Feature-Guided Reversible Network for Point Cloud Denoising

Point clouds inevitably introduce noise during the acquisition process, which severely impacts subsequent applications and downstream tasks. It is crucial for efficient and accurate point cloud denoising. However, existing methods struggle to simultaneously maintain strong expressive power and high fidelity to geometric details under complex noise interference. This paper proposes a feature-guided reversible neural network method for point cloud denoising. This method significantly improves the performance of point cloud denoising by extracting effective point cloud feature representations. First, we design a feature extraction and sharpening module capable of progressively accumulating features and capturing rich geometric information. Specifically, we apply a global bilinear sharpening method to aggregate global descriptors at both the channel and point levels, and simultaneously propose a bilinear interaction mechanism to achieve feature sharpening and noise suppression. Second, we present a reversible neural network designed to perform noise decoupling processing, further enhancing the discrimination of features. Moreover, we use an adaptive convolution module to improve the dimensionality of the input, thereby avoiding the low-dimensional bottleneck that cannot provide practical, rich geometric structure information. Finally, we separate noisy points from clean points in the constructed prior latent space through the inverse transformation of the reversible network, thereby obtaining a noise-free point cloud. We evaluate the denoising ability of our method on benchmark synthesized and real-world datasets, and the experimental results show that our method achieves a state-of-the-art performance in point cloud denoising tasks, surpassing previous point cloud denoising methods.