Abstract:

Concerning the limitations of existing filtering methods in dealing with point cloud models with high complex geometric structure， where feature blurring artifacts degrade filtering performance， a point cloud filtering network based on adaptive feature extraction and feature fusion strategy named PFRNet （Point cloud Feature Regularization fusion Network） was proposed. Firstly， the feature information between different neighborhoods was learned by the adaptive spatial feature extractor， so as to capture the local neighborhood features of different dimensions and reduce the loss of local details. Then， the global bilinear response was introduced from the local information of the point cloud through the local feature regularized fusion， and it was regularized and fused to weaken the common features of the point cloud and enhance the sharp features. Finally， the self-correlation attention decoder was used to enhance the connection between different neighborhoods in the decoding process， improve the global perception ability of the model， and better extracting local geometric features. Experimental results show that compared with Pointfilter， PFRNet reduces Chamfering Distance （CD） and Mean Square Error （MSE） of 7.45% and 4.99%，respectively. Visualization results further confirm that PFRNet generates a point cloud model closer to the real one than other methods.

Key words: point cloud filtering, local feature, dynamic graph edge convolution, feature fusion, attention mechanism

摘要：

针对现有滤波方法在处理具有高复杂度几何结构的点云模型中存在特征模糊现象，导致最终滤波效果较差的问题，设计一种基于自适应特征提取和特征融合策略的点云滤波网络PFRNet（Point cloud Feature Regularization fusion Network）。首先，通过自适应空间特征提取器学习不同邻域之间的特征信息，从而捕获不同维度的局部邻域特征，减少局部细节的丢失；其次，通过局部特征正则化融合从点云的局部信息中引入全局双线性响应，并对它进行正则化融合，削弱点云的共性特征，增强尖锐特征；最后，通过自相关注意力解码器在解码过程中增强不同邻域之间的联系，提升模型的全局感知能力，以更好地提取局部几何特征。实验结果表明，与Pointfilter相比，PFRNet的倒角距离（CD）和均方误差（MSE）分别降低了7.45%、4.99%；可视化结果显示，PFRNet相较于其他方法能够生成更接近真实的点云模型。

关键词: 点云滤波, 局部特征, 动态图边卷积, 特征融合, 注意力机制