Recognition Of Masked Faces Via AlexNet,ResNet-50,and Inception V3 Models

Since the outbreak of Coronavirus pandemic in December 2019,there is an increasing concern for the development of better-masked face detectors.This stems from the fact that everyone has to be protected against the spread of the virus.However,measures taken to prevent the spread of the virus poses challenges to systems and organizations as the existing systems cannot match faces with mask more efficiently.Developing masked face recognition systems would help organizations to identify and recognize faces in masks more efficiently and improve the security system across the globe.For the purpose of this study,a customized dataset that can ease the recognition and identification of masked and unmasked faces of individuals was generated.This was necessitated by the unavailability of large face datasets for masked and unmasked faces,and the alignment of existing datasets to Caucasian/white faces and lacks Aethiopian/black faces.Theoretically,recognizing a masked face is a challenge because,70%(the lower face and the middle face region)of the facial features of human is covered with mask,leaving 30%(Upper face)features uncovered.This study explored AlexNet,Res Net-50,and Inception V3 models by using 30% of the facial features(Upper Face)of the masked faces in recognizing Aethiopian/black race and Caucasian/white race faces with Surgical masks,N95 masks and Fabric masks.In designing a Masked Face Recognition System,the images obtained were preprocessed by normalization and randomization which minimized the variation for each input image and changed each image to the same measurements to aid feature extraction process.Principal component analysis(PCA)was utilized to obtain the main features of the face image to perform the weighted processing,to build the database for the mask face recognition systemThe findings of the research indicate that the CNN models produce excellent recognition accuracy on masked faces and unmasked faces.However,Resnet-50 achieves a higher recognition accuracy compared with AlexNet and Inception V3.Analysis of the models' performance on Aethiopian and Caucasians showed AlexNet model achieved a 35.2%recognition rate on Aethiopians and a 56% recognition rate on Caucasians.Res Net-50 achieved a 60% recognition rate on Aethiopian and 40% on Caucasians.Inception V3 also achieved 62.5%on Aethiopian and 37.5 on Caucasian.It was deduced from the study that Inception V3 performs better in recognizing Aethiopian faces than Resnet 50 and AlexNet model,and AlexNet also performs better in recognizing Caucasians than Inception V3 and Res Net-50.The datasets generated in this study produced excellent results with the CNN models,and can serve the purpose of masked face recognition for security and surveillance systems.