Age-invariant face recognition using gender specific 3D aging modeling

Abstract

The age-invariant face recognition (AIFR) is a relatively new area of research in the face recognition domain which has recently gained substantial attention due to its great potential and importance in real-world applications. However, the AIFR is still in the process of emergence and development, offering a large room for further investigation and accuracy improvement. The key challenges in the AIFR are considerable changes of appearance of facial skin (wrinkles, jaw lines), facial shape, and skin tone in combination with the variations of pose and illumination. These challenges impose limitations on the current AIFR systems and complicate the recognition task for identity verification especially for temporal variation. In order to address this problem, we need a temporally invariant face verification system that would be robust vis-a-vis several factors, such as aging (shape, texture), pose, and illumination. In this study, we present a 3D gender-specific aging model that is robust to aging and pose variations and provides a better recognition performance than the conventional state-of-the-art AIFR systems. The gender-specific age modeling is performed in a 3D domain from 2D facial images of various datasets, such as PCSO, BROWNS, Celebrities, Private, and FG-NET. The evaluation of the proposed approach is performed on FG-NET (the most referred database in the AIFR studies) and MORPH-Album2 (the largest aging database) by using the VGG face CNN descriptor for matching. In addition, we also test the effects of linear discriminant analysis (LDA) and principal component analysis (PCA) subspaces learning in our face verification experiments. The proposed AIFR system is evaluated both on the pose corrected and background composited age-simulated images. The experimental results demonstrate that the proposed system provides state-of-the-art performance on FG-NET (83.89% of rank-1, 43.24% of TAR) and comparable performance to the state-of-the-art on MORPH-Album2 (75.27% of rank-1, 96.93% of TAR).