IEEE Geoscience and Remote Sensing Magazine - June 2021 - 40

convolutional features through suitable (attention) weights
that emphasize effective features and suppress useless ones.
Likewise, spatial attention modules focus on image regions
that are more informative. Even though ablation studies
seem to support the importance of attention modules, in
general, it is not obvious how spatial attention, in particular,
helps to achieve a better image restoration.
The SAR-dynamic dilated convolution network proposed
in [63] is conceptually similar to the SAR-DRN, the
main innovation being the introduction of two five-layer
dense blocks. Dense connections [93] are known to facilitate
a better propagation of features in very deep networks
and hence reduce the vanishing gradient problem, a major
issue for the training of very deep CNNs. However, the network
used here is only 12 layers deep. Dense connections
are also employed in the multiconnection network with
wavelet features proposed in [65], now in the context of a
much deeper network consisting of 32 layers. Indeed, the
explicit goal of the method, based on the analysis of previous
literature, is to use a deeper architecture to extract more
expressive features. The computational load is reduced by
means of simplified dense connections: five-layer blocks
are used, as in [63], but only the final layer in the block
receives inputs from all preceding layers. Then, the same
structure is replicated at the block level, thus constructing a
hierarchical, multiconnection network. A further expedient
is to compute preliminarily a single-level wavelet transform
of the image to deal with a more compact four-subband input
and gain the freedom to use different losses for low and
high frequencies. With the very same aim of increasing the
network depth, a simple 24-layer architecture inspired by
the residual network is used in [64], with skip connections
at every other layer to prevent vanishing gradients.
A relatively deep network is proposed in [66], inspired
by the U-Net architecture [94] originally proposed for image
segmentation. This is an encoder-decoder network: in
the encoder, the image is repeatedly subsampled to extract
rich contextual features. The decoder then expands the features
back to the image size. Moreover, to preserve image
details, several skip connections link the two branches of
the " U " at the same sampling level so as to inject high-resolution
details into the output. The architectural choices are
well supported by ablation studies. It is worth underlining
that the network is trained on realistic SAR data obtained
by injecting simulated speckle into a deeply multilooked
SAR image. Although the speckle is simulated, the statistics
of the clean SAR signal are preserved. Skip connections
are also the key idea behind [67], where a 28-layer residual
encoder-decoder network [42] is used. Again, the goal is to
preserve image details and reduce the vanishing gradient
problem, but features are now added, as in a residual network,
rather than concatenated, as in the U-Net.
ORIGINAL DOMAIN NO-RESIDUAL ARCHITECTURES
Intelligent detection (ID) using a CNN [40] is one of the earliest
CNN-based image despeckling methods. The proposed
40
network has a residual architecture, but, unlike previous
methods, it aims to estimate the noise content from the
original domain image. Therefore, the denoised image is
obtained by taking the ratio, rather than the difference,
between the input image and the estimated speckle. This
approach makes full sense, considering the multiplicative
nature of noise. Of course, a pointwise ratio of images may
easily produce outliers in the presence of estimated noise
values close to zero. However, a tanh nonlinearity layer
placed right before the output performs soft thresholding,
thus avoiding serious shortcomings. Despite the good performance,
no other techniques followed this path, except for
some trivial variations [68], [69] of the ID-CNN. The network
itself is quite standard, with eight convolutional layers,
batch normalization, and rectified linear units.
In [70], the same authors propose the ID-GAN. Although
GANs are not plain CNNs, we consider this to be
a direct deep learning-based method because the actual
despeckling engine is just the generator subnetwork. The
generator takes the noisy input as a seed and generates a
new image that must appear virtually speckle-free to pass
the scrutiny of the discriminator. Eventually, the generator
is trained with a composite loss, which includes not
only the adversarial term but also the usual 2
, loss to ensure
fidelity to the original image, and the perceptual loss
proposed in [95], based on deep features extracted by an
independent pretrained Visual Geometry Group 16 [96]
model. Apart from this, the generator is quite standard: a
symmetric eight-layer CNN with an autoencoder structure.
As in many other cases, the model (trained on synthetic
data) is not available online, which is unfortunate given the
difficulty of GAN training. A similar method, with minor
architectural variations, is proposed in [71] with a TV loss
in place of the perceptual loss.
A multiplicative noise model is also adopted in [73]-
[75], all of which are by the same authors. In the first two
proposals, a 10-layer plain CNN is used, while in the multiobjective
network in [75], two-layer residual blocks are
also considered, bringing the network to a total of 17 layers.
Rather than architecture, the main focus is on the loss
function, which is crafted to capture the statistical peculiarities
of SAR images and speckle noise. We briefly mention
[76] and [77], where three-layer networks are used-a
plain CNN and a multilevel perceptron, respectively-with
results of limited interest.
Finally, [78] addresses the important problem of generalization.
A method trained on a given data set may perform
badly on mismatching target data, such as images acquired
by a different sensor or with a different number of
looks. Sometimes, retraining the network anew is not feasible,
so an unsupervised fine-tuning strategy is proposed.
Two 17-layer DnCNN-like networks are trained in parallel
to reproduce the clean image and the residual noise, respectively,
and specific loss terms are used for the two tasks.
Then, to adapt to new unseen data, the network is finetuned
using the very same noisy input as a proxy for the
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
