Masked Face Recognition Using Principal component analysis and Deep learning

After COVID-19 pandemic most people all around the world are wearing face masks in public which makes traditional face recognition methods insufficient to recognize a face with mask. According to NIST2020 [1] report, face recognition accuracy of existing face recognition methods has a 20-50% of error rate due to the face covered by face mask, therefore, video surveillance systems and other face authentication systems are now not working efficiently. Recently many researchers have tried to solve this problem using mainly deep learning methods (CNN, FaceNet). Some researchers have used features from occluded (area covered by face mask) part and non-occluded part of the face while some researchers discard the occluded part and only used the non-occluded part of the face to recognize. For both of these two cases, the facial features of the occluded part of the face cannot be analyzed or cannot be used to recognize the face, as a result, these methods have high error rate. This paper proposed a reconstructive method to get partially reconstructed features of the occluded part of the face, then existing deep learning method has used to recognize the face. First the occluded part of a face is discarded and Principal component analysis (PCA) is applied to the remaining part of the face. Next, the principal components or the most significant Eigenvector and its corresponding weights are calculated to reconstruct the occluded part. Proposed method cannot reconstruct the occluded part completely but it can provide enough features of the occluded part which improves the recognition accuracy up to 15%.