The image to image translation has been a different task to change a particular aspect of a given image to another, such as changing the facial expression of a person having transitional expressions depending upon the moods.
The task of the image to image translation has experienced significant improvements with the following generative adversarial networks that result from changing hair color to seasons of scenery images and also ranging its efficiency by reconstructing photos from edge maps.
These GAN models learn to translate images from one domain to the other from the given image which can denote the attribute terms as a meaningful feature inherent in an image such as gender or age, hair color, and attribute value as a particular value of an attribute.
We can further denote this domain as a set of images that share the same attribute value.
These domain settings can enable us to perform multi-domain image-to-image translation and more interesting tasks which changes images according to the relevant attributes from multiple domains.
However, existing models are both ineffective and inefficient in such multi-domain image translation tasks. The fully utilized training data is likely to limit the quality of generated images and has also resulted in failure.
Furthermore, they turned incapable of training domains jointly from different datasets where each dataset is partially labeled.
As a solution to such problems, StarGAN, a scalable and novel approach has the capacity to learn mappings among multiple domains. This model takes training data of multiple domains and learns the mappings using a single generator between all available domains.
This simple idea of a generator takes both domain information and image as inputs instead of learning a fixed translation and learns to translate the image into the corresponding domain flexibly. To represent domain information this model uses a label as a binary or one-hot vector.
During training, a target domain label is randomly generated and trains the model to flexibly translate an input image into the target domain. By doing so, the domain label can be controlled and can translate the image into any desired domain at the testing phase.
Remarkable results have been shown through Generative adversarial networks (GANs) in various computer vision tasks such as image translation, image generation, super-resolution imaging, and face image synthesis.
A typical GAN model consists of two modules, namely a discriminator and a generator. The discriminator would learn to distinguish between fake and real samples, whereas the generator would learn to generate fake samples that are indistinguishable from real samples.
The proposed approach leverages the adversarial loss in order to make the generated images as realistic as possible.
Both the discriminator and generator have provided the prior studies with class information to generate samples conditioned on the class. Other recent approaches focused on generating particular images which are highly relevant to a given text description.
The idea of conditional image generation has also been successfully applied to domain transfer, superresolution imaging and photo editing. A scalable GAN framework can flexibly steer the image translation to various target domains just by providing conditional domain information.