Processing and representation of multispectral images using deep learning techniques

Abstract

Computer vision is a scientific discipline that has been developed in recent decades due to technological advances in acquisition devices together with the increase on computational capabilities. The reduction in prices of hardware, both acquisition and processing, allows this technology to be available to most users. Additionally, there is a technological advance that allows sensors to be sensitive to different spectra, including smart mobile devices. Computer vision is defined as a field of study that develops multiple techniques to ensure that machines can "see" and "understand" information in images or videos of any spectrum, using mathematical models that process, analyze and interpret digital information extracted from images. With the advance of convolutional neural networks (CNN), the usage of machine learning based techniques has made great progress in recent years. Specifically, many techniques have been developed to implement a process similar to the visual reasoning of human vision, to performtasks such as detection, recognition, segmentation, coloring, filtering, improvement, similarity, etc. using CNN. This thesis presents a series of CNN-based techniques applied to images of different spectra, especially the near infrared spectrum (NIR) and the visible spectrum. Among the techniques implemented are: perform similarity detection between images of VISIBLE and NIR spectra, colorization of NIR images, estimation of normalized difference vegetation index (NDVI) using only one band of the spectrum and eliminate the haze present in the images. It should be noted that to implement these techniques, generative adversarial models have been used in their standard, conditional, stacked and cyclic variants, which are the latest generation in these type of networks.