Constellation Loss em Modelos de Reconhecimento Facial

Abstract

Feature extracting models for facial recognition systems have become objects of in-depth study over the past few years. In order to find a better discrimination of objects and, consequently, a better separability of classes, loss functions, such as Triplet Loss, were designed to be used in conjunction with Siamese Neural Networks. However, networks with these functions suffer from a slow convergence by considering only two classes (positive and negative) at each learning iteration, thus, they are not appropriate when there is a large number of classes in the dataset. Recently, a loss function called Constellation Loss was proposed in order to minimize these problems. In this work, a model for facial recognition using a Convolutional Neural Network (CNN) as a backbone and Constellation Loss as a loss function is proposed. To validate the model, two public databases were used and comparisons were made with different loss functions and CNNs architectures. It is also proposed in this work the use of an approach for the construction of batches, which allows network training with a reduced memory usage. The results obtained indicate that Constellation Loss is a promising technique when compared to the other loss functions evaluated, reaching average values of AUC (Area Under The Curve) equal to 99.9% in the Olivetti Faces dataset and 98.7% in the challenging Labeled Faces in the Wild (LFW) dataset. The effectiveness of the method could be certified, enabling its application to facial recognition systems.