CSA-GAN: Cyclic synthesized attention guided generative adversarial network for face synthesis

摘要

Generative Adversarial Network (GAN) is one of the recent developments in the area of deep learning to transform the images from one domain to another domain. While transforming the images, we need to make sure that the background information should not influence the learning process. The attention-based networks are developed to learn the saliency maps and to prioritize the learning based on the important image regions. We develop a new Cyclic Synthesized Attention Generative Adversarial Network (CSA-GAN) in this paper by incorporating the cycle synthesized loss with the attention network. The use of attention guidance as well as cycle synthesis objective reduces the learning space more towards the optimum solution. It also improves the rate of convergence. The proposed method is tested for Sketch to Face synthesis over CUHK and AR benchmark datasets. We also experimented for thermal to visible face synthesis over WHU-IIP dataset. The proposed CSA-GAN observed promising performance for face synthesis in comparison with state-of-the-art GAN methods.