ℓ1-graph based local regression for super-resolution

Abstract

Example-based methods are popular in the single-image super-resolution technology. Among these methods, nearest neighbor-based algorithms are attractive for their simplicity and flexibility. These algorithms are mostly designed based on the nearest neighbor estimation, which has been shown very poor in generalization according to leaning theories. The weak generalization performance of nearest neighbor estimation lowers the performance of nearest neighbor-based algorithms, in both the visual experience and statistical index. To fix the problem, we introduce a local regression method where the local training sets are adaptively generated by applying the ℓ1-graph to the nearest neighbor-based algorithms. The ℓ1-graph based local regression method improves the generalization performance of nearest neighbor-based estimation, which further enhances the performance of nearest neighbor-based algorithms in super-resolution. The experimental results have shown that, the nearest neighbor-based algorithms are improved by our method.