| 摘 要: 基于图卷积的草图语义分割方法在图构建与传播过程中是扁平化的,邻域聚集的图卷积无法获得足够的全局信息。提出基于层次图卷积网络的草图语义分割方法,构建不同层次的图对节点特征传播或聚合。另外,为了缓解由于过多的图卷积模块导致训练过程过平滑的问题,引入了随机缩放特征和梯度的正则化方法。与基准的图卷积模型相比,改进的模型在CreativeSketch数据集上比次优的SketchGNN模型,P-Metric平均高出1.42个百分点,C-Metric平均高出2.87个百分点。研究结果表明,该模型可以有效提高分割准确率。 |
| 关键词: 矢量草图 图节点分类 图神经网络 层次图卷积 语义分割 |
|
中图分类号: TP391
文献标识码: A
|
|
| Sketch Semantic Segmentation Method Based on Hierarchical Graph Convolutional Network |
|
JIN Jiahui, WU Zizhao, WANG Yigang
|
(School of Humanities, Arts and Digital Media, Hangzhou Dianzi University, Hangzhou 310018, China)
jjh000902@163.com; wuzizhao@hdu.edu.cn; yigang.wang@hdu.edu.cn
|
| Abstract: Graph convolution-based sketch semantic segmentation methods employ a flat graph construction and propagation process, where neighborhood aggregation in graph convolution fails to capture sufficient global information. This paper proposes a sketch semantic segmentation method based on a hierarchical graph convolutional network, which constructs graphs at different levels to propagate or aggregate node features. Additionally, to mitigate the ove-r smoothing issue during training caused by excessive graph convolution modules, a regularization method involving random scaling of features and gradients is introduced. Compared to the baseline graph convolution model,the improved model compared with suboptimal SketchGNN model achieves an average increase of 1.42 percentage points in P-Metric and 2.87 percentage points in C-Metric on the CreativeSketch dataset. The research results
demonstrate that the proposed model effectively enhances segmentation accuracy. |
| Keywords: vector sketch graph node classification graph neural network hierarchical graph convolution semantic segmentation |
