Multi-frame image super-resolution reconstruction based on GPOF registration and L1-norm

Abstract

We propose a robust approach of producing a high-resolution (HR) image from a sequence of low-resolution (LR), blurred and noisy images. In the proposed approach, we specifically focus on the motion model of Gaussian Pyramid Optical Flow (GPOF) registration which achieves the sub-pixel precision and enables large pixel motions, while keeping the size of the integration window relatively small. In the process of super-resolution reconstruction, our method is based on the use of L1-norm both in the measurement term and the regularization term called Bilateral Total Variation (BTV) as the prior model to penalize high spatial frequency signals and preserve edges. Specially, we introduce the Median "shift and add" idea to initialize the HR image value in the iterative steps for the optimization of the objective function, when the motions between LR frames are pure translations and the blur is space invariant.