Processing and Representation of Multispectral Images Using Deep Learning Techniques

Abstract

This thesis has implemented innovative techniques in the field of computer vision using visible and near-infrared spectrum images, applying deep learning through convolutional networks, especially GANs' architectures, which are specialists in generating information and also includes meta techniques -learning to tackle the problem of determining the similarity of images of a different spectrum. In this research, with this type of convolutional networks, different supervised and unsupervised techniques have been created to solve challenging problems, like detect the similarity of patches of different spectra (visible-infrared), colorized images of the near-infrared spectrum, estimation of vegetation index (NDVI) and the haze removal present on RGB images using NIR images. For all these techniques different variants of the GAN's networks, such as standard, conditional, stacked, and cyclic have been used. Also, a metric-based meta-learning approach has been implemented. It should be mentioned that together with the implementation of adversarial network models, the use of multiple loss functions has been proposed to improve the generalization and increase the effectiveness of the models. The experiments were performed with paired and unpaired images, given the different supervised and unsupervised architectures implemented, respectively. The experimental results obtained in each of the approaches implemented in the doctoral work compared with the techniques of the state of the art were shown to be more effective.