Face recognition under occlusion for user authentication and invigilation in remotely distributed online assessments

This study focuses on face recognition under uncontrolled conditions as a second biometric factor in order to multi factor authenticate(MFA) in online assessment. Obtained results of this project indicate reasonable accuracy to address the issue of occlusion using AR, MUCT and UMB Datasets, utilizing deep learning and the previous approach based on feature extraction (shallow method). The shallow method accuracy improvement includes HOG by 4%, in comparison to Gabor Sparse Representation based Classification (GSRC) method and by 9% using Gabor. Shallow method can handle occlusion issue in the lack of occlusion dictionaries and sufficient training sample. Modified ResNet as a deep learning method is used to be able to improve accuracy comparing the best member of the SRC family, Structured Sparse Representation based Classification(SSRC) by 3% on average.