Quaternion neural networks towards Real-time image processing

Training neural networks has become an almost daily activity for researchers working on different fields. Preparing accurately the training patterns often appears to be fundamental in obtaining a good balance between learning performance and time needed to reach them. In this contribution, we explore a paradigm to organize patterns for training quaternion neural networks for image processing tasks. The basic idea is to exploit the working principle of Cellular Nonlinear Networks, where local interactions are fundamental, to determine an efficient learning of multidimensional neural network, thus merging the main characteristics of the two architectures. A robustness analysis and practical applications are also presented.