Abstract
Face recognition systems often exhibit a performance gap between genders, typically performing better on men due to dataset biases. This paper introduces FaceGenderID, a "gender-aware" pipeline that utilizes gender-specific triplet loss training to fine-tune state-of-the-art models. By routing images through specialized convolutional neural networks (CNNs) based on estimated gender, the system significantly improves verification accuracy for both male and female subjects.
Introduction
While Deep Convolutional Neural Networks (DCNNs) have revolutionized face recognition, they frequently inherit demographic biases from training data. Large-scale datasets like VGGFace2 are often skewed toward male and Caucasian subjects, leading to a "demographic gap." Models become highly proficient at identifying males but struggle with females, who often present higher intra-class variability due to factors like makeup and hairstyles. FaceGenderID addresses this by explicitly incorporating gender information into the training and inference stages.
Proposed Methodology
The FaceGenderID system employs a modular architecture designed to treat gender as a distinct domain.
The pipeline consists of four primary stages:
- Preprocessing & Estimation: Faces are detected and passed to a fine-tuned ResNet-50 Gender Estimator.
- Model Selection: Based on the estimation, images are routed to specialized models: CNN Male, CNN Female, or a CNN Balanced model (used for cross-gender comparisons).
- Triplet Loss Fine-Tuning: The core innovation involves fine-tuning feature embeddings using a triplet loss algorithm. This process minimizes the distance between an "Anchor" and a "Positive" image of the same person while maximizing the distance to a "Negative" image of a different person within the same gender category.
- Similarity Computation: Euclidean distance between embeddings determines if two images represent the same identity.
Experimental Setup
The researchers utilized the VGGFace2 dataset, which contains over 3.3 million images. They noted a significant bias, with the dataset being 59.3% male. To train the specialized DCNNs, they created balanced triplet sets (75,000 triplets per gender) to ensure the models learned gender-specific features effectively.
Results and Discussion
The FaceGenderID system consistently outperformed baseline models. By specializing models for specific gender domains, the network learns to ignore gender-specific "noise" and focus on identity-distinguishing features.
Key performance metrics include:
- ResNet-50 Architecture: AUC improved from a 97.4% baseline to 97.7% with automatic gender estimation.
- VGG-Face Architecture: AUC rose from 94.0% to 94.9%.
- Closing the Gap: The "CNN Female" model showed a 2.1% AUC improvement for VGG-Face, indicating that specialized training is particularly effective for the demographic that typically suffers from lower baseline performance.
Conclusion
FaceGenderID demonstrates that gender-dependent domains can significantly enhance face recognition accuracy. While the current system relies on fine-tuning, future work may involve training architectures from scratch on balanced datasets. Additionally, addressing the "Gender Estimator" bottleneck and expanding the approach to other covariates like ethnicity and age remain critical steps toward developing truly fair biometric systems.
Summary updated