Identifying facial landmarks, action units and emotions using deep networks

The goal of this thesis it to use deep neural networks, specifically Convolutional Neural Networks (CNNs) to predict facial landmarks, facial action units and emotions and to study the results of intermediate experiments while doing so. Learning the different features of facial images has always been a difficult task and primarily involves using hand-crafted features which would almost definitely ignore some information related to the different dynamics of facial features. We train our network model using the raw facial images and study its effectiveness in predicting facial landmarks, action units and emotions. In this thesis we learnt that CNNs are highly effective in predicting facial landmarks and AUs, mainly because of their ability to learn features from raw images. We also established that feature sets which can effectively outline the different properties of a face are more useful in classifying facial emotions than either images or facial landmarks.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2016

Author dc:creator

• Prabhu, Namrata

Contributors dc:contributor

• Hoiem, Derek

• Convolutional Neural Networks (CNN)
• Facial Landmarks
• Expressions
• Action Units

Statement dc:rights

• Copyright 2016 Namrata Prabhu

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/92967

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/92967