Improving Super-Resolution Reconstruction with Regularized Extreme Learning Machine Networks

The primary objective of Super-Resolution reconstruction methods is, given an input low resolution image, to estimate a high resolution image, with more pixels, without losing high frequency components, as what happens in simple interpolation methods. To this end, we propose the learning of multiple Regularized Extreme Learning Machines networks, using patches of a dataset composed by low and high resolution images. Each patch in the training database is associated to a cluster, based on gradient information presented in the patches, and a network is trained for each cluster. The proposed method generates similar or better super-resolution reconstruction results than similar state of the art algorithms, based on PSNR and SSIM metrics, while using a training dataset more than two orders of magnitude smaller, showing the learning power of the proposed method.