Super resolução single image utilizando redes extreme learning machine

Abstract

Super-resolution methods aim to increase the spatial resolution of an image while maintain ing the highest possible fidelity between the estimated image and the original source from which that image was taken. The need to increase the resolution of an image arises due to the process of image formation through image acquisition devices. These devices sample an image at fixed rates, depending on their internal hardware, and add blurring and noise effects to the sampled image. To circumvent the low sampling rate of the hardware of these devices, it is more advantageous to develop software solutions capable of increasing the resolution of the images after the capture, through super resolution algorithms. This work proposes a single-image super-resolution algorithm based on machine learning techniques, where an external database is used to learn a model that relates low and high-resolution images. The method is based on multiple regressors in the form of single-hidden-layer feed-forward neural networks trained by extreme learning machine, applied in subspaces of the training set generated by clustering techniques. Pre and post-processing reconstruction techniques and the training of reshaping masks are used to refine the result of reconstruc tion. The proposed method stands out for the low training time and the ability to be used in an ordinary computer, without GPUs and large amounts of RAM, while delivering results that compete with important works in the literature