| Summary: | In recent years, Generative Adversarial Networks (GANs) have gained worldwide interest and have marked a breakthrough in deep learning, encouraging detailed studies in generating artificial images. A new Generative Adversarial Networks (GAN) is proposed to unveil how Human visual perception takes place, focusing on how human beings perceive images, firstly, coarse structures and then their details. The network called 3G-AN consists of three generation stages and a single Discriminator. In this paper, a novel three-branch generator is proposed, which takes into account Coarse, Medium, and Fine structure of a given image. Coarse RGB decomposition image provides the general structure, while Medium RGB stage provides general-fine structure. Finally, Fine RGB decomposition provides fine details of the image. The proposal is tested on MNIST, CIFAR10, and Celebrity faces databases, generating realistic images with almost no anomalies. The RGB decomposition into coarse, medium, and fine, allows to understand the composition of an image from a structural point of view. The qualitative analysis carried out in this research paper outperforms the six most competitive models existing in the literature.
|
|---|