Image Generation Using StyleVGG19-NST Generative Adversarial Networks

Abstract

Creating new image styles from the content of existing images is challenging to conventional Generative Adversarial Networks (GANs), due to their inability to generate high-quality image resolutions. The study aims to create top-notch images that seamlessly blend the style of one image with another without losing its style to artefacts. This research integrates Style Generative Adversarial Networks with Visual Geometry Group 19 (VGG19) and Neural Style Transfer (NST) to address this challenging issue. The styleGAN is employed to generate high-quality images, the VGG19 model is used to extract features from the image and NST is used for style transfer. Experiments were conducted on curated COCO masks and publicly available CelebFace art image datasets. The outcomes of the proposed approach when contrasted with alternative simulation techniques, indicated that the CelebFace dataset results produced an Inception Score (IS) of 16.57, Frecher Inception Distance (FID) of 18.33, Peak Signal-to-Noise Ratio (PSNR) of 28.33, Structural Similarity Index Measure (SSIM) of 0.93. While the curated dataset yields high IS scores of 11.67, low FID scores of 21.49, PSNR of 29.98, and SSIM of 0.98. This result indicates that artists can generate a variety of artistic styles with less effort without losing the key features of artefacts with the proposed method.