Efficient Neuromorphic Vision Pipelines for Egocentric Perception on Low Power Systems

Convolutional Neural Networks have seen great success when applied to a wide range of computer vision problems. However, they are still significantly outperformed by the human visual system in both efficiency and ability. This naturally leads to a desire to incorporate mechanisms found in the human visual system into CNNs to improve their performance and efficiency. Current research at the University of Glasgow is exploring the utility of a biologically inspired software retina and an extension of the log-polar image transform to reduce the memory requirement and training time of CNNs as well as improving their scale and rotation invariance. This work intends to investigate the development of a novel CNN architecture that is specifically designed to process the output of the software retina in order to maximize its computational efficiency as well as its performance in classic computer vision problems. Simultaneously this research will offer a more complete computational model of a biologically plausible vision system. Accordingly, the objective of this research is to design a novel Convolutional Neural Network architecture that optimally processes the multiresolution output of a Software Retina and also improve the interface between the two, as listed below