2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

Comparative Study of Implementation of Very

Deep Super Resolution Neural Network and
2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC) | 979-8-3503-5874-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/AESPC59761.2023.10389860

Bicubic Interpolation for Single Image Super

Resolution QualityEnhancement
Ganesan P1*, M.Ravichandran2, B.S.Sathish3, L.M.I.Leo Joseph4, G.Sajiv5, R.Murugesan6
1*
Dept. of Electronics and Communication Engineering, Vidya Jyothi Institute of Technology, Hyderabad, India.
[email protected]. ORCID: https://orcid.org/0000-0001-6179-632X
2
Dept. of Computer Science Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and
Technology, Chennai, India [email protected]
3
Dept. of Computer Science Engineering AI, Sri Venkateshwara College of Engineering, Bengaluru, India.
[email protected]
4
Dept. of Electronics and Communication Engineering, SR University, Warangal, Telangana, India.
[email protected]
5
Saveetha School of Engineering,SIMATS, Chennai, India.
[email protected]
6
Dept. of Electronics and Communication Engineering, Narshima Reddy Engineering College, Secunderabad, India.
[email protected]

Abstract—To improve the resolution and quality of a undergone a paradigm shift as a result of the progress of
low level image, single image super resolution (SISR) deep learning, particularly CNNs.By learning the intricate
is a challenging endeavor in the turf of computer mappings between low-resolution and high-resolution
vision. It is essential to many applications, including image pairs, deep learning models have demonstrated
image editing, security systems, medical imaging, and outstanding performance in SISR [5][6][7]. Convolutional
others. In order to increase the spatial quality of a low neural networks, such as the widely used SRCNN (Super-
level image, SISR attempts to approximate the high Resolution Convolutional Neural Network), are one
frequency details that are missing from the image. The common architecture. In order to extract and learn visual
field was formerly based on manually created features features, SRCNN employs numerous layers of
and interpolation techniques, but current convolutions. A reconstruction phase is then used to
developments in deep learning have completely generate the high-resolution outcome. This
changed it. The suggested work compares the groundbreaking work laid the stage for later
effectiveness of bicubic interpolation and neural developments.Researchers employ quantitative
network based very deep super resolution for single measurements like the SSIM and PSNR to assess the
image super resolution image quality improvement. efficacy of SISR algorithms[8]. By contrasting the created
Utilizing blind and complete reference picture quality high-resolution to the original high-resolution images,
measures, the effectiveness of the two approaches is these metrics assess the quality of the images. Deep
evaluated. The effectiveness of the very deep super learning-based SISR techniques have outperformed
resolution for single image is clearly demonstrated by conventional methods by a wide margin in recent years,
the experimental results. producing impressive results. These techniques enable the
creation of high-resolution, aesthetically acceptable
Keywords—single image super resolution. image quality images from low-resolution inputs. The handling of high
assessment. bicubic interpolation. very deep super upscaling factors, the elimination of artifacts, and
resolution increasing the computational effectiveness of the methods
are still problems.
I. INTRODUCTION
In image processing and computer graphics, the
The development of deep learning algorithms is largely bicubic interpolation algorithm is frequently used to resize
responsible for the recent breakthroughs in single image or resample images [11]. Compared to more
super-resolution. Because low-resolution images lack straightforward techniques like nearest-neighbor or
high-frequency information, single image super-resolution bilinear interpolation, it offers a higher-quality image
is a difficult challenge [1][2]. From the provided low- interpolation. A smoother and more aesthetically
resolution input, the intention is to predict a reasonable acceptable image is produced via bicubic interpolation,
high-resolution image. The handcrafted features and which attempts to estimate pixel values based on nearby
optimization algorithms used in traditional methods known pixel values. A deep learning model created
frequently produced meager results [3][4]. SISR has
979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

specifically for picture super-resolution tasks is the Very- synthesis to resolution

Deep Super-Resolution (VDSR) neural network. The locate  neighbor
technique of improving the resolution and quality of low- comparable embedding
resolution photographs is referred to as super- patches in high- .
resolution[12]. resolution
The main principle of VDSR is to train a non- image.
linear mapping between low and high-resolution pixels Deep learning-  Outstanding  SRCNN,
using a deep CNN. By efficiently utilizing the power of based results when VDSR,
deep learning, VDSR seeks to generate high-resolution learning intricate and EDSR
image which is visually sharper and more detailed than mappings  Generative
conventional interpolation-based approaches. The quality between low- adversarial
of a single low-resolution image is improved by a process and high- networks
called single image super-resolution (SISR) [13][18]. To resolution (GANs)
accomplish this goal, numerous strategies and methods pictures
have been created over time. Some of the most common Hybrid  Hybrid
strategies and procedures utilized in SISR are illustrated methods approaches
in Table I. utilize several
methodologies
TABLE I. TECHNIQUES AND METHODS USED IN SISR such as Interpolation +
Technique / Salient Features Example interpolation for DL
Method upscaling and
Interpolation  Computationally  bicubic fine-tuning with
based efficient, interpolati a deep learning-
provide on based network
ambiguous  bilinear or
results and miss interpolati reconstruction-
intricate visual on based
details.  Nearest optimization.
Neighbor Self-supervised  Train models for
Interpolati super-resolution
on using the -
Edge-based  By identifying  Canny intrinsic
and boosting edge structure and
edge detector information
information,  bilateral found in the
reconstruct high- filtering low-resolution
resolution edges images
and textures. themselves.
 Apply edge-
preserving filters
to increase the II. MATERIAL AND METHODS
final image's In order to estimate pixel values between known
sharpness. data points, the bicubic interpolation technique is
Reconstruction-  Represent SISR  Sparse frequently employed in computer vision and computer
based as an coding graphics. It is a development of bilinear interpolation and
optimization  Non-local offers a more precise illustration of the underlying
problem, with means continuous function. Bilinear or bicubic interpolation
the objective of  Total techniques are a novel way to enhance resolution devoid
locating the variation of using machine learning. The value of the interpolated
highest- regularizat pixel is determined using a weighted average of sixteen
resolution image ion nearby pixels in this technique. A two-dimensional cubic
that best polynomial is frequently employed as the bicubic function
matches the in the interpolation process. It can be stated this way:
input's low f(x, y) = ∑[i=0 to 3] ∑[j=0 to 3] a_ij * x^i * y^j
resolution. where x and y are the coefficients that determine the
Example-based  Utilize methods  sparse weights given to each pixel, while x and y are the
such as patch coding- distances from the estimated position to the surrounding
matching and based pixels. Here is an explanation of the bicubic interpolation
patch-based super- algorithm.
Step 1: Given an input image and a desired output size.
979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

Step 2: Using scaling factors, determine the location of

each pixel in the input image.
Step 3: Locate the sixteen pixels that surround the
computed place in the input picture (four rows by four
columns).
Step 4: The weights for each of the sixteen pixels should
be generated depending on how far away they are from
the estimated position.
Step 5: Employ the weighting factors to the values of the
pixels and then employ a bicubic function to get the Fig 1.Architecture of VDSR neural network
interpolated value.
Step 6: Put the computed value in the output image's In the suggested work, the RGB original picture is
equivalent pixel. transformed to a YCbCr color model, and the channel ‘Y
Step 7: Continue with steps 2 through 6 until the whole ‘is utilized to create the residual image. The comparison
output image has been interpolated. of RGB and YCbCr color spaces is shown in Table II.
An image super-resolution task-specific deep
learning model is the Very-Deep Super-Resolution TABLE II. COMPARISON OF RGB AND YCBCR SPACE
(VDSR) CNN architecture. The technique of improving Aspect RGB YCbCr
the resolution and superiority of low-resolution Components R, G and B Luminance (Y),
photographs is referred to as super-resolution. When just Chrominance (Cb,
one low-resolution picture is available for improvement, Cr)
SISR presents a difficulty that VDSR is particularly Representation Each channel Luma (Y)
designed to address. represents represents
The main principle of VDSR is to train a non- intensity levels intensity, Chroma
linear mapping between low- and high-resolution pixels (Cb, Cr) represent
using a deep CNN. By efficiently utilizing the power of color difference
deep learning, VDSR seeks to produce high-resolution Channel 0 to 255 (8-bit) or 16 to 235 (Y), 16
pictures that are visually crisper and more detailed than Range 0 to 1 to 240 (Cb, Cr)
conventional interpolation-based approaches. Multiple Usage Commonly used Commonly used
convolutional layers with a sizable number of trainable for display and for video encoding
parameters make up the VDSR architecture. A low- rendering and transmission
resolution input picture is sent into the network, which Human Matches the Takes into account
then processes it through numerous convolutional layers Perception perception of differences in
and rectified linear unit (ReLU) activation functions. The human vision human visual
network can record intricate picture patterns because of perception
the non-linearity that is introduced by ReLU activation. Storage Requires more Requires less
The depth of VDSR is its main innovation. The network Requirements storage space for storage space for
can learn incredibly complex features and take advantage color color
of the hierarchical structure of visual representations by representation representation
stacking a lot of convolutional layers—typically more Application Graphics, Video encoding,
than 20 layers. In order to produce high-frequency details computer vision, television
during the super-resolution phase, VDSR needs to be able digital imaging broadcasting
to collect both local and global context, which is made
possible by the deep architecture [19].
The suggested method solely used the Y channel to train
VDSR is taught in the training phase utilizing
the VDSR. Consider the brightness of pictures with high
pairings of low and high-resolution images in a supervised
and low resolutions, respectively, as Yhr and Ylr. The
way. The network gains the ability to reduce MSE loss
residual image is computed as Yres=Yhr-Ylr by applying
between the high-resolution pictures used as the ground
Ylris as input data to the VDSR network and training it.
truth and its forecasted high-resolution output. Using back
After training, the calculated residual image and the
propagation and optimization techniques like stochastic
upsampled low resolution picture are added to recreate the
gradient descent (SGD), the network's weights are
high resolution image. The high quality image is finally
changed. VDSR may be used to progress the resolution of
converted back to its original RGB color model. This is
hidden low-quality pictures once it has been trained. A
the final, high-resolution picture that the VDSR network
low-resolution image is fed into the network, processed
has created. Downscaling (image resizing) is used in the
through its deep layers, and output as a high-resolution
proposed work to convert high resolution to low
image [17]. Compared to the original low-resolution
resolution images using multiple scaling factors, including
input, the final high-resolution image shows improved
0.1, 0.2, 0.3, 0.4, and 0.5. High frequency picture
details, sharpness, and overall quality. Figure 1 depicts the
components are lost during the process of downscaling.
architecture of the VDSR neural network.
The reference picture and low resolution image are related
in size by the scaling factor. Single picture super
979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

resolution becomes progressively ill-posed as the scaling III. EXPERIMENTAL RESULTS AND DISCUSSION
factor rises. This is brought on by the loss of the image's
high frequency components. The experimental outcome of the proposed
Utilizing blind and complete reference picture quality methodology for SISR is explained as follows. Figure
measures [21][22], the effectiveness of the VDSR 3(a) represents the test image to validate the efficiency of
network and bicubic interpolation approaches is the propose method. The test image (football.jpg) is
examined. collected from IAPR TC-12 Benchmark.In the proposed
 Full reference: The preeminence of an work, downscaling (image resize) is applied for the
experimental image may be assessed by transformation of high resolution to low resolution
contrasting it with a perfect (reference) image image with different scaling factors as 0.1, 0.2, 0.3, 0.4
[9][10][14]. MSE, for instance. and 0.5. During downscaling process, there is a loss of
 No reference (Blind): It is possible to estimate high frequency components of the image. The scaling
the eminence of a test image without factor links the size of low resolution and reference
comparing it to any reference images image. Fig 3(b) illustrates the low resolution image with
[15][16][20]. For instance, BRISQUE, NIQE. scaling factor of 0.2.

• Acquire test image (high resolution) from

Step 1 database

• Preprocessing of test image

Step 2

• Obtain low resolution image from high

Step 3 resolution one
Fig. 3. Outcome of the proposed method for noisy
• Apply bicubic interpolation to perform SISR image
Step 4

The bicubic interpolation is applied to the low

• Do SISR- VDSR neural network resolution image 2(b) to obtain the high resolution
Step 5
version. Using interpolation techniques like bilinear or
bicubic is a novel way to increase resolution without using
• Transformation of an image from RGB to
YCbCr Space machine learning. The value of the interpolated pixel is
Step 6
determined using this approach by taking a weighted
average of sixteen nearby pixels. Fig. 4 shows the results
• Split image in YCbCr space into three channels
(Y,Cb,Cr) of the bicubic interpolation procedure.
Step 7

• Start training VDSR using Y channel only

Step 8

• Apply Ylr as input to the VDSR network

Step 9

• Compute the residual image as Yres= Yhr - Ylr

Step 10
Fig. 4. Outcome of the bicubic interpolation
• Reconstruction of high resolution image by
adding the computed residual image and To calculate a residual image, or the difference
Step 11 upsampled low resolution image between high and low resolution photos, VDSR uses a
residual learning technique. High frequency features are
• Transformation of high resolution image into present in the resultant residual picture, which is essential
Step 12 original RGB color model
for SISR. In the suggested work, the RGB original picture
is transformed to a YCbCr color model, and the channel
• Evaluate the quality of SISR images using full ‘Y ‘ applied to create the residual image. In chrominance
Step 13 reference and blind image quality metrics
channels, the color information is represented by Cb
(chrominance blue) and Cr (chrominance red). Fig. 5
demonstrates the three channels of the YCbCr variant.
Fig. 2. Proposed Methodology for SISR

979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

Fig. 8. The high resolution image in RGB space

Fig. 5. Three channels of low resolution YCbCr

image (a) Y (b) Cb (c) Cr

The luminance (Y) channel only trained using VDSR

because human observation is very subtle to variations in
intensity (luminance) than to variations in chrominance.
IfYhr and Ylr as the luminance of high and low resolution
of images respectively. Ylris applied as input to the VDSR
network and train to compute the residual image as
Yres=Yhr- Ylr. This is illustrated in fig 6.

Fig. 6. The residual image obtained from the low

resolution Y image

After training, the high resolution image is

reconstructed by summing the computed residual image
and upsampled low resolution image. The high resolution
image in YCbCr space is shown in fig 7.
Fig. 9. The outcome of the proposed method for
different scaling factors as 0.1, 0.2, 0.3, 0.4 and 0.5.

In the proposed work, downscaling (image resize) is

applied for the transformation of high resolution to low
resolution image with different scaling factors as 0.1, 0.2,
0.3, 0.4 and 0.5 (row 1, row 2 and so on). This is
illustrated in fig 9. The column (a) indicates low
resolution image. The outcome of bicubic interpolation
Fig. 7. The high resolution image in YCbCr space and VDSR network is displayed in column (b) and
column (c) respectively.
Now the high resolution image is transformed into
original RGB color model. This is the outcome of ultimate The performance of bicubic interpolation and VDSR
high resolution image produced by VDSR network. This network methods are analyzed using blind and full
is reflected in fig 8. reference image quality metrics. This is illustrated in table
III and IV respectively.

979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

TABLE III. THE QUALITY ASSESSMENT FOR BICUBIC

INTERPOLATION SISR
Image SF=0.1 SF=0.2 SF=0.3 SF=0.4 SF=0.5
Quality
Parameter
PSNR 23.0429 25.1807 27.0303 28.5039 30.3828
MSE 0.0033 0.0026 0.0016 0.0014 0.0009
SSIM 0.8696 0.8923 0.9205 0.9276 0.9525
MultiSSIM 0.8295 0.8873 0.9422 0.9481 0.9790
NIQE 7.0143 6.0782 4.9664 5.4089 4.7542
PIQE 97.6723 94.8901 90.2080 82.6985 44.1532
BRISQUE 72.8537 56.4893 56.9074 53.6140 28.6345

TABLE IV. THE QUALITY ASSESSMENT FOR VDSR NETWORK SISR

Image SF=0.1 SF=0.2 SF=0.3 SF=0.4 SF=0.5
Quality
Parameter
PSNR 25.0408 26.1378 28.6453 29.1400 31.2235
MSE 0.0031 0.0024 0.0014 0.0012 0.0007 Fig. 10 (b). The assessment of proposed method using
SSIM 0.8696 0.8930 0.9286 0.9359 0.9602 blind metrics
MultiSSIM 0.8447 0.8887 0.9509 0.9557 0.9835
NIQE 7.4485 6.4770 5.0997 5.4275 5.9823
PIQE 94.9058 90.0828 83.4393 68.6292 17.2849 IV. CONCLUSION
BRISQUE 65.2791 52.8281 50.4896 46.1819 34.6121
Two separate methods are utilized to translate
The analysis of efficiency of bicubic and VDSR low-resolution inputs into high-resolution outputs: bicubic
method for SISR using both blind and full reference interpolation and Very-Deep Super-Resolution (VDSR)
image quality assessment is graphically illustrated in fig neural networks. Both approaches strive to improve the
10. image quality, but they do so using distinct strategies and
to varied degrees of success. Although it is a quick and
easy way to upsample low-resolution photos, bicubic
interpolation has limits when it comes to capturing fine
details. Alternatively, VDSR is a deep learning-based
method that necessitates training yet may provide
excellent, intricate, and crisp super-resolved pictures.
VDSR is slower and more computationally intensive than
bicubic interpolation, though. The effectiveness of both
strategies for SISR was validated using the proposed
method's blind and complete reference picture quality
metrics. The VDSR network performs well in every
quality metric. The proposed work has a wide range of
applications, including security, surveillance and image
and video compression. As future scope, the
amalgamation of best features of deep learning methods to
formulate a new method for efficient SISR.

Fig. 10 (a). The assessment of proposed method using REFERENCES

full reference metrics [1] J. Wu, Y. Wang and X. Zhang, "Lightweight Asymmetric
Convolutional Distillation Network for Single Image Super-
Resolution," in IEEE Signal Processing Letters,doi:
10.1109/LSP.2023.3286811.
[2] J. D. V. Ouwerkerk, "Image super-resolution survey", Image and
Vision Computing, vol. 24, no. 10, pp. 1039-1052, 2006, ISSN
0262–8856.
[3] Y. Yoo, D. Kim and J. Lee, "A Performance Comparison of Super
Resolution Model with Different Activation Functions", KIPS
Transactions on Software and Data Engineering, vol. 9, no. 10, pp.
303-308, 2020.
[4] Y. Li, Y. Iwamoto, L. Lin, R. Xu, R. Tong and Y. Chen,
"VolumeNet: A lightweight parallel network for super-resolution
of MR and CT volumetric data", IEEE Trans. Image Process., vol.
30, pp. 4840-4854, 2021.

979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.
 2023 IEEE 3rd International Conference on Applied Electromagnetics, Signal Processing, & Communication (AESPC)

[5] Q. Cai, J. Li, H. Li, Y. Yang, F. Wu and D. Zhang, "TDPN: [16] Kalist, V..: Possiblistic- Fuzzy C-Means Clustering Approach for
Texture and detail-preserving network for single image super- the Segmentation of Satellite Images in HSL Color Space. Procedia
resolution", IEEE Trans. Image Process., vol. 31, pp. 2375-2389, Computer Science. 57(2015) 49-56.
2022. [17] Sajiv G. Comparative Study of Possiblistic Fuzzy C-Means
[6] Q. Wang, Q. Gao, L. Wu, G. Sun and L. Jiao, "Adversarial multi- Clustering based Image Segmentation in RGB and CIELuv Color
path residual network for image super-resolution", IEEE Trans. Space. International Journal of Pharmacy and Technology. 8 (1);
Image Process., vol. 30, pp. 6648-6658, 2021. 2016; 10899-10909.
[7] K. Prajapati et al., "Direct unsupervised super-resolution using [18] V. R. Simi, D. R. Edla and J. Joseph, "A no-reference metric to
generative adversarial network (DUS-GAN) for real-world assess quality of denoising for magnetic resonance
data", IEEE Trans. Image Process., vol. 30, pp. 8251-8264, 2021. images", Biomed. Signal Process. Control, vol. 70, Sep. 2021.
[8] Y. Guo, X. Wu and X. Shu, "Data acquisition and preparation for [19] Stpień and M. Oszust, "A brief survey on no-reference image
dual-reference deep learning of image super-resolution", IEEE quality assessment methods for magnetic resonance images", J.
Trans. Image Process., vol. 31, pp. 4393-4404, 2022. Imag., vol. 8, no. 6, pp. 160, Jun. 2022.
[9] Sathish BS et al. Color Image Segmentation based on Genetic [20] K. Ding, K. Ma, S. Wang and E. P. Simoncelli, "Image quality
Algorithm and Histogram Threshold. International Journal of assessment: Unifying structure and texture similarity", IEEE Trans.
Applied Engineering Research. 10 (6); 2015; pp. 5205-5209. Pattern Anal. Mach. Intell., vol. 44, no. 5, pp. 2567-2581, May
[10] Sajiv G. Unsupervised Clustering of Satellite Images in CIELab 2022.
Color Space using Spatial Information Incorporated FCM [21] Kalist, V., Ganesan, P., Leo Joseph, L.M.I., Sathish, B.S.,
Clustering Method. International Journal of Applied Engineering Murugesan, R. (2020). Image Quality Analysis Based on Texture
Research. 10(20); 2015. pp. 18774-18780. Feature Extraction Using Second-Order Statistical Approach.
[11] W. Muhammad, S. Aramvith and T. Onoye, "SENext: Squeeze- Lecture Notes in Networks and Systems, vol 107. Springer,
and-Excitation Next for Single Image Super-Resolution," in IEEE Singapore.
Access, vol. 11, pp. 45989-46003, 2023, doi: [22] Ganesan, P., Sathish, B.S., Vasanth, K., Vadivel, M., Sivakumar,
10.1109/ACCESS.2023.3274679. V.G., Thulasiprasad, S. (2020). Color Image Quality Assessment
[12] S. Anwar, S. Khan and N. Barnes, "A deep journey into super- Based on Full Reference and Blind Image Quality Measures.
resolution: A survey", ACM Comput. Surv., vol. 53, no. 3, pp. 1-34, Lecture Notes in Networks and Systems, vol 107. Springer,
May 2021. Singapore.
[13] Y. Romano, J. Isidoro and P. Milanfar, "RAISR: Rapid and
accurate image super resolution", IEEE Trans. Comput. Imag., vol.
3, no. 1, pp. 110-125, Mar. 2017.
[14] K. Palanivel, B. S. Sathish, V. Kalist and K. B. Shaik:Performance
of fuzzy based clustering algorithms for the segmentation of
satellite images — A comparative study. 2015 IEEE Seventh
National Conference on Computing, Communication and
Information Systems (NCCCIS). Coimbatore, 2015, pp. 23-27.
[15] O. Ieremeiev, V. Lukin, K. Okarma and K. Egiazarian, "Full-
reference quality metric based on neural network to assess the
visual quality of remote sensing images", Remote Sens., vol. 12,
no. 15, pp. 2349, Jul. 2020.

979-8-3503-5874-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Rochester Institute of Technology. Downloaded on February 18,2025 at 01:45:10 UTC from IEEE Xplore. Restrictions apply.