A Comparative Study of Face Recognition under Pose Variation

Abstract

Face recognition algorithms enable computational devices to recognize faces. It has a widespread application in commerce, law enforcement and can be effectively used in criminal identification, healthcare, advertising, access and security, payments and other different areas. Face recognition has many advantages over other biometric techniques, such as intrusiveness, user consent, surveillance, and it is completely software-based. However, it also has some limitations, such as large template size, distinctiveness, accuracy, and stability. The purpose of this work is to find and compare the accuracy of three different facial recognition algorithms against the face datasets having considerable pose variation in the x-axis (left and right variation). The methodology for this study covers data collection (the process of collecting facial datasets), pre-processing (the process of cleaning data and making it ready for further operations), processing (to apply learning algorithms either to design a template or extract features), training, and classification. To train and test a model with every single image instance available in the dataset, we have used the K-Fold cross-validation method where the dataset is divided into k folds, and each fold will go for training as well as for testing. Based on the conducted experiments, we have observed that the Local Binary Pattern Histogram algorithm has outperformed the other two algorithms by obtaining 79.806 accuracy against the FEI dataset of faces and 61.477 of accuracy against the CVL dataset of faces.