Identificação de objetos utilizando visão Neuromórfica e redes neurais convolutivas

Abstract

The recognition and identification of objects through computer vision, using deep learning algorithms, are able to provide accurate results, but it is a challenging process, and a high computational cost. This costly process is a consequence of feel characteristics of traditional images, such as high dimensionality, the spatial resolution of images that is as large as their amount of pixels and the robust architecture of hardware that require high performance. However, the biological visual system accomplishes this task with extreme mastery, sophistication, high performance and low energy cost; its high capacity is the fruit of years of evolution and improvement. Neuromorphic engineering uses a totally bioinspired approach to try to achieve the efficiency of biological nature and hence embrace the frontiers of knowledge in neuroscience. Combining a neuromorphic vision sensor with bioinspired algorithms, we developed a computer vision system for classification and identification of objects using convolutional neural networks. The developed learning model presented an accuracy of 75.31% using a dataset of images recorded by a neuromorphic camera (DVS-128).