Contribuições em Super-Resolução Multi-Frame Bayesiana Utilizando Informação Local e Registro Não Paramétrico

• Ano de defesa: 2021
• Autor(a) principal: Nascimento, Thaís Pedruzzi do
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Super-resolução multi-frame regularizada; registro demons; extreme learning machine; super-resolução local; subject.br-rjbn; Engenharia Elétrica
• Link de acesso: http://repositorio.ufes.br/handle/10/15101

Regularized multi-frame super-resolution aims to generate a high-resolution image from several low-resolution images, from the alignment of them and the use of regularization, which weighs noise suppression and edge information restoration. In this context, three approaches are proposed in this thesis, using local information and Demons registration: the hybrid approach, Demons super-resolution, and the patch-based approach. Regarding the hybrid approach, multiple single-layer neural networks were trained by the Regularized Extreme Learning Machine algorithm and implemented jointly with the multi-frame super-resolution methods Bilateral Total Variation (BTV) and Iterative Re-Weighted Super-Resolution (IRWSR), resulting in HyBTV and HyIRWSR, respectively. For Demons super-resolution two methods were proposed: D-BTVIR and D-IRWIR, which combine Demons registration with BTV-based an IRWSR-reconstruction, respectively. And for the patch-based approach (PB), which consists of aligning the patches, scan and super-resolve them individually, three methods were proposed: PB - no classification, PB - smoothness, and PB - variance. For PB - no classification all patches are super-resolved by IRW, which is based on IRWSR. For PB - smoothness, the patches are classified as homogeneous or not, by using smoothness as metrics, and super-resolved via bicubic interpolation, if homogeneous or via IRW, otherwise. PB - variance is based on the same procedure, but considering variance as the homogeneity metric. Experiments were conducted using simulated deformation in 119 images, from Set5, Se14 and B100. PSNR, SSIM and the execution time were analyzed by the use of Friedman, Nemenyi and Wilcoxon hypothesis tests, besides visual analysis. The Wilcoxon tests suggested a better performance of HyBTV over BTV with 99% reliability (p-value = 1.26 × 10−11), and of HyIRWSR over IRWSR (p−value = 3.17 × 10−6), besides better time-quality trade-off of HyBTV over IRWSR, which is, on average, 3.6 times slower than the first one. For the Demons-based approach, the Wilcoxon test suggested better performance of D-BTVIR over BTV with p-value = 3.52 × 10−7, and the same thing was noted considering D-IRWIR over IRWSR, with p-value 2.95×10−8. The Nemenyi test suggested statistical equivalence between D-BTVIR and IRWSR, however, D-BTVIR was 7.2 times faster. Nemenyi test also suggested better performance of the three variations of PB, when compared to BTV and IRWSR, besides statistical equivalence with each other, with PB - variance being the fastest one of them. Finally, the visual analysis supported the results from the hypothesis tests.