Super-Resolution Based on Clustered Examples

Abstract

In this paper, we propose an improved version of the neighbor embedding super-resolution (SR) algorithm proposed by Chang et al. [Super-resolution through neighbor embedding, in Proc. 2004 IEEE Computer Society Conf. Computer Vision and Pattern Recognition(CVPR), Vol. 1 (2004), pp. 275–282]. The neighbor embedding SR algorithm requires intensive computational time when finding the K nearest neighbors for the input patch in a huge set of training samples. We tackle this problem by clustering the training sample into a number of clusters, with which we first find for the input patch the nearest cluster center, and then find the K nearest neighbors in the corresponding cluster. In contrast to Chang’s method, which uses Euclidean distance to find the K nearest neighbors of a low-resolution patch, we define a similarity function and use that to find the K most similar neighbors of a low-resolution patch. We then use local linear embedding (LLE) [S. T. Roweis and L. K. Saul, Nonlinear dimensionality reduction by locally linear embedding, Science 290(5500) (2000) 2323–2326] to find optimal coefficients, with which the linear combination of the K most similar neighbors best approaches the input patch. These coefficients are then used to form a linear combination of the K high-frequency patches corresponding to the K respective low-resolution patches (or the K most similar neighbors). The resulting high-frequency patch is then added to the enlarged (or up-sampled) version of the input patch. Experimental results show that the proposed clustering scheme efficiently reduces computational time without significantly affecting the performance.