Orthoimage Super-Resolution via Deep

Convolutional Neural Networks

Vladimir Berezovsky Yunfeng Bai Ivan Sharshov
Information Systems and Technologies Information Systems and Technologies Information Systems and Technologies
Northern (Arctic) Federal University Northern (Arctic) Federal University Northern (Arctic) Federal University
Arkhangelsk, Russia Arkhangelsk, Russia Arkhangelsk, Russia
0000-0003-1694-5896

Roman Aleshko Kseniya Shoshina Irina Vasendina

Information Systems and Technologies Information Systems and Technologies Information Systems and Technologies
Northern (Arctic) Federal University Northern (Arctic) Federal University Northern (Arctic) Federal University
Arkhangelsk, Russia Arkhangelsk, Russia Arkhangelsk, Russia
0000-0002-1235-8937 0000-0002-6716-4419 0000-0003-4504-0981

Abstract— Using high resolution (HR) images collected from to achieve a level of accuracy unattainable with other remote
UAV, aerial craft or satellites is a research hotspot in the field data Earth sounding. After identification and differentiation of
forest areas analyzing. In practice, HR images are available for individual breeds, schemes can be obtained distribution and
a small number of regions, while for the rest, the maximum crown contouring can be done for each breed. So maybe
density various around 1 px/m. HR image reconstruction is a valuable information was obtained, for example, the number
well-known problem in computer vision. In recent years, deep of trees of each species, the volume of wood of each tree. This
learning algorithms have achieved great success in processing information is of great importance for modern forestry.
images, thus we introduced them into the field of processing However, there are still unsolved problems related to the use
orthoimages. At the same time, we noticed that orthoimages
of multispectral images for forest analysis. Difficulty
generally have colorful blocks of different sizes. Based on this
feature, we did not directly apply the classic algorithm, but
classification of breeds can be explained by the following
made some improvements. Experiments show that the effect of factors: a huge variety of breeds on a small area, variability of
our proposed method is equivalent to that of the classic spectral parameters of old trees, healthy trees and trees,
algorithms, however, in the pre-processing stage, it saves time partially or completely shaded by other trees. Researchers
significantly. An approach to the forest areas analyzing, should use the latest data processing techniques to achieve
including image segmentation and the tree spices classification higher accuracy [7]. The use of super resolution in the
is proposed. The results of numerical calculations are presented. problems of processing satellite snapshots is one of the
advanced areas of using the machine learning apparatus in
Keywords—Orthoimages, Forest Area Analysis, Super- processing images [8-10].
Resolution, Deep learning.

I. INTRODUCTION
For effective tracking of forest use, it is necessary to
develop tools that allow detailed accounting of forest Fig. 1. Some orthoimages
resources. Today, the introduction of space-based imagery
and high-resolution systems, coupled with the development of II. PROPOSED METHOD
methods for automated interpretation of digital images, has There are many solutions which allows forest areas
become one of the main engines for the digital transformation analyzing, including image segmentation and the tree spices
of production processes. However, many years of experience classification. It’s important achievement, because these 3
in the use of satellite monitoring systems have revealed methods allow you to track the state of the forest and its
several shortcomings, including insufficient spatial resolution cutting. These methods require more information to divide
of data, long waiting times for surveys, dependence on trees from each other and retrieve features of a tree. There-
cloudiness and other atmospheric processes. The use of fore, they use high resolution images for teaching and
unmanned aerial vehicles over the past few years has shown processing which can be made by UAV or aerial craft.
the possibility of solving problems requiring high detail.
The solution to this problem can be HR satellite imagery
Today, in fact, there are no complete solutions that allow
and geological data, that can help a neural network to
solving problems in subject areas, such as forestry, with a
reconstruct the geometry and solve the segmentation and
sufficient level of reliability due to the low reliability and
classification tasks. And in this case the researcher will not be
detail of the existing remote means of automated obtaining
limited with the research area.
data on forest resources [1-6].
But HR satellite images are available for a small number
Satellite data can definitely be useful for harvesting
of regions, while for the rest, the maximum density various
organizations timber, protection or study of forests, namely
around 1 px/m. It is not enough for the NN to separate trees
for: calculation of available resources; assessment of forest
from each other and determine the tree characteristics. It leads
quality; logging planning; assessing the diversity of breeds in
us to the problem of a low-quality dataset that cannot be used
a particular area; reforestation works. Thanks to additional
for the task described above. But we can use super-resolution
channels (yellow, extreme red and NIR -2) satellites improve
methods to up-sample pixel density of these images.
the capabilities separation of various tree species. High spatial
resolution of images, e.g. satellite WorldView-2, allows you
 HR image reconstruction is a well-known problem in the rescaled VGG feature maps from the discriminator. The
computer vision. Its solutions have been widely used in network was trained on 350 thousand images from the
medicine, video and photo processing, social networks and ImageNet database and tested on Set5, Set14 and BSD100
wearable gadgets. Currently, there are several approaches, datasets. The new architecture was able to outperform
including using multiple LR images to restore one HR images. SRResNet on the BSD100 dataset.
But they are not suitable for solving this problem, since it is
not always possible to obtain several images of the same area. However, SRGAN had several weaknesses, some of
Therefore, the best solution seems to be the choice of which were eliminated later.
approaches to restore one image from one LR image. Based on it, the ESRGAN architecture [15] was created. It
Initially, the researchers created small networks with a low introduced a new Residual-in-Residual Dense Block (Fig 3),
level of super-resolution. These networks were not so deep which had a larger number of neurons and was easier to learn.
and had encoder-decoder architecture (Fig. 2). The first In addition, batch normalization has been removed, because it
several layers retrieve futures from an image and then they leads to artefacts in the generated image. Also in this work, an
pass to the decoder which reconstruct the image using inverse improved discriminator is described to obtain a more
convolution. perception beautiful image. These improvements made it
possible to improve the result of the generator: the sharpness
of the images and the number of small details in the
reconstructed image increased.

Fig. 3. The architecture of the RRDB.

Fig. 2. Encoder-decoder architecture.
Further improvements to this architecture included minor
But in 2015, an article appeared that proposed a way to modifications to improve the perceptual quality of the
create deep convolutional networks for solving the SR generated image [16]. For these purposes, the new residual
problem [11,12]. The goal of the researchers was to create a dense block has been introduced to augment the network
method for fast real-time image processing on the CPU. Their capacity without increasing its complexity. Also, noise has
experiments showed that network performance can grow with been added to the outputs of the blocks, the purpose of which
the number of hidden layers in the model. They also analyzed is to increase the number of small details in the image when
the network performance depending on the color model PSNR training with the GAN method. This approach made it
calculation. As the result, they proposed a lightweight possible to increase the perceptual index, but the PSNR
architecture that can be easily trained for the needs of the task. quality of the generated image decreased at the same time.
The big breakthrough was a variant of the idea of Another version of GAN for SR is PEGAN [17]. This
competition between neural networks in the form of an paper proposes a SR method based on feature fusion network.
adversarial-generative network which proposed by Dong et For the generator the researchers used the ensemble for feature
al.[13]. The concept is to train a generator that tries to extraction and details reconstruction in the wavelet space. Jing
reconstruct a high-resolution image. et al. used YСbCr color model for images and decided to SR
The result of the generator is evaluated by a discriminator only Y channel as the one which impacts more for human eye.
– a neural network that is trained to detect the generated The other channels are amplified by interpolation the LR
images, separating them from the original ones. Thus, by image. Such method showed 29 db of PSNR with 2 factor
passing information from the discriminator to the generator, it amplification and 25.5 db with 4 factor amplification on the
is possible to train the generator to reconstruct images Remo-A dataset.
correctly. The generator is a CNN consisting of residual We decided to use the ESRGAN architecture, because it
blocks (convolution, batch normalization, dense, convolution has higher pixel-wise performance than ESRGAN+. High
and batch normalization layers). Their input and output are resolution orthoimagery data [18] were taken as a reference
connected to a common image processing path. This modular dataset. The dataset provides high quality data in GeoTIFF
architecture allows balance adjusting of the network format. We selected images made in the RGB model from the
performance and the resources spent on the image processing. dataset, and then cut them into squares with a side of 100 m
The discriminator was built according to the guidelines of and a density of 10 px/m.
Radford et al. [14]. It contains 8 convolution layers with an III. SUPER-RESOLUTION
increasing number of 3 × 3 filter kernels, increasing by a factor
of 2 from 64 to 512 kernels as in the VGG network. The A. Related Work
resulting 512 feature maps are followed by two dense layers Deep learning has plenty of impressive algorithms for
and a final sigmoid activation function to obtain a probability image super-resolution. In 2014, the Super-Resolution
for sample classification. Convolutional Neural Networks (SRCNN) was first proposed
The researchers created a new loss function to solve the by Chao Dong et al.[19]. SRCNN is the first deep learning
problem of restoring finer details of an image. They algorithm which works better than traditional methods for
formulated the function as the sum of VGG loss, which image super-resolution. Since its invention, it has quickly
characterizes the loss of details of the generated image, and succeeded in processing microwave image, low frequency
radar image and underwater image [26][21][25]. In 2016, the
inventors of SRCNN proposed a more efficient algorithm values of these pixels are the same, no matter which
named Fast Super-Resolution Convolutional Neural Networks interpolation method is used, the same result will be obtained.
(FSRCNN) [20]. Their new approach is faster than the And it can be conceived that if the values of these pixels are
SRCNN more than 40 times [20]. In 2016, Wenzhe Shi et al. close, no matter which interpolation method is used, the close
proposed Real-Time Single Image and Video Super- result will be obtained. According to the comparison
Resolution Using an Efficient Sub-Pixel Convolutional described above about the three interpolation methods, it is not
Neural Network (ESPCN) [27] to increase the resolution from difficult to find that the nearest neighbor interpolation method
LR to HR only at the very end of the network and super- is suitable for interpolating in regions with relatively small
resolve HR data from LR feature maps. They proposed an value differences of pixels, the bicubic interpolation is suitable
efficient sub-pixel convolution layer to learn the upscaling for interpolating in regions with large value differences of
operation for image and video super-resolution, which pixels, and the bilinear interpolation is suitable for regions
eliminates the need to perform most of the SR operation in the where the differences is neither large nor small.
far larger HR resolution. In 2016, Jiwon Kim et al. proposed
a highly accurate SR method based on a very deep It is a common phenomenon that there are similarly
convolutional network [22]. Their algorithm is named VDSR colorful patches on rock orthoimages. In perspective of pixel
value, this phenomenon means the value of these pixels in the
in short. The contribution of VDSR to image super-resolution
is the strategy of residual-learning and gradient clipping. colorful block are close. Based on this phenomenon and
VDSR gives a significant boost in performance. In 2016, understanding of the characteristics of the three interpolation
Jiwon Kim et al. proposed another algorithm which is named methods, and inspired by our previous work [30], we propose
deeply-recursive convolutional network (DRCN) [23]. DRCN a hybrid interpolation algorithm to interpolate rock
applies same convolutional layer repeatedly and has a orthoimages to the desired scale, which is the first step for the
receptive field of 41 by 41, which guarantees improvement of first category of deep learning algorithms of image super-
performance. In 2017, Christian Ledig et al. proposed a super- resolution. The hybrid algorithm can achieve almost the same
resolution generative adversarial network (SRGAN) which interpolation performance as only using the bicubic
applies a deep residual network (ResNet) with skip- interpolation algorithm, however, it greatly reduces the
connection and diverge from mean squared error (MSE) as the amount of calculation.
sole optimization target [24]. In 2017, inspired by VDAR and In previous work, we combined bilinear interpolation and
DRCN, Ying Tai et al. proposed Image Super-Resolution via bicubic interpolation to make a hybrid interpolation algorithm,
Deep Recursive Residual Network (DRRN)[28]. DRRN this time we combined three interpolation algorithms. Thus, it
applies deeper network structure to obtain performance can save more time. This time we also applied statistical
improvement. In 2018, Xintao Wang et al. proposed Enhanced methods to determine thresholds. The work process of the new
Super-Resolution Generative Adversarial Networks hybrid interpolation method can be summarized into the
(ESRGAN) [29]. They thoroughly studied the network following steps. The first step is to determine the similarity of
architecture, adversarial loss and perceptual loss of SRGAN the four nearest pixels around the to-be-interpolated pixel. We
and improved it. Based this work, they won the first place in used formula (1) to measure the similarity of the four pixels
the PIRM2018-SR Challenge. [30]. The similarity of each color channel is calculated
The algorithms mentioned above can be divided into two separately.
categories. One category is to enlarge images to the desired
size through interpolation, after that, use interpolated images
from training set as input and the original high-resolution
images as labels to train a neural network, and then use the
trained neural network to improve to -be-processed target
images, such as SRCNN, VDSR, DRCN. The other is to use
LR as input without interpolation processing and original
high-resolution images as labels to train a neural network,
such as FSRCNN, ESPCN. The difference between the two
categories is whether interpolation preprocessing is required Fig. 4. Pixel arrangement of bilinear interpolation or nearest-neighbor
for the input images of neural networks. interpolation and bicubic interpolation.

B. Proposed Super-Resolution Algorithm

Combining the characteristics of orthoimages, we (1)
proposed an algorithm that mixes three interpolation
algorithms described to deal with them, and then drew on the
first category of deep learning algorithms for image super- where iT and jT are the coordinates of the to-be-interpolated
resolution to process interpolated image. Combining the pixel in the target image, Fi and Fj are the coordinates of the
characteristics of orthoimages, we proposed an algorithm that to-be-interpolated pixel in the original image. However, Fi and
mixes three interpolation algorithms described to deal with Fj are often not integers, thus the pixel point P(Fi , Fj ) cannot
them, and then drew on the first category of deep learning be found in the original image directly. Therefore, these two
algorithms for image super-resolution to process interpolated numbers must be found accordantly interpolation procedure.
orthoimages. Here we chose VDSR to improve interpolated
orthoimages. Since our main improvement was on the The second step is to determine which interpolation
interpolation part, we will mainly introduce our improved part algorithm is going to be applied. Fig. 5 is the logical diagram
here. The interpolation process could be seen as the value of of this method.
to -be-interpolated pixel is estimated by the value of several
pixels around it [30]. Thus, it can be concluded that if the
 We chose Python as the programming language for our
experiment. Experimental platform equips with two 10-core
Intel Xeon E5-2680v2 processors. Our experiment was
divided into two parts. The first part was about selection of the
thresholds and performance of the hybrid interpolation
algorithm. The second part is to examine the impact of the
interpolation algorithm on the performance of the first
category of deep learning algorithm for image super-
resolution. We choose VDSR as benchmark in second part.
We applied the bicubic interpolation algorithm to enlarge
selected 40 images randomly from 400 reduced images to 72
Fig. 5. Logic diagram of the proposed method.
* 72 with our proposed algorithm using different threshold
pairs in Table 1. And we calculated the peak signal-to-noise
It saves time to independently calculate the value of three ratio (PSNR) of the images obtained by the two algorithms.
channels. For example, the red channel satisfies the judgment
condition, and the green channel and the blue channel are not TABLE I. THRESHOLDS SETTING
satisfied. In the method proposed by Huipeng Wang et al [31], T1\T2 300 450 600 750 900 1050 1350 1500 1650 1800 1950
the pixel values of the three channels are all calculated by the
150 * * * * * * * * * * *
bicubic interpolation method. And in our mode, pixel value of
red channel is calculated with bilinear interpolation, and the 300 * * * * * * * * * *
pixel values of the green and blue channels are calculated 450 * * * * * * * * *
using the bicubic interpolation method. 600 * * * * * * * *
Based on the above analysis, we proposed the new hybrid 750 * * * * * * *
interpolation algorithm to deal with single image, and its flow 900 * * * * * *
chart is shown on Fig. 6. The bicubic interpolation algorithm has the best
performance in the traditional interpolation algorithm, and it
is also the standard preprocessing method in the first category
of deep learning algorithm. Thus, in step 6 we chose to use the
images obtained by the bicubic interpolation algorithm as
benchmark to find the threshold pair we need, allowing us to
save time while making the processing images obtained by our
hybrid interpolation algorithm as close as to that of the bicubic
interpolation algorithm. The PSNR is a widely-used metric for
quantitatively evaluating image restoration quality, and is at
least partially related to the perceptual quality [19]. The
formula of PSNR is:

 2
𝑃𝑆𝑁𝑅 = 10 log10 (𝐺𝑀𝐴𝑋 /𝑀𝑆𝐸) 

where:
1
 𝑀𝑆𝐸 = 𝑖=1 ∑𝑗=1[𝑓(𝑖, 𝑗) − 𝑓𝑏 (𝑖, 𝑗)] 
∑𝑁 𝑀 2

𝑁𝑀

where f(i, j) represents values of the pixels of a to-be-

valuated image, fb(i, j) represents values of the pixels of a
benchmark image. The parameter i represents the index of row
in the image and M is the number of rows of the image. The
parameter j represents the index of column and N is the
number of columns. GMAX is the maximum signal value in the
benchmark image.
We prepared the data sets with the hybrid interpolation
algorithm and the bicubic interpolation algorithm when
magnification factor was 2 and 4, respectively. Thus, there
were four training sets and four testing sets. Corresponding to
them, four trained models for the hybrid interpolation
algorithm and the bicubic interpolation algorithm when
magnification factor was 2 and 4, respectively were obtained.

Fig. 6. Flow chart of our proposed hybrid interplation algorithm.

 system, and parallel with support for the use of Intel® Xeon
Phi TM coprocessors. For running on Many Integrated Core
(MIC) architectures the OpenMP technique has been used. A
number of routines are optimized for running in the offload
mode of the Xeon Phi TM operation. Comparison of
computational performance when using a coprocessor and
without is presented in Fig.9.

Fig. 7. The average execution time performance comparison at a

magnification factor of 4.

Fig. 9. Acceleration with the Intel® Xeon Phi TM coprocessor in offload

image interpolation

Fig. 8. The PSNR performance comparison at a magnification factor of 4.

We find that the processing results of the two algorithms

almost completely overlap and compared with the benchmark,
the PSNR value have improved, which shows that our
preprocessing algorithm does not cause the first category of
deep learning algorithm adverse effects (Fig. 8).
C. Computing details
NArFU Computing cluster (HPC NArFU) with a peak Fig. 10. Host only, mic only, host+mic and overall cluster speed-up versus
number of computing
performance of 17.6 Tflops, has a hybrid architecture
consisting of twenty 10-core dual-processor nodes with an The interpolation of the orthoimages was carried out using
Intel Xeon processor, eight of which have Intel Xeon Phi dedicated written code, using the technique of message
coprocessors. The nodes are connected by the high- passing (MPI) with the help of spatial partitioning of the
performance interconnect Infiniband 56. The computing calculation area. It can be executed both on one processor
cluster has the following characteristics: system, and parallel with support for the use of Intel® Xeon
 20 computing nodes. Phi TM coprocessors. For running on Many Integrated Core
(MIC) architectures the OpenMP technique has been used. A
 Each node has two 10-core Intel Xeon E5-2680v2(2,8 number of routines are optimized for running in the offload
GHz) processors and 64GB of RAM. mode of the Xeon Phi TM operation. Comparison of
 At eight nodes, the Intel Xeon Phi 5110P (8GB, 1.053 computational performance when using a coprocessor and
GHz, 60 core) math coprocessors are additionally without is presented in Fig.9.
installed. To solve the bottleneck problem in transferring data to
memory, we used the collection of chunks into a large image
 Internal computer network for calculations: Infiniband
for its subsequent processing. An appropriate optimization
56 Gb/s.
was carried out, containing scalar tuning, vectorization with
 Network file system FEFS (Fujitsu Exabyte File SoA and memory optimization. OpenMP threads on the
System) with a capacity of more than 50 TB and a coprocessor were started in offload mode. Data between nodes
throughput of 1.67 GB/s (13.36 Gb/s). were synchronized with MPI processes ran on CPU nodes.
Scaling properties of code shown in Fig. 10.
 The cluster performance on the CPU in the LINPACK
test is 8.02 Tflops; on the CPU + Xeon Phi 7.68 Tflops, IV. CONCLUSION
the cumulative 15.7 Tflops.
An approach to the forest areas analyzing, including image
The interpolation of the orthoimages was carried out using segmentation and the tree spices classification is proposed.
dedicated written code, using the technique of message The results of numerical calculations are presented. Our main
passing (MPI) with the help of spatial partitioning of the contribution is to improve the pre-processing algorithm by
calculation area. It can be executed both on one processor using the characteristics of the orthoimage and reduce the time
required for pre-processing. When the amplification factor is
2 and 4, it can save 19.95% and 38.68% of the time. The [18] USGS EROS Archive - Aerial Photography - High Resolution
experiment demonstrates that it saves time significantly for Orthoimagery (HRO), reference:
https://www.usgs.gov/centers/eros/science/usgs-eros-archive-aerial-
the hybrid interpolation algorithm using statistical methods to photography-high-resolution-orthoimagery-hro, last accessed
select the thresholds, however, the effect is not significantly 14/02/2022.
lower than that of the bicubic interpolation algorithm. [19] C., Dong, C. C. Loy, K. He, X. Tang, “Image super-resolution using
Although we only selected VDSR in the first category of deep deep convolutional networks” IEEE Transactions on Pattern Analysis
learning algorithms for image super-resolution for testing, we and Machine Intelligence, vol. 38(2), pp. 295-307, 2015
believe that our interpolation algorithm is generally effective [20] Y. Dai, T. Jin, Y. Song, H. Du, “SRCNN-based enhanced imaging for
for first category. In addition, this method is a good choice for low frequency radar”, In: 2018 Progress in Electromagnetics Research
Symposium (PIERS-Toyama), pp.366-370. IEEE, 2018.
image super-resolution for other field, such as medical
[21] C. Dong, C. C. Loy, X. Tang, “Accelerating the super resolution
images. convolutional neural network”, In: European conference on computer
vision, pp. 391-407. Springer, Cham, 2016.
[1] S.E. Franklin, O.S. Ahmed, G. Williams, “Northern conifer forest [22] J. Kim, J. K. Lee, K. M. Lee, “Accurate image super-resolution using
species classification using multispectral data acquired from an very deep convolutional networks” In: Proceedings of the IEEE
unmanned aerial vehicle”, Photogrammetric Engineering and Remote conference on computer vision and pattern recognition, pp. 1646-1654,
Sensing, vol. 83(7), pp.501-507, 2017. 2016.
[2] E. Lopatin, A. Lopatina, “Assessing and mapping energy biomass [23] J. Kim, J. K. Lee, K. M. Lee, “Deeply-recursive convolutional network
distribution using a UAV in Finland”, Biofuels, vol. 8(4), pp. 485-499 for image super-resolution”, In: Proceedings of the IEEE conference
2017 on computer vision and pattern recognition, pp. 1637-1645, 2016.
[3] J.C. White, N.C. Coops, M.A. Wulder, M. Vastaranta, T. Hilker, P. [24] C., Ledig, L. Theis, “Photo-realistic single image super-resolution
Tompalski, “Remote sensing technologies for enhancing forest using a generative adversarial network”. In: Proceedings of the IEEE
inventories”, Canadian Journal of Remote Sensing, vol. 42(5), pp. 619- conference on computer vision and pattern recognition, pp. 4681-4690,
641, 2016 2017.
[4] C. Hüttich, M. Korets, S. Bartalev, V. Zharko, D. Schepaschenko, A. [25] Y. Li, C. Ma, T. Zhang, J. Li, Z. Ge, Y. Li, S. Serikawa, Underwater
Shvidenko, C. Schmullius, “Exploiting growing stock volume maps for image high definition display using the multilayer perceptron and color
large scale forest resource assessment: Cross-comparisons of ASAR- feature-based SRCNN. IEEE Access, vol. 7, pp. 83721-83728, 2019
and PALSAR-based GSV estimates with forest inventory in central [26] P. Shah, M. Moghaddam, “Super resolution for microwave imaging: A
Siberia”, Forests, vol. 5(7), pp. 1753-1776, 2014. deep learning approach” In: 2017 IEEE International Symposium on
[5] T. Sankey, J. Donager, J. McVay, J. B. Sankey, “UAV lidar and Antennas and Propagation & USNC/URSI National Radio Science
hyperspectral fusion for forest monitoring in the southwestern USA”, Meeting, pp. 849-850. IEEE, 2017.
Remote Sensing of Environment, vol. 195(15), 2017. [27] W. Shi, J. Caballero, J. Totz, “Real-time single image and video super-
[6] S. Holm, R. Nelson, G. Ståhl, “Hybrid three-phase estimators for large- resolution using an efficient sub-pixel convolutional neural network”,
area forest inventory using ground plots, airborne lidar, and space In: Proceedings of the IEEE conference on computer vision and pattern
lidar”, Remote Sensing of Environment, vol. 197, pp. 85-97, 2017. recognition, pp. 1874-1883, 2016.
[7] О. Hamdan “Using WorldView-2 multispectral data to identify areas [28] Y. Tai, J. Yang, X. Liu, “Image super-resolution via deep recursive
of tree species suitable for commercial timber harvesting”. residual network”, In: Proceedings of the IEEE conference on
GEOMATICS vol. 4, 2011 computer vision and pattern recognition, pp. 3147-3155, 2017.
[8] K. Jiang, Z. Wang, P. Yi, J. Jiang, “A progressively enhanced network [29] X. Wang, et al, “ESRGAN: Enhanced super-resolution generative
for video satellite imagery superresolution”, IEEE Signal Processing adversarial networks. In: Proceedings of the European conference on
Letters, vol. 25(11), pp. 1630-1634, 2018. computer vision (ECCV) workshops, pp. 0-0, 2018.
[9] M. Deudon, et al. “Highres-net: Recursive fusion for multi-frame [30] Y. Bai, V. Berezovsky, “Digital rock core images resolution
super-resolution of satellite imagery”, arXiv preprint enhancement with improved super resolution convolutional neural
arXiv:2002.06460, 2020. networks”. In: Proceedings of the 2020 12th International Conference
on Machine Learning and Computing. , pp. 401-405, 2020.
[10] J. Anger, T Ehret, C de Franchis, G. Facciolo, “Fast and accurate multi-
frame super-resolution of satellite images”, ISPRS Annals of [31] H. Wang, L. Zhou, J. Zhang, “Region-based bicubic image
Photogrammetry, Remote Sensing & Spatial Information Sciences vol. interpolation algorithm”, Computer Engineering, vol. 36(19), pp. 216-
5(1), 2020. 218, 2010.
[11] G. Barbastathis, A. Ozcan, G. Situ, “On the use of deep learning for
computational imaging”, Optica, vol. 6, pp. 921-943, 2019.
[12] C. Dong, C. C. Loy, K. He, X. Tang, “Image super-resolution using
deep convolutional networks”, IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 38(2), pp. 295-307, 2016.
[13] A. Radford, L. Metz, S. Chintala, “Unsupervised representation
learning with deep convolutional generative adversarial networks. In:
International conference on image and graphics, pp. 97-108. Springer,
Cham, 2017.
[14] N. C. Rakotonirina, A. Rasoanaivo, “ESRGAN+ : further improving
enhanced super-resolution generative adversarial network. In ICASSP
2020-2020 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pp. 3637-3641. IEEE, 2020.
[15] X. Wang, K. Yu, S. Wu, et. al.: “ESRGAN: Enhanced super-resolution
generative adversarial networks”, Lecture Notes in Computer Science
(including subseries Lecture Notes in Artificial Intelligence and
Lecture Notes in Bioinformatics), vol. 11133 LNCS, pp. 63-79, 2019.
[16] C. Ledig, L. Theis, F. Huszár, et.al., ”Photo-realistic single image
super-resolution using a generative adversarial network” In: 2017 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR). pp.
105-114, 2017.
[17] C. -W. Jing, Z. -X. Huang,, Z. -Y Ling, ”An image super-resolution
reconstruction method based on PEGAN”, in IEEE Access, doi:
10.1109/ACCESS.2022.3142049, pp. 1-1, 2017.