Multi-frame super-resolution reconstruction via kernel regression regularized sparse learning

Abstract

A novel kernel regression regularized adaptive sparse (KR-RAS) model is presented in this paper for multi-frame super-resolution (SR) reconstruction, by incorporating KR estimation and the clustering-based dictionary learning into a unified sparse reconstruction framework. The basic idea behind our model is to exploit both the global structural self-similarity throughout all frames as prior constraints, and sparsity constraints, to regularize the ill-posed reconstruction for better estimation. In the proposed method, normalized steering kernels are introduced as features for structural clustering of image patches, to aggregate more structurally similar patches for dictionary learning. Furthermore, KR estimation is extended from local neighborhood to the global neighborhood that is constituted by similar patches from any position of all frames, so more accurate regression estimation of pixel values is possible. Extensive comparisons of experimental results on real video sequences show that the performance of the proposed method outperforms the state-of-the-art methods both subjectively and objectively in most cases.