Variational Bayesian Super Resolution Acceleration Using Preconditioned Conjugate Gradient

The high computational complex of Super Resolution (SR) is a focused topic in many imaging applications, which involves to solve huge sparse linear systems. Solving such systems usually employs the iterative methods, such as Conjugate Gradient (CG). But in most variational Bayesian SR algorithms, CG method converges slowly with the coefficient matrix being ill-conditioned and takes long execution time. In this paper, we propose Preconditioned Conjugate Gradient (PCG) to solve the problem and analyze the performance of the different PCG solvers, Jacobi and incomplete Cholesky decomposition(IC). Experimental results demonstrate that the new method achieves accelerations compared with the traditional one while maintaining high visual quality of the reconstructed HR image, and, especially, the IC solver has a better performance.