Masked face detection with anti-spoofing

Abstract

Modern facial recognition models have excellent performance identifying cleaned, unobstructed faces. However, limitations arise when these models are faced with novel occlusion conditions. This is a concern as occluded faces are common, especially during the Coronavirus Pandemic where facial masks are required in most settings. Masked faces hinder the performance of facial recognition models in carrying out important tasks. In this project, we will dive into details on modern neural network architecture that deals with occlusion conditions and understand their limitations. The focus is primarily on two recent research, FROM and TDMPNet architecture, that have made significant advancement in detecting occluded faces. The project will leverage on the key techniques learned to better detect occlusion patterns on masked faces. Our results show that the Attention Map produced has good performance in detecting occlusion patterns but further fine tuning is necessary.