Multi-Granularity Graph Representation and Augmentation for Image

Graph Neural Networks (GNNs) process graph- structured data by learning features and patterns through neural networks, applied to tasks such as clustering, classification, and prediction. Current research primarily employs image-based data augmentation methods, but these methods struggle to capture the complex structures of graphs and multi-dimensional node data. This study proposes a graph data augmentation method within a multi-granularity computational framework, including random cropping, random flipping, and the dropping of nodes and edge relationships, aiming to enhance the learning and generalization capabilities of GNN models. Experimental results indicate that these methods significantly improve model robustness and data distribution, demonstrating good training effects and interpretability. It is hoped that these methods will play a role in computer vision tasks using GNNs.