Abstract: In CAD,B-rep is widely used for its precise representation,but its unstructured nature limits deep learning models based on it.Existing methods focus heavily on geometric information,improving performance but increasing complexity and data cost,while insufficient topology attention restricts generalization and overall effectiveness.To address the above issues,a hierarchical mask and multi-scale feature fusion-based network was proposed,which supported both model classification and segmentation.Specifically,a hierarchical mask geometric encoder was designed to eliminate information redundancy,reducing the network's over-reliance on information and enhancing robustness and generalization.Furthermore,a multi-scale adaptive topological encoder was introduced to adaptively extract multi-scale features,capturing topological structure information from local to global.Extensive experiments demonstrated the advanced performance and effectiveness of the new model on multiple public classification and segmentation datasets.

Key words: boundary representation, model classification, segmentation, graph convolutional neural network, mask, topology

摘要： 在CAD领域,边界表示(B-rep)因其精确性被广泛采用,但其非结构化特性使基于B-rep的深度学习模型较少。现有方法多侧重几何信息的描述,虽提升性能,却增加了复杂度和数据提取成本,且对拓扑信息关注不足,限制了模型的泛化能力和整体表现。为解决上述问题,本文提出了一种基于层次掩码及多尺度特征融合的CAD模型表征网络,同时支持模型分类和分割。具体而言,通过设计层次掩码几何编码器来消除信息冗余,减少网络对信息的过度依赖,提高了网络的鲁棒性和泛化性;同时,通过设计多尺度自适应拓扑编码器来自适应地提取多尺度特征,能够捕获从局部到全局的拓扑结构信息。广泛实验证明了新模型在多个公开分类和分割数据集上的先进性和有效性。

关键词: 边界表示, 模型分类, 分割, 图卷积神经网络, 掩码, 拓扑