融合空间域特征与频域解耦的多粒度轻量级点云分类

Key Points

• This research aims to address deficiencies in existing point cloud classification networks by developing a novel classification method that integrates spatial and frequency domain features.
• Utilized a multi-dimensional multi-branch feature fusion architecture.
• Employed improved edge convolution operators and an increasing neighbor number sequence for multi-scale geometric feature aggregation.
• Incorporated a channel attention mechanism for adaptive enhancement of salient features.
• Achieved overall accuracy of 93.0% on the ModelNet40 dataset, improving 1.0% over existing methods.
• Achieved overall accuracy of 82.4% on the ScanobjectNN dataset, improving 3.4% over existing methods.
• Reached instance average intersection over union of 85.94% on the ShapeNet Part dataset, improving 0.8% over existing methods.