Style Transfer to Enhance Data Augmentation for Facial Action Unit Detection

Abstract

In recent decades, advances in fields such as robotics, computer graphics, and computer vision have led to an increase in interest in the study of image recognition among scientists. One of the difficulties in image recognition is obtaining a sufficiently large and well-balanced dataset to improve the effectiveness, robustness, and reliability of machine learning models. This paper explores a research approach that addresses this issue by style transfer and observes how different styles affect the recognition rates on facial recognition tasks. This study claims that synthetic data augmentation can improve the model’s performance. The experiments were conducted on three different datasets by transferring two different image styles over the datasets, and it has been observed that the synthetic data generation on facial datasets proves to be a tool that significantly improves the accuracy of machine learning models in image recognition on micro-mimic movements (facial action units) of the face when the original data is not distorted much and has a lower impact when a more complex style is applied.