A Small-Scale Image U-Net-based Color Quality Enhancement for Dense Point Cloud

Efficient compression for 3D point clouds is crucial due to their massive data volume. Quality enhancement can significantly improve the compression efficiency of 3D point clouds. In this study, we propose a neural network-based quality enhancement method for color attributes of dense 3D point cloud. Our approach involves the extraction of 3D patches from an input distorted point cloud, followed by their conversion into 2D images using a specific scan order. We then propose an efficient U-Net-inspired neural network, namely SSIU-Net, to enhance the quality of these 2D images. Finally, the processed 2D images are converted back to 3D patches, allowing for the point cloud reconstruction. Experimental results demonstrate that the proposed method, when implemented in both G-PCC and V-PCC, achieves competitive results. For example, 0.146 dB, 0.417dB and 0.270 dB PSNR gains can be achieved for Luma (Y), Chroma difference of blue (Cb) and Chroma difference of red (Cr) components compared with G-PCC, corresponding to 4.1%, 9.4% and 12.5% BD-rate savings. The source code is available at https://github.com/xjr998/SSIU.