You are on page 1of 6

Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)

IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

Quality Enhancement for Drone Based Video


using FPGA
1Y.
Vedavyas, 2S. Sri Harsha, 3M. Sai Subhash, 4S. Vasavi
1,2,3,4
Department of Computer Science & Engineering, VR Siddhartha Engineering College, India
2022 International Conference on Electronics and Renewable Systems (ICEARS) | 978-1-6654-8425-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICEARS53579.2022.9751731

1
yvedavyas4077@gmail.com, 3sriharshaseelamneni448@gmail.com, 3 saisubhash654@gmail.com,
4
vasavi.movva@gmail.com
particularly good at capturing high-detail videos, they are
Abstract—Nowadays Drones are being widely used for commonly employed in various applications. In the vast
surveillance and various other activities. The video majority of circumstances, these videos must be seen in real
stream produced by the drone can be disturbing or can time like to protect against trespassing, border
contain noise data which might reduce the quality of the monitoring. It is difficult to monitor the area if the drone
video stream. The video stream can be enhanced so that cannot deliver sharp videos. Also, the observer of the drone
there is no disturbance in the video stream. The video video may miss a mandatory scene in the video, which
enhancement can be done in real-time with the help of could result in significant damage.
field programmable gate array (FPGA) which reduces The paper will focus on the model used to enhance the
the processing time with low energy consumption. Our drone video using FPGA in real time. The prototype will
project mainly focuses on enhancing the quality of the include various techniques like CLAHE, Saturation
video stream using enhanced super-resolution Adjustment, Gamma Reduction. The Contributions can be
generative adversarial networks (ESRGAN), summarized as follows:
contrast-limited Adaptive histogram equalization • We present the quality enhancement for drone-based
(CLAHE), Gamma Correction and Saturation Video prototype to enhance the drone video generated
Adjustment by integrating the image source in the in real-time. The model takes the drone video as input
drone with the FPGA. and is subjected to different stages of the model.
• The extracted frame from the video is sent to the
Index Terms—Image Processing, Drone Video, Field ESRGAN model, where the frame is enhanced and
Programmable Gate Arrays, Quality Enhancement, sent for gamma correction and saturation adjustment.
Enhanced Super-Resolution Generative Adversarial
Networks, Contrast-Limited Adaptive Histogram
Equalization, Gamma Correction, Saturation A. MOTIVATION
Adjustment . In a wide variety of applications, such as military and
scientific research, real-time surveillance and remote
I. INTRODUCTION monitoring are now becoming increasingly important.
Drones are used to reach remote places that humans cannot
The term "video enhancement" refers to the process of reach, and these drones can cover large areas in short time
enhancing the quality of a video. In the video enhancement effectively. Drone Coverage in a fraction of time is much
procedure, there are a variety of techniques to choose from. more effective than using the traditional approaches in any
Enhancing video qualities such as brightness, contrast, and given situation. These drones are often limited with
color are the most basic approaches. In addition, cost-effective cameras and the demand for real-time video
procedures such as Noise Reduction and Video De-noising streaming for surveillance applications is greater than ever,
are commonly employed in video enhancement. In a wide necessitating computationally expensive image processing
variety of applications, such as military, scientific architectures as well as innovative encoding and decoding
research, real-time surveillance and remote monitoring are techniques.
now becoming increasingly important. The demand for
real-time video streaming for surveillance applications is B. PROBLEM STATEMENT
greater than ever, necessitating computationally expensive Today Surveillance is done using drones, the images or the
image processing architectures as well as innovative videos which are collected from the drones may contain the
encoding and decoding techniques. disturbances due to the camera which is attached or due to
As, Drones are being extensively used around the globe, the network issues. Hence the video stream which is
it is necessary to enhance the quality of the video generated produced by the drone needs to be enhanced before it can be
by these drones, to perform different analytics later on used for any surveillance applications in real time scenarios.
the video generated and which will relatively increase The enhancement is done at the drone itself using Field
the accuracy of the analytics. The sensors of these cameras programmable Gateway array (FGPA) rather than
enhancing at receiver’s end.
are not crisp because the cameras installed inside the
drones must be cost-effective. Furthermore, these low-cost C. SCOPE
cameras do not include advanced video enhancement The scope of the proposed system enhances the video
technologies. Despite the fact that these cameras are not recorded by drones such as UAV's gliders and from go

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 29

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 19,2023 at 14:32:59 UTC from IEEE Xplore. Restrictions ap
Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

pros camera’s during flight time and can be used for The model in [6] will begin by converting the supplied
security surveillance for the places where humans cannot go image to a binary sequence. The image layers are
for gaining the information and is limited to day light with transformed into binary sequence since any hardware
clear conditions. device can only work with binary values. Noise reduction,
picture enhancement, and image restoration were among the
D. OBJECTIVES
image processing techniques employed. 8-bit 3 RGB values
• Enhance the quality of the video stream generated using are used to represent each image pixel. The complete image
FPGA. is then converted to a series of pixels that can be accessed
using the row and column values, and the parallelism
• Reduce the noise generated in the drone video stream.
method allows for faster operation and performance.
• Adjust the color saturation of the drone video stream. Instead of using the usual retinex method, [7] uses a
• Configure the contrast and the brightness of the video low-cost, high throughput retinex algorithm to improve the
quality of a real-time video stream. Using FPGA, they were
stream using FPGA.
able to achieve a throughput of 60 frames per second for a
E. ADVANTAGES 1920*1080 image with zero latency (0.2 MS). The entire
operation is divided into seven modules, each of which
• The model reduces the noise and distortion caused, by performs an image improvement algorithm for a real-time
using the histogram equalization and is more efficient. video stream. [8] uses the adaptive histogram equalization
• The enhancement of the video stream produced by the with dynamic clip-threshold (AHEwDC) technique to
provide good real-time performance. They used FPGA to
drone is done in real time. improve the visibility of low-light portions in the video
• High quality contrast along with smoother edges is sequence. At a threshold frequency of 75.84 MHz, the
achieved by using the model. technology they proposed can handle 30fps full HD video at
30 frames per second. Actual images and improved images
F. APPLICATIONS were compared using various AHEwDC thresholds, and the
optimized clip threshold produced the best enhanced
• The enhanced video sequence can be used in aerial
outcome of the input stream for all inputs. Invariant feature
photography in a wide variety of domains. extraction for object detection and classification is
• Scientific research in areas where human access is described in [12].
restricted like mines.
III. PROPOSED SYSTEM
• Increase the efficiency of the research by covering more
region is less time compared to the traditional Figure 1 Presents Architecture of the proposed system.
approaches.
• 24/7 surveillance of the borders to ensure no A. PROPOSED ARCHITECTURE
trespassing.
• The enhanced video stream makes the search and rescue
process more effective.

II. LITERATURE SURVEY


In [1] they proposed a fast and efficient real-time video
defogging method based on the pyramid model, which
addresses issues with existing defogging algorithms such as
overestimation of transmittance, color distortion of the sky
area, and poor real-time performance. The texture was
refined by combined bilateral filtering, and the image was
returned to its original size. Finally, the degraded video was
restored using a combination of the atmospheric light
scattering model and the inter-frame video defogging
theory. Histogram equalization is commonly employed in
image improvement because of its simplicity and
efficiency; it alters the improved image's mean brightness
and introduces a lot of noise and distortion. As a result, [2]
used a fuzzy intensity measure and adaptive clipping Fig. 1. Architecture of proposed system.
histogram equalization to eliminate noise and distortion. [3]
built a deep learning image enhancement model that The proposed architecture of Quality Enhancement for
improves image quality and reduces noise. Two deep Drone Based Video Using FPGA consists of a drone
learning blocks comprise the model. The stacking denoising connected to FPGA. The video stream that is generated by
auto encoder, which comprises of three denoising auto drone is fed to FPGA which enhances the video quality,
encoder layers, is used to boost the contrast and recover the model starts with gathering input data of aerial videos and
noise in the first block, and the second block is used to the data undergoes pre-processing to remove the noise, later
reduce the noise in the image in the second block. key frames are extracted from the aerial video and these

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 30

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 19,2023 at 14:32:59 UTC from IEEE Xplore. Restrictions ap
Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

individual frames are enhanced and put together to form the B. DATASET COLLECTION
enhanced video, and the final video is sent to the receiver in The datasets which we used are UG2 Dataset [11] and
real time. UAV Dataset. UG2 Dataset consists of videos collected
Each individual frame extracted from the drone video from UAV, Glider, and Ground videos. It consists of 684
undergoes through the same process. Initially, the frame is videos with 30fps frame rate. It was published in IEEE
given as input to the trained ESRGAN model, then the Winter Conference on Applications of Computer Vision
output of the model is subjected to CLAHE, Gama 2018. UAV Dataset consists of videos collected from
Correction, and saturation adjustment respectively. Finally, different UAV's. It consists of 50 videos of 70250 frames
the frame obtained after the model is free from noise and all with 30 fps frame rate.
the parameters of the frame are optimal, so that the analysis • link - https://github.com/rosauravidal/UG2-Dataset
can be performed without any ambiguity. C. SOFTWARE REQUIREMENTS
Figure 2 presents Flow Chart of the proposed model
• Python Programming language
• TensorFlow
• OpenCV
• Python Image Library (PIL)
• Verilog
D. HARDWARE REQUIREMENTS
• Drone
• FPGA- ALTERA FPGA Cyclone II
• 4 Gigabytes of Read-Only Memory
• Available Storage Capacity of 15 Gigabytes
E. CLAHE
Contrast Limited Adaptive Histogram Equalization
(CLAHE) is a version of Adaptive Histogram Equalization
(AHE) that prevents the contrast from being over-amped.
CLAHE works on tiles, which are small areas of an image
rather than the complete image. To remove the false
boundaries, the surrounding tiles are blended using bilinear
interpolation. This algorithm is used to enhance image
contrast.
F. GAMMA CORRECTION
Controls the image's overall brightness. Images that haven't
been properly adjusted can appear bleached out or
excessively dark. Changing the amount of gamma
correction affects not just the brightness, but also the
red-green-blue ratios. We remove this using Gamma
correction because many bits are assigned to bright tones
and few bits are assigned to dark tones in the image using
(1)
O= I^(1/G) (1)
Where I is the input image, G is gamma value, O is output
image.
G. COLOR SATURATION ADJUSTMENT
The Method is linked to a correlation between light
brightness and color saturation discovered by light
simulation over exposure. When the brightness of the light
changes, so does the colour saturation. The YCbCr colour
model is used to calculate changes in colour saturation and
brightness.
H. GAN-GENERATIVE ADVERSARIAL NETWORKS
A Convolutional Neural Network (CNN) is a Deep
Learning system that can take an input image and assign
importance learnable weights and biases to various
aspects/objects in the image while also distinguishing
between them. A generative adversarial network (GAN) is
Fig. 2 Flow Chart of the proposed model an ML model that pits two neural networks against each
other in order to improve prediction accuracy. GANs are

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 31

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 19,2023 at 14:32:59 UTC from IEEE Xplore. Restrictions ap
Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

often unsupervised and learn by cooperating in a zero-sum Step 3: Gamma Correction is applied on the Frame to the
game. overall brightness of an image using (1).
Instead of a simple residual network, the ESRGAN model
Step 4: Saturation Adjustment is applied to the frame to
uses a Residual-in-Residual block as a fundamental
convolution block. To avoid smoothing out the artefacts in make it more vibrant and colorful using (2) and (3).
the photos, the model lacks a batch normalization mlayer in
the generator. The ESRGAN generates images with a better cb' = ((cb-128) x saturation) + 128 (2)
representation of the image artefacts' sharp edges. To better cr' = ((cr -128) X saturation) + 128 (3)
approximate the probability of an image being real or fake, Where, Y=0.2989× R+0.5866xG+0.1145* B
the ESRGAN employs a Relativistic discriminator. cb =-0.1688× R-0.3312xG+0.5000× B
Figure 3 presents Working of ESRGAN model. cr=0.5000x R-0.4184xG-0.0816× B

IV. IMPLEMENTATION

This section presents the output and results of the proposed


system. The results are obtained when the suggested model
for Quality Enhancement for Drone Based Video using
FPGA is used to test UG2 dataset videos containing noise
and blurriness.
Figure 4 presents Unenhanced drone video frame from UG2
Dataset [11].

Figure. 3 Working of ESRGAN model [12]


I. FRAME EXTRACTION
The video is a combination of different frames or images so
for enhancing the video we need to first extract the frames
which are present in the video and then we will enhance it and
then all the enhanced frames are joined together to form a Fig. 4 Unenhanced drone video frame from UG2 Dataset
[11]
new enhanced video.
The frame enhancement will be done using ESRGAN, Figure 5 presents enhanced drone video frame After
CLAHE, gamma correction, saturation adjustment. The ESRGAN model enhancement.
frame which we extracted will be sent to the ESRGAN model
which will produce the high-quality image then the frame
will be sent to CLAHE which is a variant of Adaptive
histogram equalization which helps us in improving the level
of visibility in foggy environment and also helps in reducing
the noise while doing the histogram equalization. Then the
frame will be sent for Gamma Correction which helps in
correcting the brightness in the frame. Then the frame will be
sent to the Saturation Adjustment will enhances the image by
spreading the bits uniformly in the image such as when we
have high colored images the bit concentration will be more
at that point it will reduce so that the color contrast will be
uniform throughout the frame. Fig. 5 enhanced drone video frame After ESRGAN model
J. ALGORITHM enhancement.
Step 1: Enhance the image using pre trained ESRGAN Figure 6 presents drone video frame After gamma
model. correction.
Step 2: Contrast limiting adaptive histogram equalization is
applied on the enhanced frame from ESRGAN to adjusts the
global contrast of an image by updating the image
histogram's pixel intensity distribution.

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 32

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 19,2023 at 14:32:59 UTC from IEEE Xplore. Restrictions ap
Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

REFERENCES

[1] Lu Di, Yan Limin, A real-time video de fogging


algorithm, International Conference on Display Technology
(2020).
[2] Xiangyuan Zhu, Xiaoming Xiao, Tardi
Tjahjadi, Zhihu Wu, Jin Tang, Image Enhancement using
Fuzzy Intensity Measure and Adaptive Clipping Histogram
Equalization, IAENG International Journal of Computer
Science (2019).
[3] Che Aminudin, M.F., Suandi, S.A. Video surveillance
Fig. 6 drone video frame After gamma correction image enhancement via a convolutional neural network and
Figure 7 presents the final enhanced drone video frame. stacked denoising autoencoder.
Neural Comput & Applic (2021). https://doi.org/10.1007/s
00521-021-06551-0.
[4] K. Sun, M. Koch, Z. Wang, S. Jovanovic, H. Rabah and
S. Simon, "An FPGA-Based Residual Recurrent Neural
Network for Real-Time Video Super-Resolution," in IEEE
Transactions on Circuits and Systems for Video
Technology, doi: 10.1109/TCSVT.2021.3080241.
[5] B. Paulchamy, S. Chidambaram and J. Jaya, "Proficient
Technique for Satellite Image Enhancement Using Hybrid
Transformation with FPGA," 2021 International
Conference on Advances in Electrical, Computing,
Communication and Sustainable Technologies (ICAECT),
Fig. 7 final enhanced drone video frame 2021, pp. 1-11, doi: 10.1109/ICAECT49130.2021.9392560
[6] Sowmya, K., Rakshak Udupa, T. and Holla, S., 2020.
Figure 8 presents Comparison between before and after Implementation of an FPGA Real-Time Configurable
applying CLAHE. System for Enhancement of Lung and Heart
Images. Advances in Multidisciplinary Medical
Technologies ─ Engineering, Modeling and Findings,
pp.199-213.
[7] J. W. Park et al., "A Low-Cost and High-Throughput
FPGA Implementation of the Retinex Algorithm for
Real-Time Video Enhancement," in IEEE Transactions on
Very Large Scale Integration (VLSI) Systems, vol. 28, no.
1, pp. 101-114, Jan. 2020, doi:
10.1109/TVLSI.2019.2936260.
[8] Canran Xu, Zizhao Peng, Xuanzhen Hu, Wei Zhang,
Lei Chen, and Fengwei An, FPGA-Based Low-Visibility
Enhancement Accelerator for Video Sequence by Adaptive
Histogram Equalization with Dynamic Clip-Threshold,
Fig. 8 Comparison between before and after applying IEEE Transactions on Circuits and Systems I: Regular
CLAHE Papers (2020).
[9] Min-ling Zhu, Dong-yuan Ge Image Quality
V. CONCLUSION AND FUTURE WORK Assessment Based on Deep Learning with FPGA
Implementation, (2020).
The proposed system can be used to enhance the drone [10] Incremental Model (Software Engineering)
video which can be used for many purposes, as the drone - javatpoint. (2019, August). www.Javatpoint.Com.
are being widely used for many purposes the proposed https://www.javatpoint.com/software-engineering-increme
system will be helpful in enhancing the quality of video ntal-model.
collected by the drones which have lower camera quality,
the proposed system also reduces the noise in the video [11] Vidal, Rosaura G and Banerjee, Sreya and Grm,
which is being recorded from the drone. Klemen and Struc, Vitomir and Scheirer, Walter J. Ug2: A
Future work includes improving the accuracy of the model video benchmark for assessing the impact of image
and the model can be developed to perform similarly for restoration and enhancement on automatic visual
any given climate like fog and night, which can be done by recognition 2018 IEEE Winter Conference on Applications
using the techniques like defogging and night vision in this of Computer Vision (WACV).
case.

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 33

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 19,2023 at 14:32:59 UTC from IEEE Xplore. Restrictions ap
Proceedings of the International Conference on Electronics and Renewable Systems (ICEARS 2022)
IEEE Xplore Part Number: CFP22AV8-ART; ISBN: 978-1-6654-8425-1

[12] Che Aminudin, M.F., Suandi, S.A. Video surveillance


image enhancement via a convolutional neural network and
stacked denoising autoencoder.
Neural Comput & Applic (2021). https://doi.org/10.1007/s
00521-021-06551-0.
[13] Sharma, Rajesh, and Akey Sungheetha. "An Efficient
Dimension Reduction based Fusion of CNN and SVM
Model for Detection of Abnormal Incident in Video
Surveillance." Journal of Soft Computing Paradigm (JSCP)
3, no. 02 (2021): 55-69.
[14] Huang, Wenzhun, Shanwen Zhang, and Harry
Haoxiang Wang. "Efficient GAN-based remote sensing
image change detection under noise conditions." In
International conference on image processing and capsule
networks, pp. 1-8. Springer, Cham, 2020.
[15] J.-W. Chang, K.-W. Kang, and S.-J. Kang, “An
Energy-Efficient FPGA Based Deconvolutional Neural
Networks Accelerator for Single Image Super-Resolution,”
IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 1,
pp. 281–295, 2020.
[16] X. Wang et al., “ESRGAN: Enhanced super-resolution
generative adversarial networks,” in Proc. Eur. Conf.
Comput. Vis., 2018, pp. 1–16.

978-1-6654-8425-1/22/$31.00 ©2022 IEEE 34

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 19,2023 at 14:32:59 UTC from IEEE Xplore. Restrictions ap

You might also like