Color image guided depth image reconstruction based on a total variation network

Abstract

A representative method to improve a depth image is to use an aligned high-quality color image to guide the depth image by migrating the color details to the depth image. In the process of color-guided depth reconstruction, there often is a misalignment of the edge of the color image used to guide the depth image reconstruction and the depth discontinuity of the depth image. This makes the results suffer from texture copy artifacts and blurring depth discontinuities. In this paper, we use a total variation deep network founded on deep learning and high-resolution color images. The experimental result indicates that under the guidance of high-resolution colors, the depth image recovered is closest to the ground truth in the edge contour, the PSNR and FSIM index are suboptimal for 64×, and the contour and position information recovered from the reconstructed depth image can be retained in the very low-resolution depth image.