Abstract: Local and Nonlocal image representations have shown great potential in low-level vision tasks leading to several state-of-the-art image restoration techniques. Both of these representations have their own advantages. The work towards combining these two representations seems to be minimal. The paper tries to contribute to the area of unification of these two representations. Ahybrid approach for image restoration has been proposed in this paper that combines both of these representatins. The main idea behind this approach is singular value decomposition (SVD), a bilateral variance estimation perspective. SVD of similar patches has the property of pooling both local and non-local information for estimating signal variances. This, in-turn, has led to the development of new class of image restoration algorithms. For noisy data, the algorithm makes use of iterative regularization concept; for incomplete data, it makes use of deterministic annealing-based solution along with dictionary learning. The performance of this hybrid approach will have the results that can be compared favorably with other leading image restoration algorithms.

Keywords: Deterministic annealing, iterative regularization, singular value thresholding, singular value decomposition, image denoising, image completion, patch clustering.