Perceptual Faces Completion Using SelfAttention Generative Adversarial Networks

This paper propose method based on self-attention generative adversarial networks (SAGAN) to accomplish the task of image completion wherever completed images become globally and domestically consistent. Using self-attention GANs with contextual and different constraints, the generator will draw realistic images, wherever fine details are generated within the damaged region and coordinated with the entire image semantically. To train the consistent generator, i.e. image completion network, this paper tend to use global and native discriminators wherever the global discriminator is responsible for evaluating the consistency of the whole image, whereas the local discriminator assesses the local consistency by analyzing local areas containing completed regions only. Last but not least, an attentive recurrent neural block is introduced to get the attention map regarding the missing part within the image, which is able to facilitate the subsequent completion network to fill content better. By comparison of the experimental results of various approaches on CelebA data set, our technique shows relatively good results. Traditional Convolutional GANs generate high-resolution details as a function of only spatially local points in lower-resolution feature maps. In SAGAN, details will be generated using cues from all feature locations. Moreover, the discriminator will make sure highly elaborate features in distant parts of the image are consistent with one another. Moreover, recent work has shown that generator conditioning affects GAN performance. Investing this insight, this paper tend to apply spectral normalization to the GAN generator and find that this improves training dynamics.