Variational and Deep Learning Methods in Computer Vision
Computer Vision is a field that aims to simulate the human visual system. In the last decade, with the continuous emergence of multimedia data and applications, there has been an increasing interest to exploit all this available information, which mainly consists in images and videos. Classic approaches to Computer Vision problems constitute a Bag of Tricks that have been useful for many years. With the irruption of Deep Learning, most of these techniques, all of a sudden, became old. The reasons are the impressive outperforming results of Deep Learning techniques that, taking advantage of the available data, provide an end-to-end solution that is nowadays easy to use, even for non experts users. Surprisingly, some Variational Methods, which can be considered as classical methods in Computer Vision, survived and maintained the state of the art leading in some specific tasks: Medical Imaging Registration for instance. The impact of Deep Learning applications in society is undeniable. Moreover, the profit of automatizing many processes and tedious tasks that are still nowadays realized by humans, should be taken as good news, since it would provide more free time for people... and thus, more time to live. The dark side of such automatization will rely on how this new techniques, framed in the Artificial Intelligence field, are democratized across society. This is, how useful in practice are those new emerging tools and who has access to them. Autonomous driving, Medical Imaging, earthquakes and pollution predictions are few examples of critical application fields where being inaccurate implies disastrous consequences. In such scenarios, classic approaches in Computer Vision, provides less uncertainty in outcomes. In this sense, classical methods are more robust, in particular Variational Methods which have a deep and strong mathematical foundations. Moreover, recently, adversarial attacks on Neural Networks have shown how easy is to fool Deep Learning systems, increasing skepticism for potential Deep Learning users as Medical experts. In this thesis we address Computer Vision problems in real scenarios from two perspectives with the usage of: (1) Variational Methods and (2) Deep Learning techniques. The former is a powerful tool that gives an extraordinary control over the expected outcomes with very accurate results if some hyperparameterization is carried out properly. However, this required (usually manual) hyper-parameterization constitutes a huge shortcoming in practice, and a limitation for a wide use by non experts. The later relies mainly on data and solves, until a certain point, an high-dimensional interpolation problem with astonishing results that, however, are sometimes unpredictable (and thus dangerous) when unseen data from different distribution is tested (extrapolation).
Tesis Doctoral leída en la Universidad Rey Juan Carlos de Madrid en 2019. Directores de la Tesis: Juan José Pantrigo Fernández y Emanuele Schiavi
- IA - Tesis Doctorales