Depth enhancement via non-local means filter
Ke Yang, Yong Dou, Xiaoyang Chen, Shaohe Lv, Peng Qiao · 2015
The depth image captured by a RGB-D camera is noisy and usually misses the values at some pixels, especially around the object boundaries. There are many methods take advantage of the corresponding color images to enhance the depth images. Most of them bring the texture of the color image into the depth image. In this paper, an adaptive double non-local means (ADNLM) method of depth enhancement is proposed. First, ADNLM pre-inpaint the depth image via a color-guided non-local means method; second, a depth-based non-local means method is used to denoise the pre-inpainted depth image. Experiments on the public benchmarks show that, ADNLM can avoid the color texture, and effectively enhance the depth image, especially when there are several large missing regions in the depth image. Also, the performance in terms of PSNR of ADNLM is slightly better than that of the state of the arts.