Age Invariant Face Recognition using Lightweight Deep Convolutional Neural Networks

Abstract

Age Invariant Facial Recognition is a field of study that is constantly evolving and is applied in a variety of real-world settings. The majority of age-invariant facial recognition works that make use of Convolutional Neural Networks rely on heavyweight Convolutional Neural Networks. However, training and inferring with heavyweight networks is an extremely resource-consuming task. Lightweight models have a greater advantage over the heavyweight ones, as they have less computational and memory (in terms of the number of parameters) demands, and also are proven to perform well in face recognition tasks. But no previous comparison has been made on which Lightweight Architecture is best suited for the Age Invariant Face Recognition task. Hence, this research aims at achieving three objective 1) compare the performance of 5 pre-trained lightweight Deep Convolutional Neural Network Architectures/models for the task of Age Invariant Face Recognition. 2) conducting set of experiments with different techniques before or during classification to improve the recognition accuracy and, 3) compare the performance of lightweight models pre-trained only on ImageNet for the task of AIFR. 
After conducting all experiments for objectives 1 and 2, the highest accuracies that are achieved by each of the models are 80.75% for MobileNetV1, 87.7% for MobileNetV2, 84.88% for EfficientNet, 80.65% for ShuffleNet, and 78.43% for SqueezeNet. This shows that the MobileNetV2 (pre-trained on a face dataset) is the best-suited architecture/model for the task of Age Invariant Face Recognition. The results of the experiments also show that combining the original embedding’s of the classifiers training set with embedding’s of Generative Adversarial Networks/FaceApp augmented images leads to an increase in classification accuracy given that  (1) A Support vector machine classifier with a “Radial Basis Function” kernel is used and (2) Each embedding is represented individually (the 
highest accuracy is achieved when it is used along with gender guidance). Moreover it is also shows that fine-tuning leads to an increase in accuracy. The experiments conducted to achieve objective 3 show that shows that lightweight models pre-trained only on ImageNet are not suitable for the task of AIFR.