Synthetic Forehead-creases Biometric Generation for Reliable User Verification

TL;DR

The paper addresses the scarcity of forehead-crease data for biometric verification under mask conditions by introducing a two-module diffusion-based data synthesis framework. The Subject-Specific Generation Module (SSGM) uses a Brownian Bridge Diffusion Model to produce diverse, identity-preserving, pose-conditioned synthetic samples, while the Subject-Agnostic Generation Module (SAGM) leverages an unconditional DDPM to generate new identities for privacy-preserving augmentation. An FHCVS verification backbone based on ResNet-18 with attention modules and AdaFace loss evaluates the utility of the synthetic data, with experiments showing substantial improvements when incorporating subject-specific synthetic data, achieving a cross-database EER as low as around 9.38%. The results demonstrate the value of diffusion-based synthetic data for forehead-creases biometrics, offering privacy-friendly augmentation that enhances robustness against data scarcity and pose variation, and the work contributes reproducible code and datasets to the community.

Abstract

Recent studies have emphasized the potential of forehead-crease patterns as an alternative for face, iris, and periocular recognition, presenting contactless and convenient solutions, particularly in situations where faces are covered by surgical masks. However, collecting forehead data presents challenges, including cost and time constraints, as developing and optimizing forehead verification methods requires a substantial number of high-quality images. To tackle these challenges, the generation of synthetic biometric data has gained traction due to its ability to protect privacy while enabling effective training of deep learning-based biometric verification methods. In this paper, we present a new framework to synthesize forehead-crease image data while maintaining important features, such as uniqueness and realism. The proposed framework consists of two main modules: a Subject-Specific Generation Module (SSGM), based on an image-to-image Brownian Bridge Diffusion Model (BBDM), which learns a one-to-many mapping between image pairs to generate identity-aware synthetic forehead creases corresponding to real subjects, and a Subject-Agnostic Generation Module (SAGM), which samples new synthetic identities with assistance from the SSGM. We evaluate the diversity and realism of the generated forehead-crease images primarily using the Fréchet Inception Distance (FID) and the Structural Similarity Index Measure (SSIM). In addition, we assess the utility of synthetically generated forehead-crease images using a forehead-crease verification system (FHCVS). The results indicate an improvement in the verification accuracy of the FHCVS by utilizing synthetic data.

Figures (7)

• Figure 1: (a) Capturing face images with forehead-crease using mobile devices, (b) Identifying the forehead region of interest (ROI) using YOLO-V3, (c) Generating subject-specific forehead creases images (d) Utilizing both real and synthetic forehead-crease images to train the forehead-creases verification system.
• Figure 2: Our proposed framework consists of two modules: Subject-Specific Generation Module (SSGM) and Subject-Agnostic Generation Module (SAGM). SSGM obtains image pairs for each subject from a forehead-crease image FH-V1 dataset bharadwaj2022mobile and its augmentations to train an Image-to-Image translation-based diffusion model li2023bbdm. $I_{i}^{j}$, i.e. the $j^{th}$ real pose of subject $i$, acts as a starting point $x_T$ for the pre-trained BBDM and generates subject-specific synthetic data. The SAGM samples random forehead creases from an unconditional DDPM ho2020denoising trained on unlabeled forehead creases. These random samples act as an input to the sampler in SSGM, generating new identities. The synthetic datasets, merged with the real and augmented data are subsequently used as input to the forehead-creases verification backbone bharadwaj2022mobile.
• Figure 3: Sampling Subject-Specific Synthetic Data: It sets a pose $I_{i}^{j}$ as a starting point to the BBDM li2023bbdm sampler, trained on forehead-creases image pairs. After gradual denoising, the final obtained synthetic image ${}^{syn}_{k} I_{i}^{j}$ is added to the training data of the corresponding identity $i$.
• Figure 4: Subject-Specific (SS) Samples: FH-V1 Real images (first column) of different subjects (IDs: 70, 195, 203, 207, and 239) and their synthetic counterparts (subsequent columns), highlighting the diversity in synthetic image generation.
• Figure 5: Subject-Agnostic (SA) Synthetic Data: Random sample from DDPM (left) and corresponding subject-specific samples (rest).