Resumen:
In recent years, there has been an increasing demand for high-resolution images, especially in the field of security and surveillance. Super-resolution is a technique that can be used to improve the resolution of an image. Most of these techniques are based on using a single image or a set of low-resolution images from the visible spectrum, where the high-resolution image is reconstructed by using a model that considers a degradation process. Nevertheless, images from the visible spectrum are limited by the atmospheric conditions and the availability of light. While human visual perception is limited to the visual-optical spectrum, machine vision isnot.This dissertation presents the use of images from the long-wavelength infrared spectral band, namely thermal images, for the purpose of super-resolving them. Thermal images are notaffectedbyatmosphericconditions,andtheycanbeacquiredeveninlow-lightconditions. In order to obtain a high-resolution image from a set of low-resolution thermal images, deep learning techniques are used, specifically convolutional neural networks. The results show that improving the thermal images’ resolution is possible while preserving the scene’s main features. Two main paths are tackled in the present work, the single and multi-image super-resolution, where a dataset with an extensive collection of images is collected to address this purpose. One of the main properties of this thesis is to show that thermal image super-resolution can be tackled by using the proposed architectures and validating them with the acquired public dataset used in several challenges.
