Masked face recognition using convolutional neural network

• Main Author: Wu, Peihang
• Other Authors: Yap Kim Hui
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/164453
• Institution: Nanyang Technological University

In recent years, due to the global epidemic of COVID-19, in view of the global health crisis caused by the new crown pneumonia, masks have become essential protective equipment and accessories for people to go out. This sanitation poses a dilemma for state-of-the-art facial recognition models, because they were not designed to work with masked faces. Because wearing a mask will make part of the face information be blocked, and the style and color of the mask will also increase the difficulty of the original model to recognize faces. So it is necessary to make some improvements to the original model to adapt to different situations. This dissertation aims to evaluate two different masked face recognition methods, which are convolutional block attention module and training model with simulated masks. The convolutional block attention module is mainly based on the attention mechanism, which makes the module to focus on the upper half of the face that is not obscured by the mask. The training model with simulated masks is to use data augmentation to generate a masked version dataset using the original dataset, and merge two datasets for training the network. Both methods perform well on the current dataset of simulated masks and real masks. The detailed and comprehensive experiments and in-depth analysis of the results provide valuable insights for future research on face recognition with face masks and even any occluders such as hats, eye masks, etc.