A Transition of Face Recognition to Mask Face Recognition Using Improvised Attention DL Model

With the emergence of the communicable fatal Covid-19 virus and increased level of air pollution, every individual is adapting to wear a facial mask now a days, due to which the existing Face Recognition Models are experiencing reduction in accuracy rate. The study in this paper is aimed to analyze the rate of drop in test accuracy for the task of face recognition from normal face recognition to the recognition of masked faces. Also, it targets to analyze the behavior of Convolution Block Attention mechanism with 10 different deep learning architectures as trunk branch. This paper discusses a novel idea to generate a sample space for training by various combinations of original face images, simulated masked faces and augmentation, from the original 2 datasets which improves the test accuracy of the models. The combination of InceptionV3+CBAM proves to be the best model with the peak accuracy of 88.78% & 88.62% and least model’s inference time of 512 & 923 milliseconds on Yale and Casia subset datset respectively among all other 9 implemented attention models. Moreover the proposed implementation of InceptionV3+CBAM also succeeds to achieve better accuracy as compared to existing models by using proposed set of dataset (Original face Images+ Simulated Mask faces+Augmentation) type for training.