Depth Image Denoising via Collaborative Graph Fourier Transform

Abstract

Depth maps offer partial geometric information of a 3D scene, and thus have been widely used in various image processing and computer vision task. However, due to the limitation of capturing devices, depth images usually suffer from noises. How to remove noises containing in depth images become an important problem, which will benefit many practical applications. Depth image denoising is ill-posed, whose performance largely relies on the prior knowledge of depth images. In this paper, we propose a patch-wise depth image denoising algorithm which exploits intra-patch and inter-patch correlations represented by graph. Specifically, we first cluster similar patches in a depth map, and then stack them together into a 3D group. For each patch, a fully-connected 2D graph is built to model the intra-patch correlation among pixels. Furthermore, the employ the inter-patch correlation, we construct 1D fully connected graph, in which each node represents a patch in the collected group. Finally, a collaborative filtering based on 3D graph Fourier transform (GFT) is conducted on the 3D patch group. We conduct shrinkage of the transform coefficients to get the denoised patches. Experimental results demonstrate that our proposed approach outperforms state-of-the-art image denoising methods in both objective results and subjective visual quality.