An Adaptive Learning Algorithm For Image Super-resolution

Super-resolution image reconstruction is a resolution enhancement technology that extracts higher resolution images containing more details from an image sequence of lower resolution by using digital signal processing technology, and the existing low-resolution imaging systems can be still utilized. As a technology of improving image quality significantly without changing the hardware device, super-resolution image reconstruction has a wide range of applications and has become one of the hottest image restoration research topics in the world.As an emerging imaging image modeling technique, sparse representation has been successfully used in various image restoration applications. The image restoration quality largely depends on whether the employed sparse domain can represent effectively the underlying image. In this paper, we present an adaptive learning algorithm for image super-resolution. Its implementation is guaranteed by the combination of adaptive sparse representation super-resolution (ASR_SR) and the adaptive regularization terms. We propose to learn various sets of bases from a pre-collected dataset of high quality example image patches, and then for a given patch to be processed, one or several sets of bases are adaptively selected to characterized the local sparse domain. Furthermore, a set of piecewise autoregressive (PAR) models are learned from the dataset of example image patches. For a given image patch, the best fitted PAR models are selected to regularize the image local structures. Then the image non-local self-similarity (NLSS) is introduced as another regularization term. The two additional regularization models are very helpful in enhancing image edges and suppressing noises, which can further increase the image restoration performance.Extensive experiments on image super-resolution validated that by using adaptive sparse domain selection and the adaptive regularization, the proposed adaptive learning method achieves much better results than the state-of-the-art algorithms in terms of PSNR, SSIM and visual perception.