| CNN-based | ［18］ | Fully differentiable CNN | Height map denoising network | Poor denoising effect on larger holes |
| ［19］ | GCN | Robust to high levels of noise | Neighborhood size can affect performance |
| ［20］ | Geometric dual domain graph convolutional networks | Real and virtual normals are defined | Longer training time |
| ［21］ | Feature preserving normal estimation | Automatically estimate normals and update point locations | Unsuitable for severe noise and large outliers |
Upsampling- based | ［25］ | Denoiser and upsampler combined | Effectively resist attacks from other point cloud datasets | Unsuitable for defending against black box attacks |
| ［27］ | Networks based on discrete differential geometry | Preserve features and geometric details | Incomplete datasets are not considered |
| ［29］ | Patch correlation unit and position correction unit | Consider noise and outliers in practical applications | The patch selection strategy will affect the stability of the algorithm |
| ［30］ | Graph attention convolution and edge-aware node caching | Fine-grained edge detail is preserved with high quality | GAC modules increase computational complexity |
| Filter-based | ［31］ | Edge-aware integrated network | Suitable for dense point clouds with structure-invariant scale | Training time is long |
| ［32］ | Projection denoising method based on neural network | Direct point cloud denoising using deep learning techniques | Need enough training samples |
| ［37］ | Add repulsion term and data term to the objective function | Capable of handling fine-scale features and sharp features | Depend on the quality of the input normals |
| ［38］ | Outlier recognizer and denoiser | Identify and remove points that are far from the surface | Runtime can also be optimized |
| Gradient-based | ［39］ | Score estimation network | More robust to outliers | The gradient is discontinuous |
| ［41］ | Momentum gradient ascent | The gradient field is continuous | Need to construct an effective global gradient field |
| ［42］ | GPCD++ network framework | Lightweight network UniNet | Cannot handle large pores |
| Other methods | ［43］ | Channel attention module | Stitching local features of point clouds at multiple scales | The capture of neighborhood feature information is biased |
| ［44］ | Hybrid self-attention network | Enhance local information through Transformer | Longer training time |
| ［48］ | Unsupervised machine learning | Detect outliers by isolation forests and elliptical envelopes | High time complexity |
| ［49］ | Transformer-based | Extract multi-scale local features | High computational complexity |
![PointCleanNet framework[35]](/richHtml/zgjg/2024/51/5/0509001/img_01.jpg)
