Multi-Object Vector Graph Node Classification Based on Graph Neural Networks

Vector graph is an image format that is widely used in engineering. Due to the geometric primitive representation based on mathematical equations, vector graphs have excellent characteristics such as lossless scaling, efficient storage, and convenient editing. Compared to processing bitmap renderings, abstracting vector graphics using graph structures is more efficient in deep learning models. Based on it, the classification of multiple objects in vector graphs is realized by classifying graph nodes. We implement a complete processing chain that includes graph structuring, subgraph partitioning, data enhancement, etc. Furthermore, we design a special graph neural network called VGPnet (Vector Graph-based Point Network), which uses multiple layers of graph networks to maintain local information of the nodes and extracts global information by reading out the subgraphs. Finally, both are combined to obtain the prediction results. Experiments demonstrate the feasibility of our model and show its superiority in vector graphs of two different fields, as compared to other graph neural networks, VGPnet achieves better recall, precision, and F1 metrics.