Deeply Learned Invariant Features for Component-based Facial Recognition

Abstract

Face recognition underage variation is a challenging problem. It is a difficult task because ageing is an intrinsic variation, not like pose and illumination, which can be controlled. We propose an approach to extract invariant features to improve facial recognition using facial components. Can facial recognition over age progression be improved by resizing independently each individual facial component? The individual facial components: eyes, mouth, and nose were extracted using the Viola-Jones algorithm. Then we utilize the eyes region rectangle with upper coordinates to detect the forehead and lower coordinates with the nose rectangle to detect the cheeks. The proposed work uses Convolutional Neural Network with an ideal input image size for each facial component according to many experiments. We sum up component scores by applying weighted fusion for a final decision. The experiments prove that the nose component provides the highest score contribution among other ones, and the cheeks are the lowest. The experiments were conducted on two different facial databases-MORPH, and FG-NET databases. The proposed work achieves a state-of-the-art accuracy that reaches 100% on the FG-NET dataset and the results obtained on the MORPH dataset outperform the accuracy results of the related works in the literature.