Image geopositional location in mobile phones using neural networks

Identifier:  TFM:360

Handle: https://hdl.handle.net/20.500.11797/TFM360

Authors:  Pell Vidal, Xavier

Abstract:
Deep learning in image processing and classification tasks has evolved exponentially in the last few years thanks to new advances and outperforms in existing neural networks. These networks can be trained with labelled images in order to be subsequently used to identify and classify sets of objects in it. Moreover, some libraries have been recently developed to be executed in mobile phones, so it is possible to use the phone camera to capture an image in real time, classify it with a trained neural network in order to obtain specific information on what you are seeing through the camera. In this thesis, we take a quick view to some convolutional neural networks that in conjunction with a deep-learning mobile framework let us to develop a prototype that would be able to geolocate users in a space using phone camera. This process consists in applying transfer learning to the pre-trained model MobileNet, to modify the output of the network in order to classify five objects landmarks that are in a geolocation area. This re-trained network is converted to a more light-weight version to allow its execution in a developed mobile application written in Android. Finally, we evaluate the final data to assure the proper function of the system, the accuracy of the classification process and the geolocation information that is extracted of the final project prototype.