DNF-Net: A Deep Normal Filtering Network for Mesh Denoising.

Area of Science:

• Computer Graphics
• Geometric Processing
• Machine Learning

Background:

• 3D mesh data often suffers from noise, degrading geometric accuracy and visual quality.
• Existing mesh denoising methods may require manual feature extraction or struggle with complex noise patterns.

Purpose of the Study:

• To develop an automated and effective mesh denoising method using deep learning.
• To preserve geometric features during the denoising process for high-fidelity reconstruction.

Main Methods:

• A deep normal filtering network (DNF-Net) processes local mesh patches.
• The network utilizes a novel multi-scale feature embedding unit and residual learning.
• A deeply-supervised joint loss function is employed for enhanced training.

Main Results:

• DNF-Net directly outputs denoised facet normals from noisy inputs.
• The method achieves superior denoising performance compared to state-of-the-art techniques.
• Effective feature preservation is demonstrated in reconstructed geometries.

Conclusions:

• DNF-Net offers an end-to-end solution for mesh denoising without manual feature engineering.
• The proposed network demonstrates significant improvements on both synthetic and real-world noisy meshes.
• This approach advances the state of the art in geometric data processing and noise reduction.