Clasificación de objetos usando aprendizaje profundo implementado en un sistema embebido

Abstract

Image classification is a challenging problem that receives significant attention. The incursion of frameworks for deep learning allows the construction of classifiers of greater precision and ease. The aim of this project is to classify images using three of the most popular frameworks, TensorFlow, PyTorch and Caffe. For this purpose, the operation of the Convolutional Neural Networks (CNN) has been analyzed. A CNN from the Alexnet and ResNet architecture has been used for image recognition of the Cifar-10 dataset. All this runs on a Nvidia Jetson TX2 embedded system, comparing the results with those obtained in running on a desktop computer and laptop. Where we found that the installation and configuration of the frameworks in the board Jetson TX2 depends to a large extent on the version of them and the library CuDNN. The Caffe library was found to have less runtime at the training stage. TensorFlow is the most documented and easy-to-understand library