摘要
针对Graph Transformer比较擅长捕获全局信息,但对局部精细信息的提取不够充分的问题,将图卷积神经网络(GCN)引入Graph Transformer中,得到Graph Transformer and GCN (GTG)模块,构建了能够结合全局信息和局部信息的网格分割框架. GTG模块利用Graph Transformer的全局自注意力机制和GCN的局部连接性质,不仅可以捕获全局信息,还能够加强局部精细信息的提取.为了更好地保留边界区域的信息,设计边缘保持的粗化算法,可以使粗化过程仅作用在非边界区域.利用边界信息对损失函数进行加权,提高了神经网络对边界区域的关注程度.在实验方面,通过视觉效果和定量比较证明了采用本文算法能够获得高质量的分割结果,利用消融实验表明了GTG模块和边缘保持粗化算法的有效性.
A Graph Transformer and GCN(GTG)block was obtained by introducing graph convolutional neural network(GCN)into Graph Transformer because Graph Transformer was good at capturing global information,but weak in extracting local fine-grained information.A mesh segmentation framework that combined both global and local information was constructed.Global self-attention mechanism of Graph Transformer and local connectivity properties of GCN were used in GTG block in order to capture global information and enhance the extraction of local fine-grained information.An edge-preserving coarsening algorithm was designed to constrain the coarsening to non-boundary regions in order to better preserve information in boundary regions.Boundary information was used to weight the loss function to enhance the neural network’s focus on boundary regions.In experiments,visual results and quantitative comparisons prove that the proposed algorithm can achieve high-quality segmentation results,and ablation study demonstrates the effectiveness of GTG block and edge-preserving coarsening algorithm.
作者
张梦瑶
周杰
李文婷
赵勇
ZHANG Mengyao;ZHOU Jie;LI Wenting;ZHAO Yong(School of Mathematical Sciences,Ocean University of China,Qingdao 266100,China)
出处
《浙江大学学报(工学版)》
北大核心
2025年第5期912-919,共8页
Journal of Zhejiang University:Engineering Science
基金
山东省自然科学基金:资助项目(ZR2018MF006)
浙江大学CAD&CG国家重点实验室开放课题资助项目(A2228)
青岛市自然科学基金:资助项目(23-2-1-158-zyyd-jch)。
