Improved deep face identification with multi‐class pairwise discriminant loss

Abstract

A novel method to extract discriminative deep feature representations of facial images for face identification is presented. A new ‘multi-class pairwise discriminant loss’ is devised and incorporated it into the general deep convolutional neural network learning framework in a novel way, leading to highly discriminative deep face features. The method shows significant improvement over existing deep feature extraction techniques relying on softmax or triplet loss. Moreover, the method achieves a level of accuracy on the widely used identification protocols, which are better and comparable results than other state-of-the-art methods.