Deep Features for Age-Invariant Face Recognition

The performance of a face recognition algorithm is affected as facial appearance undergoes a lot of changes with age. Thus, the development of a robust age invariant face recognition (AIFR) system is essential. Training a model from scratch requires a huge dataset and is also computationally expensive. The transfer learning approach has provided solutions to these problems. In this paper, the performance of three pre-trained models FaceNet, VGG-Face, and OpenFace are compared on the FG-NET and MORPH aging datasets. MTCNN and dlib libraries are used for face detection purposes during experimentation. The detected faces are then provided as input to the models which act as feature extractors. The feature vectors generated by the model are passed to Support Vector Machine (SVM) or K-Nearest Neighbors (KNN) classifier, and accuracy is calculated on the test data. This is used as the parameter for comparison among the models.