A Bayesian Multiframe Superresolution Iterative Method Using SSIM-based Convergence Suitable for Real Cases
Thais Pedruzzi do Nascimento1; Evandro Ottoni Teatini Salles1
1 Universidade Federal do Espírito Santo
Resumo
Multiframe superresolution is a method that generates a higher resolution image (HR estimated image) from several lower resolution images (LR images). A Bayesian approach is convenient for such purpose because it is possible to incorporate pre-known information in the form of prior. This work implements an SAR prior, which is better for softening noise, and a l1-norm and a TV prior, both known for the maintenance of edges. Besides, TVSAR and l1SAR are also implemented in order to take advantage of both possibilities. Iterative methods, namely Majorization Minimization and Expectation Maximizations, are used to manage the non-analytical nature of the l1 priors and adjust the motion parameters iteratively. SSIM and PSNR-based convergence conditions are used and compared. Previous works use full reference image quality metrics (FR-IQA) to choose the output image, which makes them unsuitable for real cases. This work investigates a direct form of choice of the estimated image, among the ones of all iterations. Such choice enables the application of our algorithm to real cases, where the HR original image is unavailable.
Palavras-chave: Multiframe Superresolution; Priors; SSIM; Convergence condition