MASKED FACE RECOGNITION USING DEEP LEARNING

Abstract

The COVID-19 crisis has brought significant changes in our lives. Social distancing has become a norm. Masks are part of our daily life. We cannot leave out homes without wearing masks. These masks have become like an integral organ for survival. But these masks have caused problems with our computer models for face recognition. Face recognition is the sub-field of computer science in which the computer matches an input face to a corresponding set of output images to deduce the identity of the face provided as input to the system. But with the masks, the face is covered from nose till neck. Only the eyes and forehead region is visible with a mask on the face. This creates a problem with existing techniques available in the domain of face recognition using computers. The computer models achieving accuracy over 95% for a face drops to a very low level when the same face is given as input with a mask on it. The predictions made by these models are no better than random predictions made by untrained models. In this project, the performance of different state-of-the-art models has been studied. A modified version of the existing dataset is utilized for training and testing these models. The modification is done by augmenting the masks on faces in the chosen part of the VGGface dataset. For faster training and testing, the concept of transfer learning plays a big role. The pre-trained models are being adapted to the modified dataset. Apart from this, a new model is also introduced which is somewhat a hybrid of the best performing models. This new model architecture is defined and then trained and tested on the modified dataset. The new model is thrown against the state-of-the-art models. The aim of developing this new model is to improve over the existing baselines present. Based on the closed observations, the research questions are answered.