Single Image Super-Resolution Using Multiple Extreme Learning Machine Regressors

This paper presents a new technique to solve the single image super resolution reconstruction problem based on multiple extreme learning machine regressors, called here MELM. The MELM employs a feature space of low resolution images, divided in subspaces, and one regressor is trained for each one. In the training task, we employ a color dataset containing 91 images, with approximately 5.3 million pixels, and PSNR and SSIM as metric evaluation. For the experiments we use two datasets, Set 5 and Set 14, to evaluate the results. We observe MELM improves reconstruction quality in about 0.44 dB PSNR in average for Set 5, when compared with a global ELM regressor (GELM), trained for the entire feature space. The proposed method almost reaches deep learning reconstruction quality, without depending on large datasets and long training times, giving a competitive trade off between performance and computational costs.