Similarity Metric for Millions of Unlabeled Face Images

Lately, image classification with unknown labels has been the focus of many research related to social media and cyber-security. In order to cluster images into labels, a similarity metric is used to compare images and based on the output a decision is made. In this article, we propose a Convolution Neural Network (CNN) based metric that achieves state of the art results. The proposed approach learns the priority features that differentiate two given input faces. The metric is trained on MSCeleb (10M faces of 100k individuals) and VGGFace2 (3.3M faces of 9k individuals) then tested using LFW (13 K faces of 5,749 individuals), MSCeleb, and VGGFace2 datasets. The network is evaluated through
$F_{\beta}$
measure and accuracy. The metric proposed in this paper achieves state of the art
$F_{\beta}$
measure of 0.94 for LFW benchmark.