Image Super Resolution with Direct Mapping and De-noising

Image super resolution attempts to extract high-resolution image using one or more corrupted low resolution images. Typical patch-wise sparse dictionary based super resolution reconstruction methods remove undesired effects but sometimes lead to blurring due to averaging of many high resolution (HR) patches. On the other hand, local self-similarity based super resolution (SR) suffers from insufficient number of HR patches and can result in distorted and unnatural SR images. In this paper, we propose a novel single image super resolution approach that reconstructs high resolution images by leveraging direct mapping i.e. one-to-one mapping between low resolution and high resolution patches. In addition, high frequency content is separated and preserved in the SR reconstructed image. Further, a K-D tree classification and knn-search algorithm is used for fast and robust search by dimensions. Incorporation of Non-Local Means filtering reduces unwanted noise as well as undesired artifacts. Finally, the proposed Gaussian weighting scheme reduces error in HR patch reconstruction process. The proposed approach is also robust for larger magnification factor beyond 2.