PROJECT TITLE :

Gated-GAN Adversarial Gated Networks for Multi-Collection Style Transfer

ABSTRACT:

Image semantic content is rendered in numerous artistic styles via style transfer. Since recently, GANs have emerged as an effective technique to style transfer by adversarially training the generator to make plausible counterfeits. Nevertheless, the mode collapse problem in classic GAN causes training to be unstable, making style transfer quality impossible to ensure. As a result, many GANs must be trained to give consumers the option of transferring more than one style from one GAN generator to another. The issues and limitations of style transmission are addressed in this work. Multiple styles can be transferred in a single model using adversarial gated networks (Gated-GANs). Encoder, gated transformer, and decoder are the three modules in the generative networks. The gated transformer may be used to create a variety of styles by varying the images that are sent into it. Auto-encoders are used to improve training stability by combining the encoder and decoder. Discriminative networks can tell if a picture is stylised or authentic based on its input. In order to enable the generative networks generate images in multiple styles, an auxiliary classifier is utilised to identify the style categories of transmitted images. As a further benefit, Gated-GAN allows for the exploration of styles learned from other artists or genres. The proposed paradigm for multi-style transfer is stable and effective, according to our comprehensive testing.