Depth map denoising using a bilateral filter and a progressive CNN

With the advantages of low cost and real-time acquisition of depth and color maps of objects, time-of-flight (ToF) cameras have been used in 3D reconstruction. However, due to the hardware shortage of ToF sensors, the depth maps obtained by ToF cameras have a lot of noise, which limits their subsequent application. Therefore, it is necessary to denoise the depth maps using a software method. We propose an algorithm for denoising depth maps by combining a bilateral filter with a progressive convolution neural network (PCNN). The algorithm takes a single depth map as input. First, the first individual network of the PCNN is used to denoise the depth map, and then the bilateral filter and the second individual network of the PCNN are used for further processing, so that the edge information of the depth maps can be retained on the basis of fine denoising. Finally, we have carried out experiments on the popular Middlebury dataset. The experimental results show that the proposed algorithm is obviously superior to traditional methods. (C) 2020 Optical Society of America