Papers
Topics
Authors
Recent
Search
2000 character limit reached

When Age-Invariant Face Recognition Meets Face Age Synthesis: A Multi-Task Learning Framework

Published 2 Mar 2021 in cs.CV | (2103.01520v2)

Abstract: To minimize the effects of age variation in face recognition, previous work either extracts identity-related discriminative features by minimizing the correlation between identity- and age-related features, called age-invariant face recognition (AIFR), or removes age variation by transforming the faces of different age groups into the same age group, called face age synthesis (FAS); however, the former lacks visual results for model interpretation while the latter suffers from artifacts compromising downstream recognition. Therefore, this paper proposes a unified, multi-task framework to jointly handle these two tasks, termed MTLFace, which can learn age-invariant identity-related representation while achieving pleasing face synthesis. Specifically, we first decompose the mixed face feature into two uncorrelated components -- identity- and age-related feature -- through an attention mechanism, and then decorrelate these two components using multi-task training and continuous domain adaption. In contrast to the conventional one-hot encoding that achieves group-level FAS, we propose a novel identity conditional module to achieve identity-level FAS, with a weight-sharing strategy to improve the age smoothness of synthesized faces. In addition, we collect and release a large cross-age face dataset with age and gender annotations to advance the development of the AIFR and FAS. Extensive experiments on five benchmark cross-age datasets demonstrate the superior performance of our proposed MTLFace over existing state-of-the-art methods for AIFR and FAS. We further validate MTLFace on two popular general face recognition datasets, showing competitive performance for face recognition in the wild. The source code and dataset are available at~\url{https://github.com/Hzzone/MTLFace}.

Citations (76)

Summary

  • The paper introduces MTLFace, a multi-task framework that jointly performs age-invariant face recognition and face age synthesis using attention-based feature decomposition.
  • It employs an Identity Conditional Module and weight-sharing strategy to preserve identity details while synthesizing high-quality age-progressed and regressed facial images.
  • Experimental results show that MTLFace outperforms state-of-the-art methods across benchmarks like AgeDB and CALFW, proving its robust performance.

An Overview of MTLFace: A Unified Multi-Task Learning Framework for Age-Invariant Face Recognition and Face Age Synthesis

The paper "When Age-Invariant Face Recognition Meets Face Age Synthesis: A Multi-Task Learning Framework" presents MTLFace, a robust framework that simultaneously addresses the challenges of age-invariant face recognition (AIFR) and face age synthesis (FAS). This initiative is driven by the need for accurate face recognition systems that function effectively across significant age variations, a task that has traditionally been handled separately through either discriminative models focusing on age-invariant features or generative models aiming to synthesize uniform age representations.

Technical Contributions and Methodology

MTLFace ingeniously consolidates AIFR and FAS into a single, multi-task learning framework. The core of the methodology lies in the decomposition of facial features into uncorrelated identity- and age-related components. This process employs an attention-based mechanism to support more precise feature separation. Unlike previous models that operate at the feature vector level, this method manipulates feature maps, enhancing spatial resolution vital for high-quality face synthesis.

A crucial innovation in this paper is the Identity Conditional Module (ICM), which leverages identity-related features to direct age synthesis at an individual level rather than across generalized age groups. This specificity surpasses existing methods that utilize one-hot encoding, offering an identity-preserved synthesis that circumvents common artifacts and errors in identity changes during synthesis tasks. The weight-sharing strategy within the ICM promotes smooth visual transitions across age progressions and regressions.

The proposed model is trained using a large cross-age face dataset collected by the authors, designed to support both AIFR and FAS tasks. This dataset is more comprehensive than existing ones, containing 1.7 million faces annotated with age and gender information. The multi-task training includes components of age estimation to guide the attention mechanism and face recognition tasks to enforce identity preservation, while a gradient reversal layer aids in learning age-invariant identity features.

Experimental Validation

The paper demonstrates that MTLFace outperforms several state-of-the-art methods across multiple benchmark datasets, including AgeDB, CALFW, CACD-VS, and FG-NET. Notably, it achieves strong numerical results in both verification and recognition rates, affirming the effectiveness of its multi-task approach in addressing the complexities of age variation in facial recognition.

Moreover, the application of MTLFace to general face recognition tasks, such as those in the LFW and MF1 datasets, showcases its robustness and generalizability beyond cross-age scenarios. The architecture maintains high recognition accuracy without compromising the interpretability granted by the generative visual outputs.

Implications and Future Directions

Practically, MTLFace offers substantial advancements for applications requiring accurate age-invariant facial recognition, such as security systems and identification of missing persons. Theoretically, the novel decomposition approach and continuous domain adaptation framework suggest a promising direction for enhancing model interpretability while ensuring high performance.

Potential future developments could explore further refinement of the attention mechanism for even more precise feature decomposition or extend the framework to handle other confounding variables in facial recognition, such as illumination changes or occlusions. Additionally, the public release of their cross-age dataset serves as an invaluable resource for advancing research across related computational vision tasks.

MTLFace represents a significant stride in multi-task learning for both theoretical exploration and practical application in facial recognition technology. Its comprehensive approach to solving age-related challenges paves the way for more adaptable and accurate recognition systems.

Paper to Video (Beta)

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Continue Learning

We haven't generated follow-up questions for this paper yet.

Collections

Sign up for free to add this paper to one or more collections.