Generating photo-realistic training data to improve face recognition accuracy.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Convolutional Neural Networks (CNNs) achieve high accuracy in face recognition but require extensive training data.
• Limited availability of large-scale face datasets hinders the performance of CNN-based systems.
• Synthetic data augmentation is explored as a solution to dataset limitations.

Purpose of the Study:

• Investigate the feasibility of using synthetic data to augment face recognition datasets.
• Propose a novel Generative Adversarial Network (GAN) capable of disentangling identity and non-identity attributes.
• Enhance face recognition accuracy by augmenting datasets with generated synthetic images.

Main Methods:

• Developed a novel GAN that disentangles identity and non-identity attributes.
• Trained an embedding network to map identity labels to a latent space.
• Conditioned the GAN on samples from the latent space to generate synthetic face images.
• Generated both in-dataset and new subject synthetic images for augmentation.

Main Results:

• Generated photo-realistic synthetic face images using advanced GAN training techniques.
• Demonstrated increased face recognition accuracy when training with augmented datasets.
• Showed that the proposed method is particularly effective for augmenting smaller datasets.
• Achieved an absolute accuracy improvement of 8.42% on a dataset with fewer than 60,000 images.

Conclusions:

• Synthetic data augmentation using the proposed GAN is a viable method to improve face recognition performance.
• The novel GAN effectively generates diverse and realistic synthetic face images.
• Augmenting smaller datasets with synthetic data offers significant accuracy gains in face recognition systems.