Reconstrução de imagens por superresolução utilizando inferência Bayesiana aproximada

• Autor(a) principal: Camponez, Marcelo Oliveira
• Data de Publicação: 2012
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal do Espírito Santo (riUfes)
• Texto Completo: http://repositorio.ufes.br/handle/10/9705
• topic: Eletrônica Industrial, Sistemas e Controles Eletrônicos; Processamento de imagens; Teoria bayesiana de decisão estatística; 621.3

Superresolution (SR) multiframe image enhancement are techniques to increase the resolution of an image without increasing the resolution of the camera used to capture them. Image processing algorithms are used, instead. A set of observable lower resolution images is first registered (with respect to one of those images) and subsequently fused into a single high resolution image, which is then filtered to improve their sharpness. In such context, this PhD Thesis proposes five Superresolution algorithms based on Bayesian statistical modeling. In the first algorithm developed, referred to as Closed Form SR, an equation is derived in closed form to calculate the high resolution image. However, the algorithm depends on a parameter λ, which corresponds to the regularization parameter in the spatial SR methods. In the case of this algorithm, however, such parameter should be heuristically set. In the second and third algorithms, referred to as SR INLA and INLA MAP SR, respectively, a nonparametric method to approximate Bayesian inference recently developed, known in the statistical literature as Integrated Nested Laplace Approximation (INLA), was applied to the Superresolution problem. The advantage of such algorithms is that the parameter λ is automatically adjusted. Finally, the last two proposed algorithms, referred to as DFT INLA and SR Closed Form DFT, respectively, are faster versions of the algorithms INLA SR and Closed Form SR, based on 2D Fourier transform. In this new implementation the computational cost, originally O(n 3 ) is reduced to O(n 2 log(n)), with a slight loss of quality of the HR image. Additionally, the size of the arrays manipulated by the algorithm is reduced from n 2 ×n 2 to n×n (the HR image size). Experiments demonstrate that the proposed algorithms are competitive, when compared to other state-of-the-art algorithms.