IEEE Geoscience and Remote Sensing Magazine - June 2021 - 38

basic assumptions on the noise modeling. We describe the
approaches in the following:
1) Log domain methods: By taking the log of the input image,
the noise is converted from multiplicative to additive,
irrespective of whether a residual architecture is used
or not. Moreover, the variance of the noise is stabilized,
which fully justifies the adoption of fixed denoisers. This
enables one to tap into the huge reservoir of ideas and
methods developed for the restoration of AWGN images.
However, one should be aware that the log domain noise
is neither white nor Gaussian but weakly correlated
Fisher-Tippett (see the " Primer on SAR Speckle " section).
TABLE 2. DESPECKLING METHODS BASED ON SUPERVISED DEEP LEARNING AND THEIR MAIN FEATURES.
REFERENCE ABBREVIATION KEYWORDS
USE OF DEEP
LEARNING
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[40]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
SAR-CNN
residual@log
MuLoG-CNN MuLoG with DnCNN
MuLoG with
FFDNet
DNN
SAR-DRN
HDRANet
SAR-DDCN
N/A
MCN-WF
N/A
REDNET
ID-CNN
N/A
FID-CNN
ID-GAN
N/A
STD-CNN
N/A
KL-DNN
MONet
MLP
N/A
deSpeckNet
CNN-NLM
CNN-NLM
N/A
SAR-NN3D
SAR-RDCP
SPB-Net
DABM
N/A
N/A
N/A
MSR-Net
MuLoG with FFDNet
Signal KLD loss
Dilated convolutions
Attention modules
Dense + dilated
ResNet/skip
Blockwise dense
U-Net
REDNET
Residual with ratio
ELU nonlinearity
Fractional TV
GAN + perceptual loss
GAN + TV loss
Attention modules
Speckle KLD loss
Speckle KLD loss
Speckle KLD loss
MLP
Residual with ratio
Direct
Direct
MuLoG-CNN Pretrained versus training Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Direct
Unsupervised fine-tuning Direct
Nonlocal + deep learning Blended
Nonlocal + deep learning,
N3 layers
MuLoG + nonlocal
Blended
Blended
Nonlocal + deep learning Blended
MAP with CNN prior
Blended
Operated by unfolded CNN Blended
MAP with CNN prior
Guided fusion
Texture map
Shearlet
Multiresolution
Blended
Blended
Blended
Blended
Blended
NOISE
MODEL
Additive
Additive
Additive
Additive
Additive
add-sig-dep University of California, Merced
add-sig-dep University of California, Merced
add-sig-dep University of California, Merced
add-sig-dep Northwestern Polytechnical University
add-sig-dep University of California, Merced
add-sig-dep N-look Sentinel
add-sig-dep BSDS500
Multiplicative UCID, BSDS500, Google Maps
Multiplicative
Multiplicative
Multiplicative UCID, BSDS500, Google Maps
Multiplicative BSD, University of California, Merced
Multiplicative BSDS500
Multiplicative UCID, BSDS500, Google Maps
Multiplicative UCID, BSDS500, Google Maps
Multiplicative University of California, Merced
Multiplicative 52-look TerraSAR-X
Multiplicative
Multiplicative N-look Sentinel
Multiplicative 25-look COSMO
Multiplicative 25-look COSMO
Additive
N/A
Unspecified Pretrained
add-sig-dep University of California, Merced
add-sig-dep Radarsat-1 sparse
add-sig-dep Fully synthetic
Additive
BSDS500
Multiplicative Northwestern Polytechnical University
Additive
Multiplicative University of California, Merced
MuLoG: multilevel logarithm with Gaussian denoising; FFDNet: fast and flexible discriminative CNN denoiser; DNN: deep neural network; KLD: Kullback-Leibler divergence; DRN: dilated
residual network; HDRANet: hybrid dilated residual attention network; DDCN: dynamic dilated convolution network; ResNet: residual network; MCN-WF: multiconnection network with
wavelet features; REDNET: residual encoder-decoder network; UCID: Uncompressed Color Image Database; ID: intelligent detection; ELU: exponential linear unit; MONet: multiobjective
network; MLP: multilayer perceptron; NLM: nonlocal means; NN3D: nearest-neighbor 3D point-to-point pattern; RDCP: recursive deep CNN prior; MSR: multiscale retinex.
http://github.com/
adugnag/deSpeckNet
http://github.com/
davin11/CNN-NLM
http://github.com/
davin11/CNN-NLM
http://github.com/
XwK-P/ID-CNN
http://github.com/
qzhang95/SAR-DRN
TRAINING SET
25-look COSMO
N-look Sentinel
N-look Sentinel
CODE
http://grip-unina.
github.io/SAR-CNN/
http://gitlab.telecomparis.fr/ring/SAR-CNN
38
IEEE
GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2021
IEEE Geoscience and Remote Sensing Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2021
