An investigation on the use of convolutional neural network for image classification in embedded systems

The study of Convolutional Neural Network (CNN) for image classification is basically carried out on high performance and parallel platforms, so that the results of the literature cannot be replied on embedded systems. The aim of our work is to investigate CNN architectures that can run in such limited platforms and still maintain or improve the results of the current approaches. To that end, we specify and evaluate the performance of several CNN frameworks using different network configurations and dataset pre-processing techniques. The results of our final approach show that its classification efficiency is close to the best results of the literature, however using a much lower computational power.