Sistema ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

 

Tesi etd-07052017-140048


Tipo di tesi
Tesi di laurea magistrale
Autore
ROMOLINI, VITTORIO
URN
etd-07052017-140048
Titolo
Implementation of a Face Recognition System using Convolutional Neural Networks
Struttura
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
COMPUTER ENGINEERING
Commissione
relatore Amato, Giuseppe
relatore Falchi, Fabrizio
relatore Gennaro, Claudio
Parole chiave
  • convolutional neural network
  • deep learning
  • riconoscimento facciale
  • face recognition
Data inizio appello
24/07/2017;
Consultabilità
completa
Riassunto analitico
The availability of large training datasets and the introduction of GP-GPUs, along with a number of algorithmic news, have fostered the recent progress in the computer vision field of Artificial Intelligence.

The latest state-of-the-art approaches for face recognition have taken advantage of such a progress by exploiting deep convolutional neural networks (DCNNs). I have examined these methods, as well as the most widely used face datasets and the techniques for face detection and alignment.

Moreover, I have selected a promising method to be developed, based on the training of DCNNs through a ''triplet loss function''.
Triplets are composed by three images: an anchor sample, a positive sample similar to the anchor, and a negative sample that instead belongs to a different class.
The investigated loss function enables the network to learn a mapping that projects the similar input images to embeddings whose Euclidean distance is smaller enough than the distance between the embeddings associated to anchor and negative samples.
In this manner, after training on triplets of face images, the similarity verification between two face samples can be reduced to the comparison of the Euclidean distance of their embeddings to a distance threshold.

I have implemented the software tools needed to realize such a technique and I have trained and tested a number of models with publicly accessible face datasets, assessing different training settings and exploring the hyper-parameters space.

Finally, I have built a face recognition system that enhances the face verification accuracy of the base method, even on strongly unaligned face images, and also surpasses the human accuracy.
File