We propose a method for frequency domain feature analysis of point clouds based on the multilayer perceptron paradigm (MLP), named PFFA-MLP. By leveraging the Fourier transform and the latest MLP as a replacement for the transformer mechanism, PFFA-MLP maps point clouds to the frequency domain using a three-dimensional (3D) discrete Fourier transform. PFFA-MLP employs lightweight stacked DGCNN modules for initial feature extraction on multi-scale point cloud signals. It also includes a frequency domain feature analysis module, based on the MLP, which analyzes and aggregates the extracted multi-scale point cloud features. Point cloud understanding experiments are conducted on the ModelNet40, ScanObjectNN, and ShapeNet Part benchmarks. On ModelNet40, PFFA-MLP achieves an accuracy of 98% compared to PointMLP, while improving the detection speed by 7.2 times. On ScanObjectNN, the accuracy reaches 94% of PointMLP, while the detection speed is improved by 3.2 times. On ShapeNet Part, it achieves 97% of PointMLP. The experimental results demonstrate that PFFA-MLP achieves a favorable trade-off between training and inference speed. This method is applicable in scenarios that require efficient point cloud analysis while maintaining remarkable detection precision.

我们提出了一种基于多层感知器范式（MLP）的点云频域特征分析方法，称为PFFA-MLP。通过利用傅里叶变换和最新的 MLP 作为变压器机制的替代品，PFFA-MLP 使用三维 (3D) 离散傅里叶变换将点云映射到频域。 PFFA-MLP 采用轻量级堆叠 DGCNN 模块对多尺度点云信号进行初始特征提取。它还包括基于MLP的频域特征分析模块，对提取的多尺度点云特征进行分析和聚合。点云理解实验在 ModelNet40、ScanObjectNN 和 ShapeNet Part 基准上进行。在ModelNet40上，PFFA-MLP相比PointMLP实现了98%的准确率，同时检测速度提高了7.2倍。在ScanObjectNN上，准确率达到PointMLP的94%，同时检测速度提高了3.2倍。在ShapeNet部分，它达到了PointMLP的97%。实验结果表明，PFFA-MLP 在训练和推理速度之间实现了有利的权衡。该方法适用于需要高效点云分析同时保持较高检测精度的场景。