Real-Time Emotion Recognition System for Dynamic Real-World Environments

Facial expressions are a direct and innate form of meaningful channel that conveys emotions non-verbally. Though considerable breakthroughs have been made in Automatic Factial Emotion Recognition (AFER), the research has been on lab-controlled datasets. Real-world scenarios pose expression-unrelated...

Full description

Saved in:
Bibliographic Details
Published in2023 16th International Conference on Developments in eSystems Engineering (DeSE) pp. 731 - 736
Main Author Madnani, Mayur
Format Conference Proceeding
LanguageEnglish
Published IEEE 18.12.2023
Subjects
Deformation
Emotion recognition
Explainability
Face recognition
Facial Expression Recognition
FER
FER 2013
FER in the wild
FER Plus
Knowledge engineering
Lighting
Predictive models
Transformers
Online AccessGet full text
DOI10.1109/DeSE60595.2023.10469007

Cover

More Information
Summary:Facial expressions are a direct and innate form of meaningful channel that conveys emotions non-verbally. Though considerable breakthroughs have been made in Automatic Factial Emotion Recognition (AFER), the research has been on lab-controlled datasets. Real-world scenarios pose expression-unrelated variations, such as illumination, occlusion, pose variations, skin-tone variation, subject identity bias, face deformation, motion blur, etc. There is a lack of data with such unconstrained conditions in contrast to the lab-controlled setting. With deep learning techniques, larger models bring significant accuracy but fail to work in real-time and are left unusable in real-world scenarios and even suffer from the explainability of the outputs. The research develops a lightweight Facial Emotion Recognition system for emotion recognition of facial images, which can operate in real-time and is suitable for use on devices with limited hardware resources. The developed FER system is comparable with previous research and SOTA models and also explainable with far fewer model parameters. With less than 600K model parameters, the proposed FER system achieves a weighted test accuracy of 60% on FER-2013 and 65% on FERPlus respectively, with an inference time of less than 560ms.
DOI:10.1109/DeSE60595.2023.10469007