TransX2GAN: Using Self-Attention and Knowledge Transfer to Generate Face Images

The success of deep learning relies on huge datasets. However, in some domains, data is not always accessible due to various constraints. Synthetic data generation helps solve the problem. Collecting face images is challenging due to issues such as privacy. Another issue is that models trained on larger datasets are generally not reused for other related tasks. Therefore, the research aims to facilitate the problems in current face image generation models, such as the lack of quality, huge computational burden, and lack of ability to use pre-trained models. The proposed work aims to utilize a Generative Adversarial Network (GAN) model that uses self-attention and knowledge transfer to generate images of faces resembling images from the CelebA dataset. The proposed model is named TransX2GAN. The ramifications are reduced training time, computational burden, and the ability to reuse deep learning models. Due to the use of self-attention, the quality of generated images has also improved considerably when compared to other models. Compared to the other models the Fréchet Inception Distance score and training time has been enhanced by 11.90% and 13.40%, respectively. Hence using transfer learning with attention alleviates some of the important problems in face image generation.