Enhancing Face Similarity Detection with GAN-Generated Synthetic Data

Abstract

Face similarity detection is a critical task in computer vision, particularly for applications like facial recognition and verification. Training robust models for this task often requires large, diverse datasets, which can be costly and time-consuming to acquire. This paper explores the potential of Generative Adversarial Networks (GANs) to augment training data with synthetic face images. We compare three GAN-based approaches-StyleGAN2 with labels, StyleGAN2 without labels, and custom GAN models-and evaluate their impact on CNN model performance across three scenarios: mixed datasets (real + generated), generated-only datasets, and fine-tuned models (first trained on real data, then fine-tuned with generated images). Our results show that while integrating GAN-generated data can enhance model performance in certain configurations, particularly when combined with real data, it may also lead to degradation. Notably, only two out of nine evaluated scenarios outperformed the baseline model trained solely on real data. StyleGAN2 without labels achieved the highest accuracy (87.90%) in the mixed-data setting and 87.06% with synthetic-only training, whereas fine-tuning the real-data model with these synthetic images produced the lowest accuracy (81.79%). These findings suggest that a balanced approach to integrating synthetic and real data is crucial for optimal performance.