Denoising of Photographic Images and Video by Marcelo Bertalmío

Author:Marcelo Bertalmío
Language: eng
Format: epub
ISBN: 9783319960296
Publisher: Springer International Publishing

However, the RMSE of all the patches is considerably higher for DnCNN than for the Oracle (2 dB). Figure 6.14 further illustrates the differences in RMSE of the DnCNN and the Oracle as a function of (for patches extracted from the denoised images) for various noise levels ( from bottom to top). As can be seen for internal Oracle error is lower, while for DnCNN error is lower. The differences are statistically significant (verified by Wilcoxon rank sum test using Matlab’s “ranksum”). Notably, this threshold does not fit the expected threshold from Fig. 6.6b. This happens because in Sect. 6.2.2 we discussed a linear mapping from noisy to clean patch, while DnCNN minimizes a nonlinear mapping between the noisy image (or patch) and its clean counterpart. Hence, the threshold might differ.

While the Oracle denoising is not an algorithm, and is linear, it is a good indication of existing information within a certain receptive field. Based on this, we conclude that for patches with relatively low PatchSNR, the learned mapping of DnCNN [31] does not manage to predict their corresponding best “clean” patch representative that resides in the receptive field of the network. This behavior is consistent for a wide range of noise levels (). However, for higher noise levels, there are not enough patches above the threshold to form substantial statistics, hence those plots are not presented in Fig. 6.14.Table 6.1Comparison of PSNR (dB) on BSD100 [5]

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.