[Paper Review] Face Aging with Contextual Generative Adversarial Nets

Si Liu, Yao Sun+5 more|arXiv (Cornell University)|Feb 1, 2018

Face recognition and analysis34 references3 citations

TL;DR

This paper proposes Contextual Generative Adversarial Networks (C-GANs) for face aging by explicitly modeling cross-age transition patterns between adjacent age groups. The method uses a conditional transformation network with residual blocks, an age discriminative network, and a novel transition pattern discriminative network to generate realistic, identity-preserving aging results that outperform state-of-the-art methods in both visual quality and cross-age face verification performance.

ABSTRACT

Face aging, which renders aging faces for an input face, has attracted extensive attention in the multimedia research. Recently, several conditional Generative Adversarial Nets (GANs) based methods have achieved great success. They can generate images fitting the real face distributions conditioned on each individual age group. However, these methods fail to capture the transition patterns, e.g., the gradual shape and texture changes between adjacent age groups. In this paper, we propose a novel Contextual Generative Adversarial Nets (C-GANs) to specifically take it into consideration. The C-GANs consists of a conditional transformation network and two discriminative networks. The conditional transformation network imitates the aging procedure with several specially designed residual blocks. The age discriminative network guides the synthesized face to fit the real conditional distribution. The transition pattern discriminative network is novel, aiming to distinguish the real transition patterns with the fake ones. It serves as an extra regularization term for the conditional transformation network, ensuring the generated image pairs to fit the corresponding real transition pattern distribution. Experimental results demonstrate the proposed framework produces appealing results by comparing with the state-of-the-art and ground truth. We also observe performance gain for cross-age face verification.

Motivation & Objective

• To address the limitation of existing GAN-based face aging methods that fail to model sequential aging transitions between adjacent age groups.
• To improve identity preservation and appearance consistency during age progression by explicitly modeling higher-order cross-age correlations.
• To enhance the realism of generated aging faces by incorporating transition pattern regularization.
• To boost performance in cross-age face verification by generating more accurate age-progression images.
• To provide a reproducible, end-to-end trainable framework for face aging with public code release.

Proposed method

• The framework employs a conditional transformation network with specially designed residual blocks to learn the mapping from a source face to a target age group.
• An age discriminative network ensures generated faces match the real distribution of each individual age group.
• A novel transition pattern discriminative network is introduced to distinguish real aging transitions (e.g., shape and texture changes) from fake ones, acting as a regularization term.
• The model is trained end-to-end using adversarial loss, with the transition pattern discriminator guiding the generation of plausible aging sequences.
• The architecture is based on DCGAN and supports both face aging and regression without modification.
• Post-processing via graph cuts and Poisson blending is applied to seamlessly integrate generated faces into real head images for qualitative evaluation.

Experimental results

Research questions

• RQ1 Can modeling transition patterns between adjacent age groups improve the realism and identity consistency of generated aging faces?
• RQ2 How does the inclusion of a dedicated transition pattern discriminator affect the quality of age progression compared to standard GANs?
• RQ3 To what extent can the proposed C-GANs improve cross-age face verification performance?
• RQ4 How well does the model generalize across diverse age ranges, especially from child to elderly?
• RQ5 Can the method generate age-progressed faces that are visually indistinguishable from real ones?

Key findings

• The C-GANs method achieves superior visual quality, with generated faces nearly indistinguishable from real images in qualitative comparisons.
• The transition pattern discriminative network significantly improves performance on contiguous age transitions, particularly in preserving facial structure and texture evolution.
• On the FGNET dataset, the EER (equal error rate) for cross-age face verification dropped to 8.72% on C-GANs-generated synthetic pairs, outperforming the 11.05% EER on CAAE-generated pairs and 17.41% on original pairs.
• The method performs well across both male and female subjects, with strong results on movie stars like Brad Pitt and Shirley Temple, accurately simulating age progression across decades.
• Performance degrades slightly when aging from children to 60+, attributed to the large appearance gap, suggesting room for improvement in extreme age transitions.
• The source code is released to support further research, enhancing reproducibility and community adoption.