An Example Based Non-local Mean Regularization For Image Super-resolution

Image super-resolution (SR) is a very useful technique for visual surveillance, high-definition TV and medical image processing. In the traditional method, the inferred high-resolution image patch is represented as a linear combination of dictionaries obtained by training images, and least square estimation or sparse regularization is used to find better solution. However, the methods proposed so far neglect the prior knowledge about local structure similarity in natural images. In this paper we introduce an adaptive regularization terms into LLE based SR framework. First, an example based image nonlocal mean regularization term is learned from the dataset of example image patches, which captures the local structure similarity between the input image and training images. Then, the image nonlocal self-similarity is used as another regularization term. In addition, we propose an iterative optimization framework to find the latent HR image. Experimental results on real surveillance images demonstrate the superiority of the proposed method over some state-of-the-art image SR approaches.