IEEE Geoscience and Remote Sensing Magazine - June 2021 - 41

clean image on one branch and, again, the residual noise
on the other branch. Experiments confirm much improved
results for off-training data.
MODEL-BASED DESPECKLING METHODS
EXPLOITING DEEP LEARNING TOOLS
Methods in this family try to blend the model-based and
data-driven approaches, with the aim to exploit the large
body of knowledge and procedures accumulated for SAR
despeckling through several decades and the great potential
of deep learning tools. A perfect example of this blending is
represented by the CNN-nonlocal means (NLM) [79], [80],
a simple and well-understood linear-filtering algorithm. The
clean target pixel is estimated as a weighted average of neighboring
noisy pixels, with weights that depend on the similarity
between the target and the estimator. In the CNN-NLM,
the similarity metric is replaced by a suitably trained CNN.
The network takes as input a patch extracted from the original
domain image and outputs a set of filter weights adapted
to the local image content. In [79], a rather standard CNN is
used with 12 convolutional layers, while in [80], a 20-layer
CNN is proposed that includes two N3
layers introduced in
[25] to exploit image self-similarities. These layers associate
the set of its K-nearest neighbors with each input feature,
which can be exploited for subsequent nonlocal processing
steps. The training is conducted on synthetic data and real
multilooked SAR images, as in [56]. The results are much
better than those of conventional nonlocal methods, such as
probabilistic patch-based (PPB) filter approaches [14], which
provides some hints about how the filter weights should be
chosen, given the underlying signal and the noise strength.
Moreover, the performance matches that of state-of-the-art
CNN-based methods, which is quite interesting considering
that the filtering engine is fully linear. The fact that, despite
the nonadditive nature of the noise, a linear-filtering method
can be competitive with highly nonlinear deep networks
may deserve further studies.
Note that the interplay between nonlocal filtering
and deep learning is the object of intense research for
AWGN denoising, e.g., [25], [27], and [98]-[100]. A first
exploratory work for SAR despeckling, inspired by [100],
is carried out in [81], where the output image provided
by the MuLoG approach with the DnCNN is eventually
subject to nonlocal refinement. Similar patches, selected
based on their likeness in the original domain SAR image,
are collected in 3D groups, subject to transform domain
shrinkage (the Haar wavelet), reprojected in the image
domain, and aggregated. Along the same lines, the SAR-
nearest-neighbor 3D point-to-point pattern [82] combines
a pretrained CNN-based despeckler (the CNN-SAR or
ID-CNN) and nonlocal 3D shrinkage in the context of
an iterative filtering procedure. At the kth iteration, the
current denoised output is combined with the input and
filtered again by the CNN; then, similar blocks undergo
nonlocal 3D shrinkage. Unfortunately, no test on real SAR
images is carried out.
JUNE 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
Another iterative method is proposed in the SAR-recursive
deep CNN prior [83], where filtering is cast as a maximum
a posteriori (MAP) problem solved by means of halfquadratic
splitting. The name of the method comes after the
prior term, estimated by means of a CNN that, to limit network
complexity, uses the same parameters in all iterations
and is therefore subject to a recursive form of training. The
network itself is relatively small, employing dilated convolutions,
residual blocks, and channel attention modules. The
fidelity term is optimized by simple gradient descent. A similar
high-level approach is followed in [84], aimed at joint SAR
imaging and despeckling, with an emphasis on the case of
corrupted raw data. A constrained optimization problem is
formulated, with restrictions on fidelity, sparsity, and detail
preservation. Then, the split Bregman method is unfolded
into a neural network, called the SPB-Net, which is trained to
solve the problem effectively and efficiently. In [85], a CNN
is used only to estimate the parameters of an autobinomial
image prior in the context of an MAP despeckling method.
Experiments show that the proposed solution provides good
despeckling results with high computational efficiency.
In [86], multiple denoisers, which are structurally similar
to the SAR-DRN, are trained in the log domain for various
levels of noise. Their outputs are then fused using a saliency
map computed on image details as an external guide. The
process is then iterated by filtering the resulting denoised
image until a convergence criterion is met. Extraordinarily
good results are claimed, with 5-10-dB gains in the peak
SNR (PSNR) compared to the state of the art, but the code
is not available for replicating experiments. A guide is used
also in [87], where a small U-Net is adopted that takes as
input not only the original image but also a texture level
map measuring the local homogeneity index, as suggested
in [101]. The idea is that the texture level should support the
denoising by providing a more accurate estimation of the
local equivalent number of looks (ENL).
Finally, in [88] and [89], multiresolution processing is
considered. In the multiscale retinex network [89], three dyadic
resolution levels are considered, and a multiresolution
denoiser with a CNN engine is used at each. The denoiser
has a simple encoder-decoder architecture, but the parameters
of the bottleneck layer flow to the upper resolution
levels by means of a long short-term memory network to
enable fast convergence. In [88], the input image is subject
to a nonsubsampled shearlet transform with two levels of
decomposition. High-frequency coefficients are then denoised
by means of a transform domain method, while
only low-frequency coefficients are denoised using an FFDNet
in the log domain.
SUPERVISED MODELS: TRAINING AND TESTING
Training is at the core of supervised deep learning methods.
A network learns to perform its task based on examples of
the input paired with the desired output. Such examples
should be large in number to ensure good generalization
and, needless to say, to be meaningful to the problem.
41
IEEE Geoscience and Remote Sensing Magazine - June 2021

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