Comparative analysis of the Light-CNN and FaceNet methods for identifying and maintaining human faces

Abstract

Maintaining the identity while synthesizing the frontal view image is the most critical step in developing a "recognition via generation" framework. To this end, this paper investigates, tests and compares the performance of two deep learning architectures: Light-CNN and FaceNet. The Light-CNN is used to learn a robust feature for face verification tasks that produces a high-level facial identity accuracy over many traditional deep learning models. FaceNet, on the other hand, is a model to maps face images into a compact Euclidean space where distances directly represent a measure of face similarity. In our comparison, we use the TP-GAN model to perform several pre-processing stages. The face features are then extracted from the synthesized face images using Light-CNN and FaceNet as 256- and 128-dimensional representations, respectively. We evaluate the accuracy performances of Light-CNN and FaceNet architectures on Multi-PIE and FEI datasets.