DArFace: Deformation Aware Robustness for Low Quality Face Recognition

TL;DR

DArFace tackles the problem of degraded performance in face recognition under real-world low-quality conditions by integrating deformation-aware training that combines differentiable global transformations and local elastic deformations. The method uses an adversarial loop to search for challenging deformations via $\psi$ and optimizes a composite loss $\mathcal{L}_{total} = \mathcal{L}_{clean} + \mathcal{L}_{trans} + \lambda_{cont} \mathcal{L}_{cont}$, where $\mathcal{L}_{cont}$ enforces identity consistency across views through a contrastive objective. Key contributions include the deformation parameterization $\psi=(\varphi, \Delta u, \Delta v, \lambda, \alpha, \sigma)$, differentiable global/local transforms, and a 75/15/10 sampling strategy, enabling robust recognition on TinyFace, IJB-B, and IJB-C without paired high/low-quality data. Extensive experiments show that DArFace achieves state-of-the-art or competitive results across high-, mixed-, and low-quality benchmarks while preserving strong performance on clean images, highlighting its practical impact for surveillance and standoff imaging scenarios.

Abstract

Facial recognition systems have achieved remarkable success by leveraging deep neural networks, advanced loss functions, and large-scale datasets. However, their performance often deteriorates in real-world scenarios involving low-quality facial images. Such degradations, common in surveillance footage or standoff imaging include low resolution, motion blur, and various distortions, resulting in a substantial domain gap from the high-quality data typically used during training. While existing approaches attempt to address robustness by modifying network architectures or modeling global spatial transformations, they frequently overlook local, non-rigid deformations that are inherently present in real-world settings. In this work, we introduce \textbf{DArFace}, a \textbf{D}eformation-\textbf{A}ware \textbf{r}obust \textbf{Face} recognition framework that enhances robustness to such degradations without requiring paired high- and low-quality training samples. Our method adversarially integrates both global transformations (e.g., rotation, translation) and local elastic deformations during training to simulate realistic low-quality conditions. Moreover, we introduce a contrastive objective to enforce identity consistency across different deformed views. Extensive evaluations on low-quality benchmarks including TinyFace, IJB-B, and IJB-C demonstrate that DArFace surpasses state-of-the-art methods, with significant gains attributed to the inclusion of local deformation modeling.

Figures (2)

• Figure 1: Comparison of our method with the current methods: (A) clean-only training uses only original images $x$, (B) saadabadi2024ARoFace uses clean + globally transformed $T_g(x)$ images for training, (C) Our method (DArFace) uses clean, globally $T_g(x)$ (e.g., rotated) and locally $T_l(x)$ (e.g., elastic deformed eye and lip region) transformed images for improved robustness.
• Figure 2: Overview of DArFace training framework. Step1 (Top Row- Adversarial Search): A clean image $x$ is transformed $T_{\psi}$ and passed through a frozen network to search for global and local parameters such that the loss $\mathcal{L}$ maximally increases. Step2 (Bottom Row-Forward Pass): Clean, globally (e.g. rotated) and locally (warped eye and lip) transformed $(x, T_g(x),T_l(x))$ inputs are fed to the model for training. Step3 (Bottom Row-Backward Pass): We minimize the total loss $\mathcal{L}_{clean}+\mathcal{L}_{trans}+\lambda_{cont}\mathcal{L}_{cont}$, i.e. angular loss on clean and transformed views and a contrastive loss that aligns global and local feature pairs.