Javascript must be enabled for the correct page display

Pre-trained Deep Convolutional Neural Networks for Face Recognition

Looije, S (2018) Pre-trained Deep Convolutional Neural Networks for Face Recognition. Master's Thesis / Essay, Artificial Intelligence.

[img]
Preview
Text
MAs2209276SiebertLooije.pdf - Published Version

Download (2MB) | Preview
[img] Text
toestemming.pdf - Other
Restricted to Backend only

Download (78kB)

Abstract

Pre-training of models is important because of the unavailability of datasets and models that are becoming more complex. We investigate two aspects of pre-training using face recognition tasks. The first is the use of models that are pre-trained on face datasets and non-face datasets. We will evaluate five pre-trained models based on their results with freezing multiple layers and on robustness. The second aspect is to investigate universal features in pre-trained deep models. This is done by evaluating the performance using only the first few layers for pre-training. This is also investigated by swapping the first layers of the models. We show that models pre-trained on face datasets achieve better results and are more robust in three face recognition tasks than models pre-trained on non-face datasets. The results with pre-training and swapping only the first layers show a significant difference between models that are pre-trained on face datasets and non-face datasets. From this, we conclude that it is important which dataset is used for pre-training the models and used for testing in face recognition. We also conclude that the first few layers of pre-trained models affect performance on face recognition.

Item Type: Thesis (Master's Thesis / Essay)
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 15 Feb 2018 08:35
Last Modified: 15 Feb 2018 08:35
URI: http://fse.studenttheses.ub.rug.nl/id/eprint/16426

Actions (login required)

View Item View Item