The Impact of Data Augmentation on Prototypical Networks for Open-Set Face Identification

Extensive research has been conducted on face recognition for large training datasets, and it continues to be an active study area. The present study aimed to examine the issue of open-set face recognition using the ITB Frontal Profile Limited Dataset (IFPLD). This dataset comprises a single frontal and profile face image for each subject. The system is trained to determine, given a profile face of an individual, which frontal face pertains to the same individual. As a result of the constrained dataset accessible for training and testing purposes, the face identification problem is considered a few-shot-open-set problem. Prototypical networks are applicable for face recognition on the IFPLD because they are one of the meta-learning techniques used to address the few-shot learning problem. This study investigates the effect of adding data augmentation to training and testing Prototypical Network on the IFPLD dataset by proposing two distinct techniques for identifying a query profile face. Our investigation using the standard model for few-shot learning (FSL) shows that method 2 achieved higher accuracy, 86.2% for 5-ways-1-shot and 66.3% accuracy for 20-ways-1-shot scenarios.