DeepOrder
 |
Deep Ordering for Vector-Valued Mathematical Morphology and Neural Networks and University of Campinas |
Team:
Santiago VELASCO-FORERO
Marcos Eduardo VALLE
Jesus ANGULO
Samy BLUSSEAU
Peter SUSSNER
João Batista FLORINDO
Abstract
Deep learning techniques achieved outstanding performance in various image processing and analysis tasks. Deep networks for vector-valued images represent an active research topic and include, for example, hypercomplex-valued neural networks. The nonlinearity played by some layers and activation functions in modern deep neural networks is closely related to mathematical morphology, a theory of image operators based on topological and geometrical. This research project aims to develop a mathematical framework encompassing mathematical morphology, hypercomplex algebras, and deep learning. As a result, we expect to devise powerful and robust machine-learning techniques for vector-valued image processing, taking into account topologic and geometric concepts.
Project's topics
Rank-operators for gradient descent approaches
Hypercomplex-valued and vector-valued networks