When Age-Invariant Face Recognition Meets Face Age Synthesis: A Multi-Task Learning Framework and a New Benchmark.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Age variation significantly impacts face recognition accuracy.
• Existing methods like age-invariant face recognition (AIFR) and face age synthesis (FAS) have limitations: AIFR lacks interpretability, and FAS can introduce artifacts compromising recognition.
• A unified approach is needed to address both challenges effectively.

Purpose of the Study:

• To develop a unified, multi-task framework (MTLFace) for joint age-invariant face recognition (AIFR) and face age synthesis (FAS).
• To enhance identity-related feature extraction for AIFR while generating visually interpretable synthetic faces via FAS.
• To improve the performance and interpretability of face recognition systems under varying ages.

Main Methods:

• Propose an attention-based feature decomposition to separate identity and age features spatially.
• Introduce an identity-conditional module for identity-level FAS, improving age smoothness.
• Utilize synthesized faces from FAS to fine-tune AIFR via a selective strategy.
• Collect and release a large cross-age face dataset and a benchmark for tracing missing children.

Main Results:

• MTLFace achieves superior performance on AIFR and FAS tasks across five benchmark cross-age datasets.
• The framework demonstrates competitive performance on general face recognition datasets ('in the wild').
• The proposed methods effectively enhance both the accuracy of age-invariant recognition and the quality of synthesized faces.

Conclusions:

• The unified MTLFace framework effectively handles age variation in face recognition and synthesis.
• MTLFace offers improved identity representation and realistic face synthesis, benefiting model interpretation.
• The released dataset and benchmark facilitate further research in cross-age face analysis and child tracing.