AQUAFace: Age-Invariant Quality Adaptive Face Recognition for Unconstrained Selfie vs ID Verification

Shivang Agarwal, Jyoti Chaudhary, Sadiq Siraj Ebrahim, Mayank Vatsa, Richa Singh, Shyam Prasad Adhikari, Sangeeth Reddy BattuAAAI 2025 · DOI

Introduction

Face recognition systems in real-world applications, such as digital identity verification for mobile banking, face a dual challenge: matching a modern, high-quality "selfie" against an older, often lower-quality "ID photo." This "selfie vs. ID" task is complicated by significant age gaps and variations in image quality, such as scanning artifacts, poor lighting, or blur. Existing margin-based loss functions typically address either age or quality in isolation. AQUAFace, presented at AAAI-25, introduces a unified framework that treats these factors as intertwined components of a sample's "recognizability."

Figure 1: Examples of face pairs categorized by recognizability (Easy, Medium, Hard) across age and quality variations.

Methodology: The AQUAFace Framework

The core innovation of AQUAFace is its adaptive loss function, which dynamically adjusts margins based on the inherent difficulty of the input pair. The framework consists of three primary components:

  1. AQUALR (Age and QUAlity-based Likelihood Ratio): Utilizing a Gaussian Mixture Model (GMM), the system estimates the joint density of comparison scores, image quality, and age labels. This produces a likelihood ratio that serves as a metric for "recognizability."
  2. Adaptive-Margin Contrastive Loss: Unlike standard fixed-margin losses, AQUAFace uses the AQUALR to scale the penalty for hard samples. Pairs with large age gaps and poor quality are assigned higher margins, forcing the model to extract more robust, identity-preserving features.
  3. Identity-Preserving Loss: The model leverages a pre-trained ArcFace backbone to maintain high discriminative power while fine-tuning for specific unconstrained scenarios.
Figure 2: The AQUAFace architecture, featuring a Siamese network, a Gaussian Mixture Model (GMM) for AQUALR estimation, and the adaptive margin contrastive loss.

Datasets and Training

To overcome the scarcity of large-scale, age-varied datasets with quality degradation, the authors introduced Syn_AM, a synthetic dataset containing 56,612 images of 14,017 subjects. GAN-based techniques were employed to simulate realistic aging and quality loss. The model was also evaluated on VerifyMe, a private real-world dataset of selfie-to-ID matches, and several public benchmarks including CALFW, CPLFW, AgeDB, and IJB-C.

Results and Performance

AQUAFace demonstrates superior performance over state-of-the-art models like MagFace and AdaFace. On the CALFW (Cross-Age LFW) benchmark, it achieved a record 97.82% accuracy for ResNet18-based architectures. Notably, the model showed significant gains in the "Hard" category of the VerifyMe dataset, where age and quality variations are most extreme.

The research highlights that combining age and quality into a single adaptive margin is more effective than balancing separate loss functions. Furthermore, the success of training on the synthetic Syn_AM dataset suggests that GAN-generated data can effectively bridge the gap for privacy-sensitive biometric training.

Conclusion

AQUAFace provides a robust solution for unconstrained face verification by explicitly modeling the relationship between age, quality, and recognizability. While the framework excels across most difficulty levels, future work will focus on improving synthetic data generation to better capture the most extreme "hard" cases of human aging and environmental degradation.

Summary updated

BibTeX
@inproceedings{agarwal2025aquaface,
  author = {Shivang Agarwal and Jyoti Chaudhary and Sadiq Siraj Ebrahim and Mayank Vatsa and Richa Singh and Shyam Prasad Adhikari and Sangeeth Reddy Battu},
  title = {{AQUAFace: Age-Invariant Quality Adaptive Face Recognition for Unconstrained Selfie vs ID Verification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i2.32165},
}

Loading related content...

Figure