Bias Elimination Network and Style Transfer for Facial Expression Recognition
Part Of
11th IEEE International Conference on Consumer Electronics - Taiwan, ICCE-Taiwan 2024
Start Page
541
End Page
542
ISBN (of the container)
979-835038684-4
Date Issued
2024-07-09
Author(s)
Yo-Shin Liu
Abstract
Humans often evaluate the emotion through reading their facial expressions and then generate corresponding responses. Thus, in the world where the AI Chatbots are popularly investigated, the ability for robots to recognize people’s facial expressions becomes a critical issue to solve. In this work, we first interpret the necessity to eliminate image bias within those datasets, which may be the stumbling block to those previous works. Then we introduce an innovative way to overcome the challenges produced by biased images and then solve the task of Facial Expression Recognition, where the Style Transfer technique is used to enhance our data for both training and inference phases. In our experiments, we show that after applying our method to the JAFFE dataset, the recognition accuracy significantly outperforms the same model trained on unenhanced ones. Our method to eliminate data bias should be generalizable to all face-related tasks and even applicable to other field of machine learning, and we hope the performance of these tasks can take a big step forward.
Event(s)
11th IEEE International Conference on Consumer Electronics - Taiwan, ICCE-Taiwan 2024
SDGs
Publisher
IEEE
Type
conference paper