ENHANCING UNDERWATER IMAGERY: ADVANCED IMAGE QUALITY

IMPROVEMENT USING CYCLEGAN FOR MARINE RESEARCH AND

PHOTOGRAPHY

A PROJECT REPORT

Submitted by

AHAMED AASIK AFRIDI S

HARESH S
KATHIRAVAN V
SIBIRAJ P

in partial fulfilment for the award of the degree

of

BACHELOR OF TECHNOLOGY

IN

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

K. RAMAKRISHNAN COLLEGE OF
ENGINEERING
(AUTONOMOUS)
SAMAYAPURAM, TRICHY

ANNA UNIVERSITY
CHENNAI 600 025

NOVEMBER 2024
ENHANCING UNDERWATER IMAGERY: ADVANCED IMAGE QUALITY
IMPROVEMENT USING CYCLEGAN FOR MARINE RESEARCH AND
PHOTOGRAPHY

UAM1711 PROJECT PHASE I

Submitted by

AHAMED AASIK AFRIDI S (8115U21AM003)

HARESH S (8115U21AM015)
KATHIRAVAN V (8115U21AM021)
SIBIRAJ P (8115U21AM050)

in partial fulfilment for the award of the degree

of

BACHELOR OF TECHNOLOGY

IN

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

Under the Guidance of

Mr. M. PONNI VALAVAN

Department of Artificial Intelligence and Machine Learning

K. RAMAKRISHNAN COLLEGE OF ENGINEERING

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

K. RAMAKRISHNAN COLLEGE OF ENGINEERING

(AUTONOMOUS)
Under
ANNA UNIVERSITY, CHENNAI
 K. RAMAKRISHNAN COLLEGE OF ENGINEERING

(AUTONOMOUS)
Under
ANNA UNIVERSITY, CHENNAI

BONAFIDE CERTIFICATE

Certified that this project report titled "ENHANCING UNDERWATER IMAGERY:

ADVANCED IMAGE QUALITY IMPROVEMENT USING CYCLEGAN FOR
MARINE RESEARCH AND PHOTOGRAPHY” is the bonafide work of Mr. S.
AHAMED AASIK AFRIDI (8115U21AM003), Mr. S. HARESH (8115U21AM015),
Mr. V. KATHIRAVAN (8115U21AM021), Mr. P. SIBIRAJ (8115U21AM050) who
carried out the work under my supervision.

Dr. B. KIRAN BALA Mr. M. PONNI VALAVAN

HEAD OF THE DEPARTMENT SUPERVISOR

ASSOCIATE PROFESSOR, ASSISTANT PROFESSOR,
Department of Artificial Intelligence Department of Artificial Intelligence
and Machine Learning, and Machine Learning,
K. Ramakrishnan College of K. Ramakrishnan College of
Engineering, (Autonomous) Engineering, (Autonomous)
Samayapuram, Trichy. Samayapuram, Trichy.

SIGNATURE OF INTERNAL EXAMINER SIGNATURE OF EXTERNAL EXAMINER

NAME: NAME:

DATE : DATE :
 K. RAMAKRISHNAN COLLEGE OF ENGINEERING

(AUTONOMOUS)
Under
ANNA UNIVERSITY, CHENNAI

DECLARATION BY THE CANDIDATES

We declare that to the best of our knowledge the work reported here in has been composed
solely by ourselves and that it has not been in whole or in part in any previous application
for a degree.

Submitted for the project Viva- Voce held at K. Ramakrishnan College of Engineering
on .

SIGNATURE OF THE CANDIDATES

 ACKNOWLEDGEMENT

We thank the almighty GOD, without whom it would not have been possible
for us to complete our project.
We wish to address our profound gratitude to [Link],
Chairman, K. Ramakrishnan College of Engineering(Autonomous), who
encouraged and gave us all help throughout the course.
We extend our hearty gratitude and thanks to our honorable and grateful
Executive Director [Link], [Link]., MBA., Ph.D., K.
Ramakrishnan College of Engineering(Autonomous).
We are glad to thank our Principal [Link], M.E., Ph.D., FIE.,
MIIW., MISTE., MISAE., [Link], for giving us permission to carry out this
project.
We wish to convey our sincere thanks to [Link] BALA,M.E.,M.B.A.,
Ph.D., Head of the Department, Artificial Intelligence and Machine Learning
for giving us constant encouragement and advice throughout the course.
We are grateful to [Link] VALAVAN, M.E., Assistant Professor
Artificial Intelligence and Machine Learning, K. Ramakrishnan College
of Engineering (Autonomous), for his guidance and valuable suggestions
during the course of study.
Finally, we sincerely acknowledged in no less terms all our staff members, our
parents and, friends for their co-operation and help at various stages of this
project work.

AHAMED AASIK AFRIDI S (8115U21AM003)

HARESH S (8115U21AM015)

KATHIRAVAN V (8115U21AM021)

SIBIRAJ P (8115U21AM50)
 ABSTRACT

Underwater image enhancement is a pivotal field in computer vision,

addressing challenges like light absorption, scattering, and color distortion,
inherent in underwater environments. This project leverages the CycleGAN deep
learning model to transform degraded underwater images into visually appealing
and functionally enhanced outputs. By training on unpaired datasets, the system
overcomes the limitations of traditional enhancement methods, such as reliance
on paired datasets, color imbalance, and detail loss. The proposed solution
provides real-time adaptability across diverse underwater conditions, supporting
applications in marine biology, archaeological exploration, underwater robotics,
and media production. Integrating advanced algorithms, the project ensures
improved visual clarity, natural color restoration, and reduced noise, fostering
advancements in underwater imaging and promoting its utility across scientific,
industrial, and creative domains. This project sets a new standard for underwater
imaging, combining deep learning advancements with practical usability to
deliver clearer, more vibrant images, fostering advancements in science,
exploration, and media production.

v
CHAPTER TITLE PAGE
NO.
NO.
ABSTRACT v

LIST OF FIGURES ix

LIST OF ABBREVIATIONS x

1 INTRODUCTION 1

1.1 Deep Learning 1

1.2 Cycle GAN 1

1.3 Underwater Image Enhancement 2

1.4 Problem Statement 3

1.5 Objectives 3
1.6 Importance of Enhancing Underwater 5
Image
1.6.1 Scientific Research and Marine 5
Biology
2 LITERATURE SURVEY 7

2.1 Experimental Review of 7

Image Enhancement And
Restoration Techniques for
Underwater Imaging
2.2 Class-Conditional Attention Gan For 8
Underwater Image Enhancement
2.3 Gan-Based Underwater Image 9
Enhancement Algorithm
2.4 ResNet and GAN Hybrid Model for 9
Improved Image Clarity
2.5 WaterPairs Dataset for Enhanced Training 10
Of GAN Models
 2.6 Comparative Study of GAN Architectures 11
For Underwater Enhancement
2.7 Gan-Base Image Enhancement For 12
Underwater Object Detection
2.8 Trends in Deep Learning-Based 13
Image Enhancement for
Underwater Applications
2.9 Underwater Image Enhancement And 14
Restoration Techniques
3 SYSTEM ANALYSIS 15
3.1 Existing System 15

3.2 Limitations of the Existing System 16

3.3 Proposed System 17

3.4 Advantages of the Proposed System 20

3.5 Versatile Applications 22

4 SYSTEM REQUIREMENTS 23
4.1 Hardware and Software Requirements 23

4.1.1 Hardware Requirements 23

4.1.2 Software Requirements 23

5 MODULE DESCRIPTION 25
5.1 Modules 25

5.2 Dataset Handling 26

5.3 Cycle GAN Architecture 27

5.4 Feature Extraction 28
5.5 Image Enhancement 28
5.6 Model Evaluation 29
5.7 User Interaction 29
6 SYSTEM DESIGN 30

6.1 Flowchart 30
6.2 Data Flow Diagram 31
6.3 Sequence Diagram 32

7 CONCLUSION AND FUTURE 33

WORK
7.1 Conclusion 33

7.2 Future Work 34

APPENDIX A : SAMPLE CODE 35
APPENDIX B : SCREEN SHOT 43
REFERENCE 45
 LIST OF FIGURES

FIGURE NO. TITLE PAGE NO.

5.2.1 Dataset Types 27
5.3.1 Cycle GAN Architecture 28
6.1.1 Flowchart 30
6.2.1 Data Flow Diagram 31
6.3.1 Sequence Diagram 32
 LIST OF ABBREVIATIONS

ABBREVIATIONS EXPANSION
GAN Generative Adversarial Network
C-GAN Cycle-Consistent Generative
Adversarial Network
CNN Convolutional Neural Network
CA-GAN Class-Condition Attention Generative
Adversarial Network
PSNR Peak Signal-to-noise ratio
SSIM Structural Similarity Index Measure
 CHAPTER 1
INTRODUCTION

1.1 DEEP LEARNING

Deep learning is a branch of artificial intelligence that mimics the workings of the
human brain to process data and create patterns for decision-making. It relies on
artificial neural networks with multiple layers, enabling computers to learn from data
without explicit programming.

• Hierarchical Learning: Deep learning extracts features layer by layer,

progressing from simple to complex representations.
• Architectures: Includes Convolutional Neural Networks (CNNs) for image
processing, Recurrent Neural Networks (RNNs) for sequential data, and
Generative Adversarial Networks (GANs) for generating realistic outputs.
• Applications: From autonomous driving to medical imaging, deep learning
has revolutionized industries by solving problems in pattern recognition, data
generation, and prediction.

For this project, deep learning plays a crucial role in restoring underwater imagery,
leveraging the capabilities of GANs to enhance degraded images effectively.

1.2 CYCLE GAN

CycleGAN is a type of Generative Adversarial Network (GAN) specifically

designed for image-to-image translation without requiring paired datasets. It consists
of two generators and two discriminators:

• Generators: Transform images from one domain to another.

1
 • Discriminators: Evaluate how well the transformed images match the target
domain.
• Cycle Consistency Loss: Ensures that if an image is converted to another
domain and back, it remains unchanged.

In this project, CycleGAN is used to enhance underwater images by mapping

degraded images (e.g., with color distortion or low contrast) to a domain of visually
appealing and functionally improved images, even when paired training data is
unavailable.

1.3 UNDERWATER IMAGE ENHANCEMENT

Underwater image enhancement addresses the visual degradation caused by the

underwater environment, such as:

• Light Absorption: Water absorbs longer wavelengths (red, orange) first,

leaving images predominantly blue or green.
• Scattering: Suspended particles scatter light, reducing clarity.
• Color Distortion: Uneven absorption of light leads to imbalanced color
representation.
• Low Visibility: Turbidity and depth reduce the overall visibility and quality
of images.

These issues impact the usability of underwater images in marine research,

photography, and navigation. Enhancing these images requires advanced
computational techniques like CycleGAN to restore natural colors, improve contrast,
and reduce noise while preserving critical details.

2
1.4 PROBLEM STATEMENT

Underwater images suffer from significant degradation due to the inherent properties
of water and environmental conditions. Challenges include:

• Severe color distortion and dominance of bluish/greenish hues.

• Reduced sharpness and visibility due to scattering and low light.
• Lack of affordable, efficient solutions for real-time enhancement of images
for applications in marine research, underwater exploration, and media
production.

The objective of this project is to develop a robust CycleGAN-based solution for

enhancing underwater imagery. The system will:

• Automate the enhancement process.

• Address challenges like color correction, clarity improvement, and noise
reduction.
• Produce visually appealing and functionally accurate outputs, supporting
applications in marine research and photography.

1.5 OBJECTIVES

The primary objective of this project is to develop an advanced, automated system

for enhancing underwater images using CycleGAN, addressing the unique
challenges of underwater environments. The specific objectives include:

1. Improving Visual Clarity

• Enhance the sharpness and detail of underwater images by reducing
scattering and haze.
• Restore fine details that are often obscured in underwater conditions.

3
2. Restoring Natural Colors
• Correct color distortions caused by differential light absorption and
scattering.
• Rebalance and restore colors, ensuring images look natural and
visually appealing.
3. Leveraging CycleGAN
• Utilize unpaired datasets to perform effective image-to-image
translation.
• Implement CycleGAN with ResNet50 to ensure high-quality feature
extraction and efficient processing.

4. Real-Time Processing Potential

• Lay the groundwork for future real-time underwater image
enhancement applications.
• Design a system architecture adaptable for mobile and hardware-
constrained platforms.
5. Supporting Diverse Applications
• Enable applications in marine biology, underwater archaeology, and
professional photography.
• Provide a solution that benefits various industries, from scientific
research to creative media.
6. User-Friendly System Design
• Ensure the system is intuitive and accessible to users without
extensive technical expertise.
• Develop an interface that allows seamless image enhancement with
minimal input.

4
 7. Generalization Across Conditions
• Train the system to adapt to varying underwater environments,
including differences in depth, turbidity, and lighting conditions.
• Guarantee robust performance across diverse datasets and scenarios.

By achieving these objectives, the project aims to address critical challenges in

underwater imaging while creating a versatile and impactful solution for both
research and practical applications.

1.6 IMPORTANCE OF ENHANCING UNDERWATER IMAGES

Underwater image enhancement holds critical importance in various scientific,

industrial, and creative fields, as it addresses the inherent limitations of capturing
clear and meaningful images in underwater environments. These enhancements are
not merely aesthetic; they serve functional purposes that directly impact a range of
applications.

1.6.1. SCIENTIFIC RESEARCH AND MARINE BIOLOGY

Enhanced underwater images are indispensable for marine biologists and

environmental scientists studying aquatic ecosystems. Clear and accurate images
allow researchers to:

• Identify and Monitor Marine Species: High-quality images enable detailed

observation of marine flora and fauna, facilitating species identification,
population assessments, and behavioral studies.
• Assess Coral Reef Health: Coral reefs are sensitive indicators of ocean health.
Enhanced images help in identifying stress indicators like bleaching, algae
overgrowth, or physical damage, supporting conservation efforts.

5
 • Study Marine Ecosystems: Understanding interactions between species and
their habitats often requires detailed visual data, which enhanced images
provide.

This chapter lays the foundation for understanding the necessity and scope of
underwater image enhancement while highlighting the objectives. The subsequent
chapters will build on these concepts and provide a detailed analysis and
implementation of the proposed solution.

6
 CHAPTER 2
LITERATURE SURVEY

2.1 EXPERIMENTAL REVIEW OF IMAGE ENHANCEMENT

AND RESTORATION TECHNIQUES FOR UNDERWATER
IMAGING

Wang et al. (2019) provide an in-depth review of underwater image

enhancement and restoration methods, assessing both traditional and advanced
techniques to address underwater imaging challenges such as light scattering, color
distortion, and low visibility. Conventional approaches like histogram equalization,
contrast stretching, and dehazing are highlighted for their widespread use, though
they often yield inconsistent outcomes in underwater environments due to their lack
of adaptability to specific underwater conditions. These methods struggle to correct
color loss caused by selective light absorption, especially at varying depths where
red wavelengths are absorbed quickly, resulting in predominantly green-blue images.
Wang et al. identify GANs, particularly CycleGAN, as promising alternatives
because they can enhance images without needing paired high-quality and low-
quality datasets—a major advantage in underwater imaging where capturing such
pairs is difficult. By experimenting with GAN-based approaches, the authors
demonstrate that GANs enhance color fidelity, reduce haze, and improve visibility
more effectively than traditional methods. The study positions GANs as powerful
tools that could either supplement or replace conventional methods , particularly
for applications that demand high-quality imagery, such as marine
research, environmental monitoring, and underwater archaeology. Their findings
underscore the potential of GANs to produce more consistent, visually accurate
underwater images, paving the way for more sophisticated image processing solutions
in aquatic sciences.
7
2.2 CLASS-CONDITIONAL ATTENTION GAN FOR
UNDERWATER IMAGE ENHANCEMENT

Wang et al. (2020) introduce the Class-Condition Attention GAN (CA-

GAN), a novel GAN model designed for underwater image enhancement that focuses
on color correction and visibility improvements. CA-GAN incorporates advanced
attention mechanisms to selectively enhance different regions within an image based
on specific underwater conditions, such as turbidity, particulate levels, and depth.
This allows the model to perform localized corrections, dynamically adapting to the
environmental features present in each section of the image. By applying targeted
color adjustments and clarity improvements, CA-GAN is able to preserve finer
image details while enhancing color accuracy, producing high-quality images
suitable for applications like coral reef monitoring and marine biodiversity studies.
CA-GAN’s effectiveness is demonstrated through its ability to outperform other
GAN models in terms of color accuracy and visibility in underwater settings where
traditional methods often struggle. However, the authors note that CA-GAN’s
performance decreases in environments with extremely high turbidity, where
particles obscure visual clarity. This limitation suggests the need for further model
refinement to enhance CA-GAN’s performance in such challenging underwater
conditions. Overall, the study underscores the potential of attention-based GAN
models in real-time underwater applications that require enhanced visual fidelity.

8
2.3 GAN-BASED UNDERWATER IMAGE ENHANCEMENT
ALGORITHM

Xu et al. (2022) present a specialized GAN model tailored for addressing

underwater-specific imaging challenges such as color distortion, low contrast, and
haze. This model leverages deep learning to autonomously identify and correct
common underwater distortions, including blue-green color shifts and haziness due
to light scattering by suspended particles. Unlike traditional methods that require
extensive manual adjustments, this GAN-based model adapts to various underwater
conditions automatically, adjusting color balance and contrast based on factors like
depth and lighting. The study highlights how this model preserves image details
more effectively than traditional approaches, making it particularly valuable for
fields that depend on high-quality visual data, such as marine biology, environmental
studies, and underwater photography. Xu et al. find that the model outperforms
conventional enhancement methods, maintaining clarity and color fidelity, which is
crucial for scientific applications where image quality impacts data accuracy. By
reducing the need for manual intervention, this GAN-based solution offers a more
efficient method for large-scale underwater image processing, suitable for both
aesthetic and functional purposes in marine research.

2.4 RESNET AND GAN HYBRID MODEL FOR IMPROVED

IMAGE CLARITY

Krishnan et al. (2020) propose a hybrid model combining ResNet and GAN
architectures to improve underwater image clarity by addressing detail loss, color
degradation, and visibility challenges. This innovative approach uses ResNet’s
residual connections within the GAN framework to enhance edge definition and
brightness, producing clearer images with balanced color. The deep
9
residual layers in ResNet aid in retaining structural details, such as textures and shapes,
which are often compromised in traditional underwater imaging. This hybrid model is
particularly valuable for tasks requiring precise detail, such as marine life monitoring,
habitat documentation, and ecological studies, where feature detection is paramount.
The study shows that this hybrid approach outperforms standard GAN models in
structural clarity and color fidelity. However, Krishnan et al. acknowledge that the
model’s adaptability remains limited in environments with rapidly changing lighting
conditions, such as during dives in shallow or deep waters. They recommend future
research to refine the model’s responsiveness to dynamic lighting, which would
enhance its applicability across a wider range of underwater environments, improving
the effectiveness of underwater imaging for researchers and media professionals alike.

2.5 WATERPAIRS DATASET FOR ENHANCED TRAINING

OF GAN MODELS

Chen et al. (2024) introduce WaterPairs, a unique dataset composed of

paired underwater images specifically designed to enhance the training of GAN
models used in underwater image enhancement and object detection. This dataset
fills a critical gap in underwater imaging, as high-quality paired datasets are difficult
to obtain due to the challenges of capturing corresponding high-clarity and low-
clarity images in aquatic environments. WaterPairs offers high-resolution reference
images that enable more precise training, allowing GAN models to learn and perform
image adjustments that are both accurate and adaptive to underwater characteristics
such as color shift, turbidity, and varying light conditions. Chen et al. demonstrate
through experiments that using paired data significantly improves GAN
performance, as the models are better equipped to generalize across different
underwater conditions,

10
achieving greater accuracy in color and contrast corrections. While WaterPairs is
instrumental for training GAN models, the dataset’s current focus on specific
underwater conditions (e.g., certain types of water bodies or depths) limits its
broader applicability across all marine environments. The study suggests expanding
WaterPairs with a wider variety of underwater scenarios to improve its versatility,
thereby enabling models trained on WaterPairs to perform effectively in diverse
underwater environments.

2.6 COMPARATIVE STUDY OF GAN ARCHITECTURES

FOR UNDERWATER ENHANCEMENT

Kim et al. (2022) conduct a comprehensive comparative study of multiple

GAN architectures, including CycleGAN, in the context of enhancing underwater
images. By analyzing various models, they aim to identify the strengths and
limitations of each architecture in addressing underwater-specific challenges like
color distortion, haze, and poor contrast due to the unique properties of light under
water. The study reveals that CycleGAN variants stand out for their adaptability to
underwater environments, as they excel in correcting color distortions and enhancing
contrast through domain adaptation. This adaptability is especially beneficial in
conditions where traditional models fail, such as in deeper or more turbid waters.
However, the study also identifies performance variability among different GAN
architectures, with some models better suited to specific types of aquatic
environments or visual characteristics. These findings highlight the importance of
model selection based on environmental factors, such as depth and particulate levels,
to achieve optimal results. The study also points to the potential for hybrid
architectures that integrate the strengths of multiple GAN types, which could further
enhance underwater image quality and consistency across varied conditions.

11
2.7 GAN-BASED IMAGE ENHANCEMENT FOR
UNDERWATER OBJECT DETECTION

Fabbri et al. (2018) investigate the use of GANs to improve underwater

image quality specifically for object detection tasks, which play a vital role in
monitoring marine life, conducting ecological surveys, and assessing environmental
changes. Underwater images are often compromised by low contrast, color
distortion, and haziness, all of which make it challenging to distinguish and
accurately detect objects. The study highlights how GAN-enhanced images
significantly improve the visibility of critical features, thus enhancing detection
accuracy and enabling more reliable identification of marine species and other
objects in underwater scenes. Fabbri et al. demonstrate that GAN-driven
enhancements provide clear, well- balanced visuals that facilitate higher precision in
tasks like tracking species movement, identifying population distributions, and
observing underwater ecosystems. The authors underscore the potential of GANs to
transform object detection in underwater environments, making it feasible to detect
even subtle or camouflaged objects with greater accuracy, which is essential for both
scientific research and conservation initiatives. However, they also caution that for
GAN models to perform optimally, high-quality and large-scale datasets are
necessary. Collecting such data in underwater environments presents challenges,
given the complexity and variability of these settings. This reliance on extensive
training data could hinder GAN applicability in real-time fieldwork where
immediate results are needed. To address this, the authors suggest that future
research could focus on refining GAN architectures to function effectively with
smaller or low-quality datasets, or on developing hybrid models that combine GAN
enhancements with object detection algorithms. Such approaches would allow for
high-accuracy

12
detection even in data-constrained scenarios, thus broadening GAN utility across
various underwater research and monitoring contexts.

2.8 TRENDS IN DEEP LEARNING-BASED IMAGE

ENHANCEMENT FOR UNDERWATER APPLICATIONS

Lepcha et al. (2023) provide a comprehensive review of recent advancements

in deep learning-based image enhancement, particularly focusing on the role of
GANslike CycleGAN in underwater imaging. Underwater environments pose
unique challenges, including light scattering, absorption, and color distortion, which
traditional image enhancement techniques cannot effectively address. GANs,
however, have demonstrated transformative potential in underwater applications by
adapting to these specific environmental challenges and significantly enhancing
image quality. The authors detail how GANs have succeeded in producing clear,
vibrant visuals that are essential for both scientific research and commercial use,
such as in marine biology and underwater filmmaking. One of the notable strengths
of GANs, as highlighted in the study, is their ability to adjust to various water
conditions, providing versatile and adaptive image enhancement. Lepcha et al. also
address the significant computational costs associated with GANs, which limit their
real-time applicability in field environments where lightweight solutions are
necessary. The authors call for further research to optimize GAN architectures to
reduce computational demands while preserving high-quality image output, which
would expand the utility of GANs to real-time applications like underwater drones
and conservation monitoring. Additionally, the study proposes that exploring multi-
model frameworks that integrate GANs with other machine learning methods could
maximize their effectiveness across different underwater scenarios, paving the way
for real-time, high-fidelity underwater monitoring systems.

13
2.9 UNDERWATER IMAGE ENHANCEMENT AND
RESTORATION TECHNIQUES

Yang et al. (2019) conduct a comprehensive review of both traditional and

GAN- based underwater image enhancement techniques, highlighting the limitations
of older methods and the potential of GANs in addressing underwater imaging
challenges. Traditional methods, such as dehazing, histogram equalization, and
contrast adjustment, are often inadequate for underwater imaging due to the complex
light absorption and scattering that occurs under water, resulting in color distortion
and low contrast. GANs, however, have shown remarkable adaptability in
overcoming these limitations by learning the unique characteristics of underwater
environments. The authors emphasize that GAN-based models excel in producing
images with improved clarity, accurate color reproduction, and enhanced contrast,
making them highly suitable for underwater applications where visual accuracy is
critical, such as marine biology, archaeology, and environmental studies. The review
also highlights the versatility of GANs, which can adjust to varying underwater
conditions and produce consistently high-quality results. However, Yang et al. note
the importance of high-quality, diverse datasets for training GANs to ensure optimal
performance across different environments. They suggest that further development
of such datasets, tailored specifically to underwater imaging, could greatly enhance
GAN effectiveness. The authors advocate for ongoing research into GAN-based
techniques as they hold the potential to redefine standards in underwater image
enhancement, ultimately supporting a wide range of scientific and practical
applications.

14
 CHAPTER 3
SYSTEM ANALYSIS

This chapter provides an in-depth analysis of the existing systems and outlines the
proposed approach for underwater image enhancement. It explores the limitations of
current methodologies and emphasizes how the proposed system addresses these
challenges.

3.1 EXISTING SYSTEM

Underwater image enhancement has traditionally relied on techniques that are

broadly categorized into conventional image processing methods and modern
machine learning-based approaches. While these systems have contributed
significantly to addressing some challenges, they fall short in comprehensively
solving the complexities of underwater imaging.

Conventional Image Processing Methods

Traditional methods, such as histogram equalization, white balance correction, and

Retinex-based techniques, aim to improve contrast, correct color distortions, and
enhance image clarity. Histogram equalization redistributes the intensity of pixels to
enhance global contrast, which can help in revealing hidden details in low-light
underwater images. However, it often amplifies noise and lacks adaptability to
varying underwater conditions. White balance adjustment focuses on correcting the
greenish or bluish color casts that dominate underwater images, but these methods
are highly dependent on manual calibration and cannot dynamically adjust to
changes in the imaging environment. Similarly, Retinex-based methods attempt to
mimic the human visual system by enhancing local contrast and brightness.

15
Emergence Of Deep Learning Techniques

Deep learning has introduced transformative advancements in underwater image

enhancement, particularly with the advent of Convolutional Neural Networks
(CNNs) and Generative Adversarial Networks (GANs). These models are designed
to learn complex transformations by analyzing large datasets. For instance, CNNs
can learn to detect and enhance features such as edges, textures, and colors in
underwater images. GANs, particularly CycleGAN and its derivatives, have proven
highly effective in unpaired image-to-image translation, enabling them to learn
mappings between raw and enhanced underwater images without requiring paired
datasets. Models like FW-GAN and CA-GAN have addressed issues such as multi-
scale texture preservation and attention-based color correction, delivering superior
results compared to traditional approaches.

3.2 LIMITATIONS OF EXISTING SYSTEMS

Despite the advancements, existing systems face several challenges that limit their
performance and applicability:

1. Color Distortion: Light absorption and scattering underwater cause severe

color distortions. While some models address these issues, they often fail to
achieve consistent results across varying depths and turbidity levels.
2. Loss of Detail: Traditional methods frequently compromise image details,
particularly textures, while some deep learning models fail to adequately
preserve these intricate features during enhancement.
3. Dataset Dependency: Many deep learning models require large amounts of
paired datasets for training. Such datasets are not only scarce but also
expensive to create due to the need for controlled imaging conditions.

16
 4. Computational Overhead: Advanced models like GANs arecomputationally
demanding, requiring significant processing power for both training and
inference. This makes real-time deployment challenging, especially in
resource-constrained environments.
5. Generalization: Models trained on specific datasets often struggle to perform
well in unseen or drastically different underwater conditions, limiting their
utility in diverse real-world scenarios.
6. Integration Challenges: Existing solutions lack seamless integration with
underwater imaging systems, making them impractical for real-time
applications like marine exploration or underwater robotics.

These limitations underscore the need for a more robust, adaptable, and efficient
system capable of addressing the multi-faceted challenges of underwater image
enhancement.

3.3 PROPOSED SYSTEM

The proposed system is designed to overcome the limitations of existing underwater

image enhancement techniques by leveraging the power of CycleGAN, an unpaired
image-to-image translation method. The system is crafted to provide a robust,
adaptable, and efficient solution for underwater image enhancement, catering to
diverse environments and varying use cases. The primary objective is to deliver
visually enhanced underwater images with improved clarity, color correction, and
detail preservation without the reliance on paired datasets.

17
Leveraging Cyclegan For Image Enhancement

CycleGAN, a Generative Adversarial Network (GAN), is a state-of-the-art deep

learning model that excels at unpaired image-to-image translation. It learns to
transform images from one domain (raw underwater images) to another domain
(enhanced underwater images) by minimizing a cycle consistency loss. This
approach allows the system to operate without the need for large paired datasets,
addressing one of the critical bottlenecks in underwater image enhancement.

The system trains CycleGAN on unpaired datasets consisting of raw underwater

images and enhanced images sourced from various underwater environments.
During the training process, the model learns to generate enhanced images that
closely resemble the target domain while maintaining the original content and
structure of the input images.

Addressing Key Challenges

1. Color Distortion Correction: The CycleGAN model is adept at learning

complex color transformations. By employing adversarial loss, the model
ensures that the enhanced images have natural and visually appealing colors,
effectively addressing the bluish and greenish tints common in underwater
imagery.
2. Detail Preservation: The system incorporates additional loss functions, such
as perceptual loss, to ensure that fine textures and details are preserved during
the enhancement process. This is particularly important for applications
requiring high-detail imagery, such as marine biology and underwater
archaeology.

18
 3. Adaptability to Diverse Environments: The system is trained on diverse
datasets representing various underwater conditions, such as shallow reefs,
deep-sea environments, and murky waters. This ensures that the model
generalizes well across different scenarios, making it versatile and reliable.
4. Computational Efficiency: The proposed system employs optimization
techniques to reduce computational overhead, enabling near-real-time
processing. This makes it suitable for integration with underwater imaging
devices and drones.

Seamless Integration And User Accessibility

The proposed system is designed to integrate seamlessly with underwater imaging

workflows. It can be deployed as a standalone application or embedded into
underwater cameras and drones. A user-friendly interface allows users to upload raw
images and retrieve enhanced outputs effortlessly. This accessibility broadens its
appeal to a diverse user base, including marine biologists, photographers, and
researchers.

Iterative Model Refinement

The system includes a feedback mechanism for continuous improvement. Users can
provide feedback on the enhanced images, which is used to refine the model through
additional training iterations. This iterative process ensures that the system adapts to
emerging challenges and maintains high performance over time.

Real-Time Applications

Real-time enhancement capabilities are integrated into the system to cater to

applications such as underwater exploration and navigation. The optimized

19
implementation ensures that image processing occurs with minimal latency,
providing immediate feedback for dynamic environments.

3.4 ADVANTAGES OF THE PROPOSED SYSTEM

The proposed system addresses the challenges of underwater image enhancement

with a robust and innovative approach, providing numerous advantages over existing
systems. These benefits span technical performance, usability, and adaptability,
making the system highly impactful in various applications.

Enhanced Color Restoration

The system employs CycleGAN to effectively restore color distortions caused by

light absorption and scattering underwater. By learning complex color mappings, the
model compensates for the dominance of blue and green hues, restoring natural tones
to underwater images. This results in visually appealing outputs suitable for
scientific analysis and artistic purposes.

Improved Detail Preservation

One of the standout advantages is the preservation of intricate details and textures in
underwater images. Conventional methods often blur or over smooth fine details
during enhancement. The proposed system integrates perceptual loss during training
to retain textures and structural details, ensuring high-quality outputs.

20
Unpaired Data Training

Unlike traditional supervised methods that require paired datasets of raw and
enhanced images, the proposed system works with unpaired datasets. This
significantly reduces the dependency on costly and scarce paired data, making the
system more practical and scalable.

Adaptability To Diverse Conditions

The system is designed to adapt to a wide range of underwater environments,

including shallow coral reefs, deep-sea regions, and turbid waters. The training
dataset includes diverse samples, enabling the model to generalize well across
varying conditions. This adaptability is crucial for applications requiring consistent
performance in different underwater scenarios.

Real-Time Processing

Optimized for computational efficiency, the system delivers real-time or near-real-

time processing capabilities. This feature is particularly valuable for dynamic
applications such as underwater navigation, real-time exploration, and live video
feeds, where immediate feedback is essential.

User-Friendly Interface

The system provides a streamlined and intuitive interface, allowing users to enhance
underwater images with minimal technical expertise. This accessibility expands its
usability to non-technical users, such as marine biologists and underwater
photographers, enabling them to focus on their primary tasks without being hindered
by complex processes.

21
Scalability

The system's modular architecture ensures that it can be deployed across various
platforms, from high-performance computing clusters to resource-constrained
underwater devices like drones and robotic systems. This scalability broadens its
applicability and allows for seamless integration with existing workflows.

Feedback-Driven Improvement

Incorporating a feedback mechanism allows users to evaluate the quality of

enhanced images and provide input for further refinement. This iterative process
ensures that the system evolves and adapts over time, maintaining high performance
as new challenges and datasets emerge.

3.5 VERSATILE APPLICATIONS

The proposed system supports a broad range of applications, such as:

• Marine Biology: Enables detailed study of marine ecosystems by enhancing

visibility of flora and fauna.
• Underwater Archaeology: Improves clarity for examining artifacts and
submerged structures.
• Recreational Photography: Produces visually stunning images for both
professional and amateur photographers.

22
 CHAPTER 4
SYSTEM REQUIREMENTS

4.1 HARDWARE AND SOFTWARE REQUIREMENTS

The system requires a robust hardware and software setup to ensure seamless
operation and efficient processing. These requirements are tailored to meet the
demands of high-resolution image processing while maintaining accessibility for a
wide range of users.

4.1.1 HARDWARE REQUIREMENTS

• Processor: A multi-core processor (Intel i5 or higher) is recommended to

handle the computational demands of CycleGAN.
• Graphics Processing Unit (GPU): A CUDA-compatible GPU (NVIDIA
GTX 1060 or higher) is essential for accelerating the training and inference
processes.
• Memory: A minimum of 16 GB of RAM is required for managing large
datasets and intermediate computations.
• Storage: At least 500 GB of disk space is necessary to store datasets, models,
and outputs.
• Camera: High-resolution underwater cameras or drones equipped with
imaging capabilities are required for capturing raw images.

4.1.2 SOFTWARE REQUIREMENTS

• Operating System: Windows 10/11, macOS, or Linux with support for GPU
acceleration.

23
• Programming Language: Python is the primary programming language for
implementing CycleGAN and other system functionalities.
• Frameworks and Libraries: TensorFlow or PyTorch for deep learning;
OpenCV for image preprocessing; NumPy for numerical operations;
Matplotlib for visualization; and Flask or Django for building the user
interface.
• Development Environment: Anaconda Navigator or Visual Studio Code for
managing dependencies and providing a streamlined development workflow.

24
 CHAPTER 5
MODULE DESCRIPTION

5.1 MODULES

The system for enhancing underwater images using CycleGAN has been divided
into the following key components to ensure systematic development and
implementation:

Dataset Handling

• Collection and preprocessing of underwater imagery datasets.

• Splitting data into training, validation, and test sets.
• Normalization and augmentation to improve model robustness.

Cyclegan Architecture

• Design and training of the CycleGAN framework.

• Implementation of unpaired image-to-image translation using
generators and discriminators.
• Optimization using cycle consistency loss and adversarial loss.

Feature Extraction

• Integration of ResNet50 for feature extraction.

• Enhancement of critical details like edges, textures, and patterns in
images.

25
Image Enhancement
• Color correction to restore natural hues in underwater images.
• Reduction of noise and scattering effects to improve clarity.
• Contrast adjustment for better visual appeal and usability.

Model Evaluation

• Assessment of model performance using metrics like PSNR (Peak

Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and
qualitative visual analysis.
• Iterative refinement based on evaluation results.

User Interaction

• Development of an intuitive interface for users to upload and

enhance underwater images.
• Integration with the CycleGAN model to provide real-time or
batch-enhanced outputs.

5.2 DATASET HANDLING

This component focuses on managing the underwater datasets required for training
the CycleGAN model. Key processes include:

• Dataset Collection: Gathering diverse underwater images from publicly

available sources and custom-collected datasets.
• Preprocessing: Normalizing image sizes, applying augmentations (e.g.,
flipping, rotation), and converting images to appropriate formats.

26
 • Dataset Statistics: Analyzing dataset diversity in terms of depth, turbidity,
and lighting.
• Dataset Split: Dividing the data into 70% training, 15% validation, and 15%
testing subsets for optimal model performance.

Fig. 5.2.1 DATASET TYPES

5.3 CYCLEGAN ARCHITECTURE

This component implements the CycleGAN framework for image-to-image

translation:

• Generator Networks: Transform degraded underwater images into

enhanced images.
• Discriminator Networks: Distinguish between real and generated images to
improve generator performance.
• Cycle Consistency Loss: Ensures the original image can be reconstructed
when transformed back, maintaining structural integrity.

27
 Fig. 5.3.1 CYCLE GAN ARCHITECTURE

5.4 FEATURE EXTRACTION

The ResNet50 model is integrated into this component to improve image quality:

• Feature Extraction: Captures high-level features like edges and textures.

• Integration with CycleGAN: Enhances the generator’s ability to produce
sharper, more detailed outputs.

5.5 IMAGE ENHANCEMENT

This component focuses on applying specific enhancements to underwater images:

• Color Correction: Adjusting color imbalances caused by light absorption

underwater.
• Noise Reduction: Suppressing noise while preserving details.
• Contrast and Brightness Adjustment: Making images visually appealing
and functionally usable.

28
5.6 MODEL EVALUATION

To assess the quality of the enhanced images, this component uses:

• Quantitative Metrics: Metrics such as PSNR and SSIM to measure image

quality.
• Qualitative Analysis: Comparing raw and enhanced images visually to
identify improvements.

5.7 USER INTERACTION

This component enables user interaction with the system:

• Upload Interface: Allowing users to upload underwater images for

enhancement.
• Processing Interface: Displaying the enhanced output along with the
original image for comparison.
• Download Option: Enabling users to save enhanced images locally.

This structured approach ensures efficient development, evaluation, and

deployment of the underwater image enhancement system while maintaining
flexibility for future enhancements.

29
 CHAPTER 6
SYSTEM DESIGN

6.1 FLOWCHART

Fig.6.1.1 FLOWCHART FOR IMAGE ENHANCEMENT USING CYCLEGAN

30
6.2 DATA FLOW DIAGRAM

Fig. 6.2.1 DATA FLOW DIAGRAM FOR IMAGE ENHANCEMENT USING

CYCLEGAN

31
6.3 SEQUENCE DIAGRAM

Fig. 6.3.1 SEQUENCE DIAGRAM FOR IMAGE ENHANCEMENT USING CYCLEGAN

32
 CHAPTER 7
CONCLUSION AND FUTURE WORK

7.1 CONCLUSION

Underwater imagery has long been a challenging domain for researchers and
professionals due to inherent environmental factors such as light attenuation, color
distortion, and scattering. This project leveraged the capabilities of CycleGAN, an
advanced generative adversarial network, to enhance underwater images by
addressing these critical issues. Through rigorous experimentation and analysis, the
proposed system demonstrated significant improvements in image quality, including
increased PSNR and SSIM scores, reduced processing time, and higher user
satisfaction.

The project proved to be transformative in the field of underwater imaging by

introducing automation and accuracy in enhancement processes. It eliminated the
need for manual editing and consistently delivered high-quality outputs. The
integration of CycleGAN's deep learning capabilities with a user-friendly interface
extended the scope of applications for marine researchers, photographers, and other
stakeholders in the underwater domain.

Key contributions of this project include:

1. Development of a robust CycleGAN model to handle underwater-specific

challenges like color distortion and haziness.
2. Quantitative analysis of image enhancement using industry-standard metrics
such as PSNR and SSIM.
3. Implementation of a scalable and efficient system suitable for real-time
applications.

Overall, this project has made a meaningful impact on underwater image processing,
laying a foundation for further advancements in the field.

33
7.2 FUTURE WORK
While the project successfully achieved its objectives, there is ample scope for future
enhancements and extensions. Some potential directions for future work include:

1. INTEGRATION OF MULTI-MODAL DATA:

• Incorporate additional data sources like sonar or hyperspectral
imaging to enhance the contextual understanding of underwater
scenes.
• Combine visual data with depth maps for more comprehensive
scene reconstructions.
2. INTEGRATION WITH MARINE APPLICATIONS:
• Develop specific modules for underwater object detection and
classification using enhanced images.
• Collaborate with marine conservation organizations to monitor
biodiversity and detect marine hazards using the system.
3. ENHANCED INTERFACE DESIGN:
• Build an interactive web-based platform or mobile application for
wider accessibility.
• Include batch processing capabilities and cloud storage for large-
scale usage.

34
 APPENDIX A

SAMPLE CODE

import streamlit as st

import subprocess

import os

import pandas as pd

from PIL import Image

import supervision as sv

from inference import get_model

st.set_page_config(

page_title="Underwater Sea Image Enhancement",

page_icon="˘
k̇",
·

layout="wide",

initial_sidebar_state="expanded"

def load_model(model_path):

"""

Loads a YOLO object detection model from the specified model_path.

Parameters:

35
 model_path (str): The path to the YOLO model file.

Returns:

A YOLO object detection model.

"""

# model = YOLO(model_path)

modeeel = model_path

model =

get_model(model_id="imageenhancement/2",api_key="eSzt9jqUwL3SzdHp0dmr")

return model

def run_detection(name, checkpoint_dir, data_dir, epoch, test_name,

result_dir,input_height,input_width):

# Formulate the command to run the detection script

detection_command = f"python test_funiegan.py --name {name} --checkpoint_dir

{checkpoint_dir} --data_dir {data_dir} --epoch {epoch} --test_name {test_name} --

result_dir {result_dir} --input_width {input_width} --input_height {input_height}"

# Run the detection command using subprocess

process = [Link](detection_command, shell=True, stdout=[Link],

36
stderr=[Link])

stdout, stderr = [Link]()

# Check if detection ran successfully

if [Link] != 0:

[Link](f"Error occurred during detection: {[Link]()}")

return False

else:

# [Link]("Detection completed successfully!")

return True

def calculate_psnr_ssim(oppath, data_dir, oppathcsv, resize, width, height):

# Formulate the command to calculate PSNR and SSIM metrics

psnr_ssim_command = f"python calc_psnr_ssim.py --input_dir {oppath} --refer_dir

{data_dir} --output_dir {oppathcsv} {'--resize --width '+str(width)+' --height '+str(height)

if resize else ''}"

# Run the PSNR and SSIM calculation command using subprocess

process = [Link](psnr_ssim_command, shell=True, stdout=[Link],

37
stderr=[Link])

stdout, stderr = [Link]()

# Check if PSNR and SSIM calculation ran successfully

if [Link] != 0:

[Link](f"Error occurred during PSNR/SSIM calculation: {[Link]()}")

return None

else:

# [Link]("PSNR and SSIM calculation completed successfully!")

return [Link](oppathcsv, "quantitive_eval.csv")

def main():

[Link](" Underwater Sea Image Enhancement")

# Sidebar for input parameters

with [Link]:

[Link]("Detection Parameters")

name = 'test'

# name = st.text_input("Name", "test")

checkpoint_dir = 'pretrained'

38
 # checkpoint_dir = st.text_input("Checkpoint Directory", "pretrained")

epoch = 95

# epoch = st.number_input("Epoch", value=95)

test_name = st.text_input("Test Name", "test1")

result_dir = st.text_input("Result Directory", "results/funiegan01")

uploaded_file = st.file_uploader("Upload Image", type=["png", "jpg", "jpeg"])

input_height = st.number_input("Height",value=256)

input_width = st.number_input("Width",value=512)

if uploaded_file is not None:

# Save the uploaded file to the data_dir

data_dir = "Inputimages"

[Link](data_dir, exist_ok=True)

image_path = [Link](data_dir, uploaded_file.name)

with open(image_path, "wb") as f:

[Link](uploaded_file.getvalue())

[Link](f"Image uploaded successfully: {uploaded_file.name}")

# Display uploaded image

[Link]("Uploaded Image")

39
 uploaded_img = [Link](image_path)

[Link](uploaded_img, caption="Uploaded Image", use_column_width=True)

if [Link]("Run Detection and Metrics"):

if run_detection(name, checkpoint_dir, data_dir, epoch, test_name,

result_dir,input_height,input_width):

oppathcsv= result_dir + '/' + test_name

oppath= result_dir + '/' + test_name + '/single/predicted'

metrics_file = calculate_psnr_ssim(oppath,

data_dir,

oppathcsv,

resize=True,

width=256,

height=256)

if metrics_file is not None:

# [Link]("Metrics calculation completed!")

# Display output image if exists

output_path = [Link](result_dir, test_name, "single", "predicted",

uploaded_file.name)

40
 if [Link](output_path):

[Link]("Output Image")

output_img = [Link](output_path)

# [Link](output_img, caption="Output Image", use_column_width=True)

model = load_model("temp")

res = [Link](output_img,

confidence=0.25

detections = [Link].from_inference(res[0].dict(by_alias=True,

exclude_none=True))

# [Link](len(detections))

bounding_box_annotator = [Link]()

label_annotator = [Link]()

# annotate the image with our inference results

annotated_image = bounding_box_annotator.annotate(

scene=output_img, detections=detections)

annotated_image = label_annotator.annotate(

scene=annotated_image, detections=detections)

41
 [Link](annotated_image, caption='Detected Image',

use_column_width=True)

class_names = [Link]['class_name']

for class_name in class_names:

[Link](class_name)

if class_name=='shark':

[Link]("Shark Ahead, Don't Proceed that Side")

else:

[Link]("You can Proceed")

else:

[Link]("Output image not found. Please check.")

# Display the calculated metrics (CSV file)

[Link]("PSNR and SSIM Metrics")

metrics_df = pd.read_csv(metrics_file)

[Link](metrics_df, layout="wide")

if name == " main ":

main()

42
APPENDIX B

SCREENSHOTS

43
44
REFERENCES

[1] Wang, Y., Song, W., Fortino, G., Qi, L.Z., Zhang, W. and Liotta, A., 2019. An
experimental-based review of image enhancement and image restoration methods for
underwater imaging. IEEE access, 7, pp.140233-140251.

[2] Wang, J., Li, P., Deng, J., Du, Y., Zhuang, J., Liang, P. and Liu, P., 2020. CA-
GAN: Class-condition attention GAN for underwater image enhancement. IEEE
access, 8, pp.130719- 130728.

[3] Xu, B., Zhou, D. and Li, W., 2022. Image enhancement algorithm based on GAN
neural network. IEEE Access, 10, pp.36766-36777.

[4] Krishnan, H., Lakshmi, A.A., Anamika, L.S., Athira, C.H., Alaikha, P.V. and
Manikandan, V.M., 2020, December. A novel underwater image enhancement
technique using ResNet. In 2020 IEEE 4th Conference on Information &
Communication Technology (CICT) (pp. 1-5). IEEE.

[5] Li, X., Teng, G., An, P. and Yao, H.Y., 2023. MT-GAN: toward realistic image
composition based on spatial features. EURASIP Journal on Advances in Signal
Processing, 2023(1), p.46.

[6] Chen, L., Dong, X., Xie, Y. and Wang, S., 2024. WaterPairs: a paired dataset for
underwater image enhancement and underwater object detection. Intelligent Marine
Technology and Systems, 2(1), p.6.

45