Evaluating the Viability of Synthetic Pre-training Data for Face Recognition Using a CNN-Based Multiclass Classifier

Sammanfattning: Today, face recognition is becoming increasingly accurate and faster with deep learning methods such as convolutional neural networks (CNNs), and is now widely used in areas such as security and entertainment. Typically, these CNNs are trained using real-face datasets like CASIA-WebFace, which was put together using web-crawling of IMDB. This can, however, lead to privacy and bias issues. Synthetic datasets made up of computer generated pictures, such as DigiFace-1M, created by Microsoft and the University of Cambridge, provide alternatives, offering large volumes of unbiased training data that respect privacy. Despite these advances, there’s been little research comparing the performance of models pre-trained on synthetic datasets versus traditional ones. In this study, we address this gap. We tested a ResNet-18 model pre-trained on a subset of real-face images from CASIA-WebFace against one trained on a subset of the synthetic DigiFace-1M dataset. We also compared these results to a model pre-trained on ImageNet, a large, general-purpose, mixed object dataset, and a model without any pre-training. The models were later evaluated on the first 100 identities in CASIA-WebFace. The findings showed that while the best performance came from pre-training on real-face datasets, the synthetic dataset also offered a viable option for multiclass face recognition. The synthetic dataset showed slightly better performance than ImageNet and significantly better performance than the model without pre-training, all while avoiding privacy issues linked to web-crawled images. More research is needed to further explore whether classification models pre-trained on larger synthetic datasets like DigiFace-1M can significantly outperform broader datasets like ImageNet, or even improve upon real-face pre-training.