![](https://csdnimg.cn/release/download_crawler_static/16629539/bg4.jpg)
Sensors 2020, 20, 7268 4 of 18
mechanism with recurrent structure, and a feedback network based on recursive learning was proposed,
providing strong early reconstruction ability while requiring only a few parameters.
Figure 2 illustrates DRFFN, where the deeply recursive fusion module (DRFM) can be utilized
circularly as a recursive unit without adding new convolution parameters. Figure 3 shows that each
DRFM is composed of two cascaded branches which also employ the recursive structure, and the
low-frequency module (LFM) and high-frequency module (HFM) are recursive units of branches,
respectively. The recursive structure allows us to design very deep networks while prevent incurring
more parameters. Though a deeply recursive network involves fewer parameters, it cannot avoid
high computational costs, which will greatly increase the risk of gradient vanishing or exploding.
Referencing [
19
,
35
–
37
] the aforementioned GRL and LRL in DRFFN are introduced to address
the problem.
Sensors 2020, 20, x FOR PEER REVIEW 4 of 19
mechanism with recurrent structure, and a feedback network based on recursive learning was
proposed, providing strong early reconstruction ability while requiring only a few parameters.
Figure 2 illustrates DRFFN, where the deeply recursive fusion module (DRFM) can be utilized
circularly as a recursive unit without adding new convolution parameters. Figure 3 shows that each
DRFM is composed of two cascaded branches which also employ the recursive structure, and the
low-frequency module (LFM) and high-frequency module (HFM) are recursive units of branches,
respectively. The recursive structure allows us to design very deep networks while prevent incurring
more parameters. Though a deeply recursive network involves fewer parameters, it cannot avoid
high computational costs, which will greatly increase the risk of gradient vanishing or exploding.
Referencing [19,35–37] the aforementioned GRL and LRL in DRFFN are introduced to address the
problem.
Figure 2. The overall architecture of DRFFN. The red dotted rectangle represents the trunk part of the
network, which consists of the four parts of FE, recursive-in-recursive (RIR), UPAND RE, and the
detailed description is provided in Section 3.2. The deeply recursive frequency fusing module
(DRFFM) is the primary unit of the network.
LFM LFM LFM
HFM
HFM HFM
Figure 3. Interior structure of DRFFM, which contains two branches: the top one is comprised of low-
frequency models (LFMs), and the bottom one is composed of high-frequency modules (HFMs).
2.3. Attention Mechanism
Spatial locations and channels in the network contribute to super-resolution in varying degrees,
and not all features are equally important to super-resolution. In general, the attention mechanism
can guide how to reallocate available resources according to the informative features of the input.
This mechanism can be designed from the two dimensions of image control and image channel, based
on which the mechanism can be divided into the two categories of spatial attention (SA) and channel
attention (CA). In SelNet [38], a novel selection unit acts as a gate between convolutional layers, which
only allows selected values from the feature maps to pass. In RCAN [12], a channel attention
mechanism is employed in each local residual block by which the model focus on selective feature
maps that are more significant for the final tasks, so that the relationships between feature maps can
be effectively modeled. In SRRAM [15], an SR network is built based on a new attention method
fusing the two mechanisms (SA and CA) with the residual attention module (RAM), which is a basic
part of SRRAM consisting of residual blocks based on SA and CA.
The VCA proposed in this paper is for both low-frequency and high-frequency cascaded
branches, where the former pays more attention to channels with low variance, while the latter
focuses on channels with high variance. The experimental results show that the channel attention
Figure 2.
The overall architecture of DRFFN. The red dotted rectangle represents the trunk part of
the network, which consists of the four parts of FE, recursive-in-recursive (RIR), UPAND RE, and the
detailed description is provided in Section 3.2. The deeply recursive frequency fusing module (DRFFM)
is the primary unit of the network.
Sensors 2020, 20, x FOR PEER REVIEW 4 of 19
mechanism with recurrent structure, and a feedback network based on recursive learning was
proposed, providing strong early reconstruction ability while requiring only a few parameters.
Figure 2 illustrates DRFFN, where the deeply recursive fusion module (DRFM) can be utilized
circularly as a recursive unit without adding new convolution parameters. Figure 3 shows that each
DRFM is composed of two cascaded branches which also employ the recursive structure, and the
low-frequency module (LFM) and high-frequency module (HFM) are recursive units of branches,
respectively. The recursive structure allows us to design very deep networks while prevent incurring
more parameters. Though a deeply recursive network involves fewer parameters, it cannot avoid
high computational costs, which will greatly increase the risk of gradient vanishing or exploding.
Referencing [19,35–37] the aforementioned GRL and LRL in DRFFN are introduced to address the
problem.
Figure 2. The overall architecture of DRFFN. The red dotted rectangle represents the trunk part of the
network, which consists of the four parts of FE, recursive-in-recursive (RIR), UPAND RE, and the
detailed description is provided in Section 3.2. The deeply recursive frequency fusing module
(DRFFM) is the primary unit of the network.
LFM LFM LFM
HFM
HFM HFM
Figure 3. Interior structure of DRFFM, which contains two branches: the top one is comprised of low-
frequency models (LFMs), and the bottom one is composed of high-frequency modules (HFMs).
2.3. Attention Mechanism
Spatial locations and channels in the network contribute to super-resolution in varying degrees,
and not all features are equally important to super-resolution. In general, the attention mechanism
can guide how to reallocate available resources according to the informative features of the input.
This mechanism can be designed from the two dimensions of image control and image channel, based
on which the mechanism can be divided into the two categories of spatial attention (SA) and channel
attention (CA). In SelNet [38], a novel selection unit acts as a gate between convolutional layers, which
only allows selected values from the feature maps to pass. In RCAN [12], a channel attention
mechanism is employed in each local residual block by which the model focus on selective feature
maps that are more significant for the final tasks, so that the relationships between feature maps can
be effectively modeled. In SRRAM [15], an SR network is built based on a new attention method
fusing the two mechanisms (SA and CA) with the residual attention module (RAM), which is a basic
part of SRRAM consisting of residual blocks based on SA and CA.
The VCA proposed in this paper is for both low-frequency and high-frequency cascaded
branches, where the former pays more attention to channels with low variance, while the latter
focuses on channels with high variance. The experimental results show that the channel attention
Figure 3.
Interior structure of DRFFM, which contains two branches: the top one is comprised of
low-frequency models (LFMs), and the bottom one is composed of high-frequency modules (HFMs).
2.3. Attention Mechanism
Spatial locations and channels in the network contribute to super-resolution in varying degrees,
and not all features are equally important to super-resolution. In general, the attention mechanism
can guide how to reallocate available resources according to the informative features of the input.
This mechanism can be designed from the two dimensions of image control and image channel,
based on which the mechanism can be divided into the two categories of spatial attention (SA) and
channel attention (CA). In SelNet [
38
], a novel selection unit acts as a gate between convolutional
layers, which only allows selected values from the feature maps to pass. In RCAN [
12
], a channel
attention mechanism is employed in each local residual block by which the model focus on selective
feature maps that are more significant for the final tasks, so that the relationships between feature
maps can be effectively modeled. In SRRAM [
15
], an SR network is built based on a new attention
method fusing the two mechanisms (SA and CA) with the residual attention module (RAM), which is
a basic part of SRRAM consisting of residual blocks based on SA and CA.
The VCA proposed in this paper is for both low-frequency and high-frequency cascaded
branches, where the former pays more attention to channels with low variance, while the latter
focuses on channels with high variance. The experimental results show that the channel attention