Masked face identification using face recognition

Abstract

This work intends to express one of the several well-known biometric authentications entitled Masked Face Identification models by applying current Face Recognition algorithms and public masked face raw data that predict beneficial use. At the end of 2019, the COVID-19 pandemic has been exotically expanding worldwide, which severely negatively harms people’s economies and well-being. Since using facial masks in social environments is now an efficient system to stop the spread of viruses, Nevertheless, appearance identification using facial masks is now a profoundly demanding duty because of the shortage of appropriate facial statistics. Here in our approach, the Deep Learning method will be executed by us to recognize the masked appearance by employing different face portions, some extra-superintendent and some owned-superintendent multi-task training facial appearance spotters, which can compact with different scales of appearance quickly and effectively. Additionally, the features are extracted by us from the masked face’s eyes, forehead, and eyebrow areas and merged with characteristics acquired from those methodologies into a combined structure for identifying masked faces. In order to process, we will perform various image processing techniques on our dataset to clean our data for better accuracy. We will train our model from scratch to perform face-mask recognition. The most important part of this project remains the data collection and data cleaning process. Using a data-centric approach, we will systematically enhance our data-set to improve accuracy and prevent overfitting by performing data augmentation and stratified sampling and keeping our model architecture constant. Finally, our proposed systems will be compared by us with multiple unions of genius appearance identification techniques among those advertised by CASIA, LFW, and owned gathered raw data, which are managed from different sources. When wearing a mask, a person’s face is hidden by 60–75%. Using only 30–40% of a person’s face, we designed a face mask recognition model with an accuracy of 99.84%. Trained on a modified CASIA dataset containing images with and without masks, the model could successfully get the embeddings of 85743 people within a few minutes and perform perfect face recognition with and without masks.