Image super-resolution based on multi-scale self-similarity and gradient constraint

In this study, we proposed a single image superresolution based on multi-scale self-similarity and gradient constraint. Instead of depending on an external set of example images required in most learning-based super-resolution algorithms, the proposed algorithm exploit self-similarity and generates example image pairs directly from the image pyramid of the input low resolution image. The algorithm is patch-based. Therefore, these images including input low resolution image are divided into patches firstly. Gradient feature of the patches is extracted. Then, similar example patch pairs are searched for each input patch. The gradient feature of high resolution patch can be estimated by weight of the similar patches. At last, the solution model constrained by gradient is set up, which can solve the weight and estimate high resolution image patch by alternated iteration. The model can reduces the influence of weight estimation error on performance of the proposed algorithm. Experimental results demonstrate that the proposed algorithm performs well.