Prior de regularização para problema de demosaicing com aplicação em CFA’s variados

Abstract

This research presents a new proposal to Demosaicing algorithms, using a more flexible approach to deal with the Color filter array (CFA) in single sensor color imaging. The proposed algorithm is structured in the inverse problems model, whose functions employ a CFA adaptive matrix-vector operational model. From this concept, the Demosaicing problem is treated as a cost function minimization with two terms, one referring to the dependence between the estimation and the data provided by the acquisition model, and other term related to features observed in images, which can be explored to form a more precise estimation, this last term is known as Prior. The established proposal is applied in regularization algorithms with focus on the high correlation among color channels (R, G, and B), and in the local smoothness of uniform regions. Both characteristics organize the Prior employed in this work. The minimization proposed is iteratively achieved through IRLS-CG, which is the combination of two efficient minimization algorithms, that presents quick responses, and the capacity to deal with L1 and L2 norm at the same time. The quality of the proposed algorithm is verified in an experiment in which varous CFA were used and a situation with 35dB gaussian noise and another one with no noise applied to the Kodak dataset, and the results were compared with state-of-the-art algorithms, in which the performance of the proposed Prior showed excellent results, including when the CFA is different from Bayer’s, which is the most commonly used pattern.