Data Augmentation for Deep Learning

Matthieu Cotting

Semester Master Project

, EPFL

January 2017

Abstract

This project aims at investigating ways of performing data augmentation in order to train a deep neural network and improve its performances. To do so, a plug-in for imageJ was implemented allowing us to perform several linear transformations as well as elastic transformations (c.f. figure). The transformations were applied on the MNIST database and the LeNet network has been used. Finally, some effects of these transformations on the performances of the network have been assessed using the following metrics : accuracy, specificity, precision and sensitivity.