Graph neural network-based visual design element association analysis and style transfer model

Abstract

The growing demand for personalized design has driven the development of advanced models for visual design element association and style transfer. However, existing methods struggle to capture complex relationships among design elements and maintain aesthetic coherence during style transfer. The research proposes a novel Modified Zebra-Driven Graph Convolutional Neural Network (ModZ-GConvoNNet) for analyzing and associating visual design elements, with a focus on improving style transfer in design tasks. A comprehensive dataset of visual design elements & images is curated using Kaggle datasets, including the COCO (123k), DTD (5.7k), and WikiArt (81.4k) datasets. Data represent diverse visual styles, textures, patterns, and layout structures from domains such as fashion, interior design, and web aesthetics. Min- Max normalization and Augmentation techniques like rotation, flipping, and scaling enhance model robustness to layout variations and the brightness and contrast of design images. Feature extraction uses canny edge detection to identify boundaries and texture mapping to capture finer details, enabling more accurate design. The GConvoNNet model uses graph-based representations to capture relationships between design elements, enhancing the understanding of their interdependencies. It incorporates the ModZ mechanism to tune the parameters of the graph convolution layers and improves learning of visual element relationships for effective style transfer. The ModZ-GConvoNNet implemented in Python 3.8 outperforms traditional style transfer methods with an SSIM of 0.972 and a PSNR of 39.215, indicating enhanced visual design style transfer and structural preservation. It provides a versatile tool for automating systems and a creative professional with a significant leap forward in design automation.