Sparse random encoder for age invariant face recognition

Identifying people with age variation is one of the challenging tasks in face recognition since aging changes, e.g. facial shape and increasing wrinkle. In this paper, we propose to learn age-invariant features by transforming face images taken in a different age to uniformly random faces through sparse auto-encoder neural networks. The aim is to build a reference set of sparse random encoders where each one is trained on images in a different age for the same person with random faces as the target values to learn age-free high-level features representation. By using random faces, the learned representation highlights precise features that are invariant across time at the neurons in the hidden layer. The proposed method evaluated on large-scale dataset known as Cross-Age Celebrity Dataset (CACD), it includes more than 160,000 face images of 2,000 different celebrities. The experiments show that the proposed method outperforms the state-of-the-art approaches on handling age variations.