Simulated and Cropped Masked FER2013:A Novel Simulated Dataset for Masked Facial Expression Recognition

Abstract

Wearing masks has become the norm after the COVID-19 pandemic, obscuring key facial regions and posing significant challenges to Facial Expression Recognition (FER) systems. Existing FER datasets, such as FER2013, lack masked face images and consist of low-resolution samples, resulting in reduced model robustness under complex occlusions. To address these limitations, this paper introduces an innovative approach that integrates image processing and data augmentation techniques to simulate realistic mask occlusions in the FER2013 dataset. Leveraging the MaskTheFace algorithm, the proposed method automatically generates the Simulated and Cropped Masked FER2013(SCMFER2013) dataset. OpenCV is utilized for facial area detection, Dlib for facial key point localization, and the Albumentations library for data augmentation. Experimental results demonstrate that models trained on the SCMFER2013 dataset achieve substantial performance improvements, particularly under challenging low-resolution and occlusion conditions.