Sorry, you need to enable JavaScript to visit this website.

Video denoising for the hierarchical coding structure in video coding

Primary tabs

Citation Author(s):
Jingning Han, Yaowu Xu
Submitted by:
Cheng Chen
Last updated:
25 March 2020 - 12:36am
Document Type:
Document Year:
Presenters Name:
Cheng Chen



In modern video codecs, video frames are coded out-of-order following a hierarchical coding structure. The naive uniform video denoising, where denoising is applied indifferently to each frame, does not improve the compression performance. In our work, we only apply denoising to frames at layer 0 and 1. The denoising leads to a significant reduction of bit rates while maintaining temporal correlation. PSNR scores of the filtered frames decrease but PSNR scores of the unfiltered frames remain or even improve. As an example, a 150 frame video is encoded and the PSNR values of the first GOP is displayed in the figure. The PSNR value of the filtered key frame has dropped about 2dB as compared with the unfiltered one, while PSNR values of following frames in this GOP increase. The reason is that the denoising of the key frame reduces the noise and improves its quality in terms of temporal correlation and thus serves as a better reference for the following frames. Meanwhile the reduction of bit rate is achieved on filtered frames. Therefore the overall compression performance is significantly improved.

In this work, a non-local mean algorithm is used for denoising. We divide the current frame to M × M (M = 32) blocks. A total of N (N = 7) candidate patches are found through motion search and combined to generate the filtered block. The proposed method is implemented in AV1 (libaom). We evaluated the compression performance on various video benchmark test sets, which cover a wide range of videos of different types: slow and fast motion, zooming, rotation, screen content and etc. The experiment is evaluated at speed 1 (–cpu-used=1 ), on the constant quality mode (–end-usage=q), with the maximum frame number of 150. The compression efficiency improvements in PSNR, SSIM and VMAF are significant across different test sets.

0 users have voted:

Dataset Files