IEEE Access (Jan 2021)
Tackling Age-Invariant Face Recognition With Non-Linear PLDA and Pairwise SVM
Abstract
Face recognition approaches, especially those based on deep learning models, are becoming increasingly attractive for missing person identification, due to their effectiveness and the relative simplicity of obtaining information available for comparison. However, these methods still suffer from large accuracy drops when they have to tackle cross-age recognition, which is the most common condition to face in this specific task. To address these challenges, in this paper we investigate the contribution of different generative and discriminative models that extend the Probabilistic Linear Discriminant Analysis (PLDA) approach. These models aim at disentangling identity from other facial variations (including those due to age effects). As such, they can improve the age invariance characteristics of state-of-the-art deep facial embeddings. In this work, we experiment with a standard PLDA, a non-linear version of PLDA, the Pairwise Support Vector Machine (PSVM), and introduce a nonlinear version of PSVM (NL–PSVM) as a novelty. We thoroughly analyze the proposed models’ performance when addressing cross-age recognition in a large and challenging experimental dataset containing around 2.5 million images of 790,000 individuals. Results on this testbed confirm the challenges in age invariant face recognition, showing significant differences in the effects of aging across embedding models, genders, age ranges, and age gaps. Our experiments show as well the effectiveness of both PLDA and its proposed extensions in reducing the age sensitivity of the facial features, especially when there are significant age differences (more than ten years) between the compared images or when age-related facial changes are more pronounced, such as during the transition from childhood to adolescence or from adolescence to adulthood. Further experiments on three standard cross-age benchmarks (MORPH2, CACD-VS, and FG-NET) confirm the proposed models’ effectiveness.
Keywords