Detección de logos en imágenes mediante redes convolucionales

Abstract

[en] This memory will talk about an application for Android and ios devices that will look for events or nearby places that will offer discounts and/or rewards interacting either click or taking photos of a certain logo. The project is separated in two parts. The first part is about how the app and the server have been developed, explaining its features and the technologies that have been used. This document will make an introduction on this, but will be focused on the second part, image recognition through Convolutional Neural Networks. We tested two different networks, a custom network advised by our tutor to this specific situation, and another net based on a pyramidal structure called AlexNet. We have prepared a dataset to build the networks. Our dataset contains 11 classes with an average of 500 pictures per class. The networks have been configured with different parameters and we’ll compare the results obtained to choose one. We’ll also explain the different techniques tested to improve predictions. Finally, we’ll talk about limitations, improvements and extensions that the app could have, such as new events or to improve the current dataset.