Discriminative Frontal Face Synthesis by Using Attention and Metric Learning

Abstract

This paper introduces a novel approach for obtaining distinctive frontal facial representations from collections of multiple facial images. The primary objective is to ensure that the profound features extracted through a deep Convolutional Neural Network (CNN) from these learned facial representations exhibit notable separability within the feature space. The acquisition of frontal facial representations capable of effectively representing entire sets of images holds significant value as it considerably reduces the number of image samples requiring processing. This acceleration proves especially advantageous during the classification testing phase. The proposed method combines three fundamental components: attention mechanisms, adversarial methodologies, and metric learning strategies. We adopt a U-Net architecture enhanced by attention modules for the facial aggregation network that generates frontal faces that approximate multiple face images within image sets. Furthermore, we employ both a discriminator network and a pre-trained facial classification network to successfully achieve the goals of adversarial and metric learning. The experimental studies on different face recognition datasets demonstrate that using only attention mechanisms and metric learning strategy is sufficient to synthesize discriminative frontal face images yielding high classification accuracies.