Example-Based Super-Resolution using Locally Linear Embedding

Example-Based Super-Resolution is a learning-based technique that attempts to recover high-resolution (HR) image according to the corresponding relation in a set of training low-resolution (LR) and high-resolution image pairs prepared in advance. The conventional learning-based method for image super-resolution usually cannot achieve the high-frequency components accurately, which are lost in the input LR image, for recovering the HR image, since it only estimates the lost information using one most similar training LR patch to the input patch, and its corresponding HR pair. Therefore, we propose to use a manifold learning method- Locally Linear Embedding (LLE) for reconstructing the input LR patch with a linear weight summation of its several most similar training LR patches, and then can recover HR patch using the same linear summation of the corresponding training HR patches. Furthermore, in order to solve the expensive computational problem in the conventional exampled-based learning method, only the patches with larger variance, which means with high-frequency components, are selected for super-resolution procedures. Finally, Experimental results show that the recovered high-resolution images by our proposed approach are much better than those by conventional method and interpolation techniques.