A Comparative Study of Face Recognition Algorithms under Occlusion

Abstract

Face recognition algorithms are used to automatically recognize human faces. It has got a wide variety of applications in many areas such as Surveillance, access, security, advertisement, healthcare, and etc. Many big tech companies have already adopted this technology and it has been proved as promising convenient biometric technology. In this paper, we are comparing the face recognition algorithms performance against the datasets that are reflecting considerable occlusion (the hidden part of the face, the face parts could be hidden with scarf, glasses, hair or any other object). We selected two publicly available datasets, the first one is the face disguise dataset that reflects major occlusion and the second one that is Specs on Faces (SoF) dataset that reflects partial occlusion. After the data collection, we run the data preprocessing techniques in which we removed the existing noise to the datasets and organized them into different sets. Afterward, we applied feature extraction algorithms and then we fed them into classifiers to get algorithm's performance. At the end of the experiments, we observed that the Local Binary Pattern Histogram (LBPH) algorithm outperforms the other two algorithms by securing 33.444% accuracy against the dataset with major occlusion and 98.504% accuracy against the dataset with partial occlusion, and Linear Discriminant Analysis (LDA) secured the second position against the dataset with major occlusion but third position against the dataset with partial occlusion.