstly, there are described state of the art methods dealing with this problem, as well as principles of neural networks with focus on convolutional ones. Finally, there is described a few models of convolutional neural network for image super-resolution to double size, which have been trained, tested and compared on newly created database with pictures of people.

This master's thesis deals with methods of increasing the image resolution. It contens as a description of theoretical principles and description of calculations which are wellknown nowdays and are usually used for increasing image resolution both description of new methods which are used in this area of image procesing. It also contens a method which I suggested myself. There is also a description of methods for an evaluation of image similarity and a comparation of results from methods which are described in this thesis. This thesis includes implementations of selected methods in programming language MATLAB. It was created an application, which realizes some methods of increasing image and evaluate their results relation to the original image using PSNR and SSIM index.

Generative Adversarial Networks have been used to obtain state-of-the-art results for low-level computer vision tasks like single image super-resolution, however, they are notoriously difficult to train due to the instability related to the competing minimax framework. Additionally, traditional ensembling mechanisms cannot be effectively applied with these types of networks due to the resources they require at inference time and the complexity of their architectures. In this thesis an alternative method to create ensembles of individual, more stable and easier to train, models by using interpolations in the parameter space of the models is found to produce better results than those of the initial individual models when evaluated using perceptual metrics as a proxy of human judges. This method can be used as a framework to train GANs with competitive perceptual results in comparison to state-of-the-art alternatives. / Generative Adversarial Networks (GANs) har använts för att uppnå state-of-the- art resultat för grundläggande bildanalys uppgifter, som generering av högupplösta bilder från bilder med låg upplösning, men de är notoriskt svåra att träna på grund av instabiliteten relaterad till det konkurrerande minimax-ramverket. Dessutom kan traditionella mekanismer för att generera ensembler inte tillämpas effektivt med dessa typer av nätverk på grund av de resurser de behöver vid inferenstid och deras arkitekturs komplexitet. I det här projektet har en alternativ metod för att samla enskilda, mer stabila och modeller som är lättare att träna genom interpolation i parameterrymden visat sig ge bättre perceptuella resultat än de ursprungliga enskilda modellerna och denna metod kan användas som ett ramverk för att träna GAN med konkurrenskraftig perceptuell prestanda jämfört med toppmodern teknik.

[pt] A indústria mineral vem presenciando, ao longo das últimas décadas, uma redução da qualidade de minério de ferro extraído e o surgimento de novas demandas ambientais. Esta conjuntura fortalece a busca por produtos provenientes do minério de ferro que atendam aos requisitos da indústria siderúrgica, como é o caso de novos aglomerados de minério de ferro. A Microtomografia de Raios-X (microCT) permite a caracterização da estrutura tridimensional de uma amostra, com resolução micrométrica, de forma não-destrutiva. Entretanto, tal técnica apresenta diversas limitações. Quanto melhor a resolução, maior o tempo de análise e menor o volume de amostra adquirido. Modelos de Super Resolução (SR), baseados em Deep Learning, são uma poderosa ferramenta para aprimorar digitalmente a resolução de imagens tomográficas adquiridas em pior resolução. Este trabalho propõe o desenvolvimento de uma metodologia para treinar três modelos de SR, baseados na arquitetura EDSR, a partir de imagens tomográficas de briquetes de redução direta: Um modelo para aumento de resolução de 16 um para 6 um, outro para aumento de 6 um para 2 um, e o terceiro para aumento de 4 um para 2 um. Esta proposta tem como objetivo mitigar as limitações do microCT, auxiliando o desenvolvimento de novas metodologias de Processamento Digital de Imagens para os aglomerados. A metodologia inclui diferentes propostas para avaliação do desempenho da SR, como comparação de PSNR e segmentação de poros. Os resultados apontam que a SR foi capaz de aprimorar a resolução das imagens tomográficas e mitigar ruídos habituais da tomografia. / [en] The mining industry has been witnessing a reduction of extracted iron ore s quality and the advent of new environmental demands. This situation reinforces a search for iron ore products that meet the requirements of the steel industry, such as new iron ore agglomerates. X-ray microtomography (microCT) allows the characterization of a sample s three-dimensional structure, with micrometer resolution, in a non-destructive analysis. However, this technique presents several limitations. Better resolutions greatly increase analysis time and decrease the acquired sample’s volume. Super-Resolution (SR) models, based on Deep Learning, are a powerful tool to digitally enhance the resolution of tomographic images acquired at lower resolutions. This work proposes the development of a methodology to train three SR models, based on EDSR architecture, using tomographic images of direct reduction briquettes: A model for enhancing the resolution from 16 um to 6 um, another for enhancing from 6 um to 2 um, and the third for enhancing 4 um to 2 um. This proposal aims to mitigate the limitations of microCT, assisting the development and implementation of new Digital Image Processing methodologies for agglomerates. The methodology includes different proposals for SR s performance evaluation, such as PSNR comparison and pore segmentation. The results indicate that SR can improve the resolution of tomographic images and reduce common tomography noise.