Javascript must be enabled for the correct page display

Exploring Automatic Emotion Recognition with Neutral Expression Subtraction and a Regression Model

Struijk, S. van der (2015) Exploring Automatic Emotion Recognition with Neutral Expression Subtraction and a Regression Model. Bachelor's Thesis, Artificial Intelligence.

[img]
Preview
Text
AI_BA_2015_StefvanderStruijk.pdf - Published Version

Download (196kB) | Preview
[img] Text
toestemming.pdf - Other
Restricted to Backend only

Download (23kB)

Abstract

The concept of what an emotion is, is explored by going through the current theories available in the field of emotion studies. These theories are Basic Emotion, Appraisal, Psychological Construction and Social Construction models, which can either be measured in a discrete or scalar manner. Then these theories are used to argue that current classification systems lack the scalability for use in more complex practical situations, as a single label only gives limited information. As answer to this problem, Neutral Expression Subtraction (NES) regression is introduced. This approach uses a neutral expression as baseline and makes it possible to train on any emotion separately by assigning belief values. To test whether this new approach has a similar performance to classification systems, three methods are used on the CK+ database to classify 7 emotions and neutral. Method 1 classifies facial expressions in the classical way. Method 2 classifies facial expressions from data where a neutral expression is subtracted from and Method 3 takes the regression approach and assigns belief values for every emotion to an expression. To compare the methods, Method 3 classifies an expression as the emotion with the highest belief value. The data is transformed by using PCA and either KNN or KNN-regression with K-Means clustering is used. Method 1 achieves 37.7%, Method 2, 45.1% and Method 3, 41.2%. These results suggest that NES is useful for expression recognition and that NES regression, using only data of one emotion, could prove useful for the future. The low overall recognition accuracy is mostly due to the limited use of pre-processing and feature extraction, which should be more advanced in future works to make more conclusive statements about NES regression. NES regression should not be judged by these results alone, because its classification performance was only for comparison reasons. The use of belief values has many more possibilities for complex system behavior than a single classified label.

Item Type: Thesis (Bachelor's Thesis)
Degree programme: Artificial Intelligence
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 15 Feb 2018 08:02
Last Modified: 15 Feb 2018 08:02
URI: http://fse.studenttheses.ub.rug.nl/id/eprint/12436

Actions (login required)

View Item View Item