Face Image Generation Using Knowledge Transfer and Self-Attention

Date of Award


Publication Type


Degree Name



Computer Science


Deep learning, Datasets, Face image generation, Generative adversarial networks


I. Ahmad


A. Ngom



Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.


The success of deep learning relies on huge datasets and access to immense computational power. However, in some domains, data is not always accessible due to privacy concerns or lack of mechanisms to collect and safely store data safely, among many other constraints. Data generation has a lot of real-world applications that are mentioned below, however a major reason behind data generation is to alleviate the problem of data scarcity. One such deep learning application that faces data scarcity is image generation, specifically face image generation, due to issues such as privacy. Face image generation has many applications, such as creating huge datasets for applications like image super-resolution, image completion, and image matting. Real-world application of face image generation includes video surveillance, biometrics, and in medical applications such as mapping facial features to genetic data. Generative adversarial networks (GANs) have been used to generate images in the recent literature. In this work, GANs are used for the task of face image generation. GANs have two networks: a discriminator network, and a generator network. The discriminator distinguishes real data from fake data. Real data is data coming from the dataset, and fake data is the data coming from the generator. The generator generates synthetic data. The proposed work aims to utilize attention and knowledge transfer to generate images of faces resembling images from the CelebA dataset. The unique point of the proposed solution is the concept of using pre-trained models as initial models for generating images in domains that lack huge amounts of data. The proposed work uses an autoencoder-based model previously trained on a related dataset. The trained layers from this pre-trained model is used in both the discriminator and the generator to eliminate the need for training from scratch. The benefits are reduced training time, computational burden, and the ability to reuse deep learning models. In addition to knowledge transfer, self-attention is also used in the GAN model to improve the quality of generated images. Self-attention drastically reduces the number of model parameters that need to be learnt separately. It also helps learn filters with more focus on the important parts of the image, which is especially important for facial image generation.