This research project combines self-attention mechanism with capsule network for high-quality image generation. Experimental results on MNIST and Cifar-10 proves the superiority of our proposed adversarial network against other convolutional generative networks.
The contributions of the research projects are as follows :
- The proposed adversarial network leverages self-attention mechanism to learn long term information as well as acquire bottleneck details through capsule module.
- The featured generative model mainly focuses on representation learning for the generator and discriminator part that makes it more robust than convolutional generator and discriminator