عنوان

Automated Recognition of Facial Affect Using Deep Neural Networks

Number

TL53847

.Language of Text, Soundtrack etc

انگلیسی

Title Proper

Automated Recognition of Facial Affect Using Deep Neural Networks

General Material Designation

[Thesis]

First Statement of Responsibility

Hasani, Behzad

Subsequent Statement of Responsibility

Mahoor, Mohammad H

Name of Publisher, Distributor, etc.

University of Denver

Date of Publication, Distribution, etc.

2020

Text of Note

159 p.

Dissertation or thesis details and type of degree

Ph.D.

Body granting the degree

University of Denver

Text preceding or following the note

2020

Text of Note

Automated Facial Expression Recognition (FER) has been a topic of study in the field of computer vision and machine learning for decades. In spite of efforts made to improve the accuracy of FER systems, existing methods still are not generalizable and accurate enough for use in real-world applications. Many of the traditional methods use hand-crafted (a.k.a. engineered) features for representation of facial images. However, these methods often require rigorous hyper-parameter tuning to achieve favorable results. Recently, Deep Neural Networks (DNNs) have shown to outperform traditional methods in visual object recognition. DNNs require huge data as well as powerful computing units for training generalizable and robust classification models. The problem of automated FER especially with images captured in the wild setting is even more challenging since there are subtle differences between various facial emotions. This dissertation presents the recent efforts I made in 1) creating a large annotated database of facial expressions, 2) developing novel DNN-based methods for automated recognition of facial expressions described by two main models of affect, the categorical model and the dimensional model, and 3) developing a robust face detection and emotion recognition system based on our state-of-the-art DNN and trained on our proposed database of facial expressions. Existing annotated databases of facial expressions in the wild are small and mostly cover discrete emotions (aka the categorical model). There are very limited annotated facial databases for affective computing in the continuous dimensional model (e.g., valence and arousal). To address these needs, we developed the largest database of human affect (called AffectNet). For AffectNet, we collected, annotated, and prepared for public distribution a new database of facial emotions in the wild. AffectNet contains more than 1,000,000 facial images from the Internet by querying three major search engines using 1250 emotion related keywords in six different languages. About half of the retrieved images were manually annotated for the presence of seven discrete facial expressions and the intensity of valence and arousal. AffectNet is by far the largest database of facial expression, valence, and arousal in the wild enabling research in automated facial expression recognition in two different emotion models. This dissertation also presents three major and novel DNN-based methods for automated facial affect estimation. The methods are: 1) 3D Inception-ResNet (3DIR), 2) BReG-Net, and 3) BReG-NeXt architectures. These methods modify the residual unit -proposed in the original ResNets- with different operations. Comprehensive experiments are conducted to evaluate the performance of each of the proposed methods as well as their efficiency using Affect and few other facial expression databases. Our final proposed method -BReG-NeXt- achieves state-of-the-art results in predicting both dimensional and categorical models of affect with significantly fewer training parameters and less number of FLOPs. Additionally, a robust face detection network is developed based on the BReG-NeXt architecture which leverages AffectNet's diverse training data and BReG-NeXt's efficient feature extraction powers.

Subject Term

Artificial intelligence

Subject Term

Biometrics

Subject Term

Computer science

Subject Term

Facial recognition technology

Subject Term

Neural networks

Hasani, Behzad

Mahoor, Mohammad H

University of Denver

Electronic name

p

[Thesis]

276903

a

Y