Deep-feature encoding-based discriminative model for age-invariant face recognition

Abstract

Facial aging variation is a major problem for face recognition systems due to large intra-personal variations caused by age progression. A major challenge is to develop an efficient, discriminative feature representation and matching framework, which is robust to facial aging variations. In this paper, we propose a robust deep-feature encoding-based discriminative model for age-invariant face recognition. Our method learns high-level deep features using a pre-trained deep-CNN model. These features are then encoded by learning a codebook, which converts each of the features into a discriminant S-dimensional codeword for image representation. By incorporating the locality information in the whole learning process, a closed-form solution is obtained for both the codebook-updating and encoding stages. As the features of the same person at different ages should have certain correlations, canonical correlation analysis is utilized to fuse the pair of training features, for two different ages, to make the codebook discriminative in terms of age progression. In the testing stage, the gallery and query image's features are encoded using the learned codebook. Then, linear mapping based on linear regression is employed for face matching. We evaluate our method on three publicly available challenging facial aging datasets, FGNET, MORPH Album 2, and Large Age-Gap (LAG). Experimental results show that our proposed method outperforms various state-of-the-art age-invariant face recognition methods, in terms of the rank-1 recognition accuracy.