MD3: Model-Driven Deep Remotely Sensed Image Denoising
, Zifan Zhu 2, Yaozong Zhang 2, Zhicheng Wang 2, Biyun Xu 2, Jun Liu 3, Shaoyi Li 4 and Hao Fang 5,*
Round 1
Reviewer 1 Report
This paper proposed a model-driven deep denoising (MD3) scheme for noise removal of remotely sensed images. The contents are arranged as follows: Firstly, the authors claimed the background, problems and contributions in the introduction part. Then, they introduced the related works to solve the problems and presented their drawbacks. Next, they importantly introduced their work, including MD3 model construction, MD3 model decomposition, and their respective solution. Finally, they employed quantitative and qualitative comparisons of the state-or-the-arts to verify the effectiveness of the proposed method. The above organized contents are easily followed by readers. My issues are listed as follows:
(1) Line 17, some words seem to be missing.
(2) Line 198, the parameter lambda is not defined.
(3) Line 486, the journal should be correctly written. Such minor revisions can further improve the quality of the paper.
Author Response
Thank you very much for your comments on our work. The responses are listed in the attachment file, please download and check it. Thank you very much again.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Image denoising has been a long-standing hot topic research in computer vision. The authors claimed that they had proposed a model-driven deep denoising (MD3) scheme to remove noise. That is, the sparsity of the neighborhood similar patches are employed to build a MD3 noise removal model. Then, it is separated into the two sub-problems by the alternating direction method of multipliers (ADMM). Next, the image component is obtained by solving a quadratic optimization problem which has a fast closed-form solution while the other sub-problem is further decomposed by the ADMM and its two parts are iteratively solved by deep learning strategies, such as one is a learnable ISTA (LISTA) encoder network, the other is a learned deep Gaussian denoiser (DGD), to speed up the model convergence to a fixed solution. The structures of this paper are arranged and readers can be easily followed. The results are faithful and can verify the effectiveness of the proposed MD3 method. My comments are as follows:
1. There are some syntax errors in the main text, for instance, in Abstract,”limits” should be “limit”. The authors should check the whole text.
2. There is a word “Remote” missing at the front of the first sentence in Introduction.
3. Remote sensing images degraded by additive white Gaussian noise (AWGN) produced by photo or electronic should be cited.
4. The spaces at the front of the sentence “where E(.) is a function to compute the mean value of a matrix” under Eq. (5) are unnecessary.
5. Line 198,“beta is a penalty parameter” should be “lambda is a penalty parameter”.
6. The tense of the sentence in the conclusion part should be seriously checked.
Author Response
Thank you very much for your comments on our work. The responses are listed in the attachment file, please download and check it. Thank you very much again.
Author Response File: Author Response.pdf
