Robust Multiframe Super-Resolution with Adaptive Norm Choice Using Difference Curvature Based BTV Regularization

Abstract

Multi-frame image super-resolution focuses on reconstructing a high-resolution image from a set of low-resolution images with high similarity. Since super-resolution is an ill-posted problem, regularization techniques are widely used to constrain the minimization function. Combining image prior knowledge with fidelity model, Bayesian-based methods can effectively solve this ill-posed problem, which makes this kind of methods more popular than other methods. Our proposed model is based on maximum a posteriori probability (MAP) estimation. In this thesis, we propose a novel initialization method based on median operator to initialize our estimated high-resolution image. For the fidelity term in our proposed algorithm, the half-quadratic estimation is used to choose error norm adaptively instead of using fixed L1 or L2 norm. Furthermore, for our regularization term, we propose a novel regularization method based on Difference Curvature (DC) and Bilateral Total Variation (BTV) to suppress mixed noises and preserve image edges simultaneously. In our experimental results, synthetic data and real data are both tested to demonstrate the superiority of our proposed method in terms of clearer texture and less noise over other state-of-the-art methods.