WBT-GAN: Wavelet based Generative Adversarial Network for Texture Synthesis
Sara Saberi Tehrani Moghadam, Reza Azmi, Maral Zarvani · 2021
Generative Adversarial Networks (GANs) and their variants have become a state-art-of-the-art method in the field of texture transfer and synthesis. GANs are effective technics in computer graphics and have been widely used for image-to-image translation. Generation of high-resolution and realistic images is the main challenge of these networks as it depends on high-frequency features such as textures and edges. Wavelet Transform (WT) as one of most powerful time-frequency transformations in image processing is able to describe images at multi-level resolution. In this study, a generative model called WBT-GAN is proposed by using the four-level WT and employing objective function for defining of its loss function. In general, WBT-GAN is an extension of the existing network Texture-GAN. Experimental results showed that these changes have improved image resolution and sharpening and have led to better texture spread.