Single image super resolution using neighbor embedding and statistical prediction model

This paper proposes learning based approaches for single image super-resolution using sparse representation and neighbor embedding. Two learning based methods are proposed to recover the high-resolution (HR) image patches from the low resolution (LR) patches. The first method, named as LeNm-SRI, is a computationally efficient approach using neighbor embedding in a partitioned feature space. In this method, the training set is updated by including details extracted from different scales of LR input image. LeNm-SRI, which uses sparse representation invariance, gives acceptable results at low computational load. In the second approach, named as LeNm-RBM, a statistical prediction model is used to predict HR feature coefficients to obtain increased performance. Separate prediction models are trained for each cluster, and the model parameters are updated with each input image, to adapt to input test image. Experimental results validate the computational efficiency and performance of the proposed methods.