Date of Publication

2023

Document Type

Dissertation/Thesis

Degree Name

Master of Science in Computer Science

College

College of Computer Studies

Department/Unit

Software Technology

Thesis Advisor

Merlin Teodosia Suarez

Defense Panel Chair

Joel Ilao

Defense Panel Member

Macario Cordel II
Jocelynn Cu

Abstract (English)

Micro-expression is an involuntary facial expression that can reveal the genuine emotion of a person. Because of their low intensity and short duration, micro-expressions are challenging to collect, resulting in few publicly available databases. There are studies that generate micro-expressions, mainly using generative adversarial networks. A few generate the full micro-expression cycle and are typically evaluated by creating recognition models. Given the variety of expressions appearing in an image in low intensity, generating micro-expressions on the image is difficult.

This works investigates the effect of using the neutral expression in generating synthetic micro-expressions. As micro-expressions begin with an expression similar to a neutral expression, starting with a neutral expression may reduce the visual artifacts, thus, creating better images. N2ME generates a full-sequence synthetic micro-expression. Action unit analysis shows a high Pearson's correlation on the generated data, specifically on AU4, AU6, AU7, AU10, AU12, and AU14. Optical flow analysis shows that subtle movements between the onset and apex frames shown in the generated data were comparable to the ones shown in the real data. Combining the generated and real data shows a positive impact on the recognition of emotion classes that previously suffered in performance due to a low number of samples. In addition, the incorporation of pseudo-AU27 shows a significant improvement in generating synthetic neutral expressions.

Abstract Format

html

Language

English

Upload Full Text

wf_yes

Embargo Period

8-9-2023

Download

Share

COinS