See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/377176154

Image enhancement using CNN

Article in International Journal of Advances in Electrical Engineering · January 2024

DOI: 10.22271/27084574.2024.v5.i1a.47

CITATIONS READS

0 27

1 author:

Tarushi Gupta
Chandigarh University
1 PUBLICATION 0 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tarushi Gupta on 18 September 2024.

The user has requested enhancement of the downloaded file.

 International Journal of Advances in Electrical Engineering 2024; 5(1): 10-14

E-ISSN: 2708-4582
P-ISSN: 2708-4574
IJAEE 2024; 5(1): 10-14 Image enhancement using CNN
© 2024 IJAEE
www.electricaltechjournal.com
Received: 08-11-2023 Tarushi Gupta
Accepted: 13-12-2023

Tarushi Gupta DOI: https://doi.org/10.22271/27084574.2024.v5.i1a.47

Computer Science Department
AITCSE AML Chandigarh Abstract
University, Punjab, India This study explores Image Enhancement using CNN experiments conducted using the LOL dataset has
the potential to greatly increase the quality of photographs taken in low light. It has been demonstrated
that CNN-based low-light picture enhancement techniques perform better than conventional techniques
in terms of both quantitative and qualitative criteria. CNNs are excellent for enhancing images in low
light because they can discover intricate correlations between picture data. A CNN is usually trained on
a dataset of paired low-light and improved pictures as part of CNN-based low-light image
improvement techniques. CNN gains the ability to create improved images that resemble the ground
truth images in the dataset by extracting characteristics from photos taken in low light. There are
several benefits that CNN-based low-light picture enhancing techniques offer over conventional
techniques. First, CNNs perform better than conventional techniques because they can understand
intricate correlations between visual data. Secondly, fresh low-light pictures can be improved using
CNN-based techniques without requiring paired training data. They become more useful and broadly
applicable as a result. Third, improved photographs appear more realistic and natural because CNN-
based techniques may retain semantic information in low-light photos.

Keywords: Deep learning, convolutional neural network (CNN), LOL dataset

Introduction
Pictures have an impact, on our lives acting as a powerful tool for communication and
storing information. However the quality of pictures can often suffer due to factors like
lighting, noise or limitations in the cameras abilities. Enhancing images is a field in
computer vision and image processing that aims to improve their quality while preserving
crucial details and information. [1]
Traditional techniques for enhancing images have relied on designed algorithms and filters
which often struggle to produce high quality results across various situations. In years
Convolutional Neural Networks (CNNs) have emerged as a game changing technology in
image enhancement. [2] They offer a data driven approach to the problem by learning patterns
and features from datasets. CNNs have showcased capabilities in tasks like image
classification, object detection and segmentation making them adaptable, for enhancing
images. [3]
Image enhancement using Convolutional Neural Networks (CNNs) represents a
groundbreaking approach in the field of computer vision and image processing. Images are
pivotal in various domains, including surveillance, medical imaging, photography, and more.
However, poor lighting conditions, noise, and sensor limitations can often degrade image
quality. [4]
CNNs bring several advantages to image enhancement. Firstly, they excel at feature learning,
automatically extracting hierarchical features from raw pixel data. This feature learning
capability enables them to capture intricate patterns, textures, and structures within images,
facilitating the enhancement of relevant aspects like edges, textures, and object contours.
Secondly, CNNs are remarkably adaptable to diverse scenarios. [5] Unlike traditional methods
that often require manual parameter adjustments or specific tuning, CNN-based models can
enhance images under various lighting conditions, weather situations, or with different
Correspondence
camera sensors without human intervention. [6]
Tarushi Gupta
Computer Science Department Understanding CNN Architecture
AITCSE AML Chandigarh Convolutional Neural Networks (CNNs) can be pivotal for grasping their significance in
University, Punjab, India modern computer vision and image processing.

~ 10 ~
International Journal of Advances in Electrical Engineering https://www.electricaltechjournal.com

CNNs are a class of deep learning models specifically hierarchical representation of the input, enabling the
designed to process grid-like data, such as images and network to capture both low-level details and high-level
videos. [7] They are inspired by the visual processing that abstractions. The combination of convolutional and pooling
occurs in the human brain, where neurons respond to stimuli layers allows CNNs to learn and represent complex patterns
only in specific receptive fields, allowing for the recognition within images efficiently. [14]
of patterns and shapes in visual input. [8] One of the key advantages of CNNs is their ability to learn
At the core of CNNs are convolutional layers, which apply a hierarchical features. In deeper layers of the network,
set of learnable filters (kernels) to the input image. These neurons respond to more complex and abstract features,
filters slide across the input, capturing local features and such as object parts or entire objects. This hierarchical
creating feature maps. The filters are learned through the representation is what makes CNNs particularly effective in
training process, allowing the network to automatically image recognition tasks, as they can learn to recognize
identify and extract relevant features from the input data. objects at various levels of granularity. [15]
This feature extraction capability is critical in tasks like Furthermore, CNNs often include fully connected layers at
image classification, object detection, and image the end of the architecture to perform classification or
enhancement. [9] regression tasks based on the extracted features. These
CNNs also incorporate pooling layers that down sample the layers aggregate information from the previous layers to
feature maps, reducing the spatial dimensions while make final predictions. [16]
preserving important features. [10] Pooling helps in creating a

Fig 1: CNN Architecture

Role of Deep Learning in Object Recognition images may exhibit diverse attributes, ensuring CNN-based
Deep learning, particularly through Convolutional Neural enhancement techniques remain versatile and effective
Networks (CNNs), has significantly reshaped the landscape across a range of conditions. [17]
of image enhancement by offering powerful tools capable of The inherent capability of CNN architectures to model
improving image quality across various domains and intricate relationships between input and output is
challenges. [11] CNNs excel in this domain due to their instrumental in image enhancement tasks. By understanding
innate capacity to learn intricate representations and the complex dependencies between low-quality images and
patterns from data, directly impacting the quality of their desired high-quality versions, CNNs can effectively
enhanced images. The fundamental strength of CNNs lies in restore missing details, reduce noise, and enhance overall
their ability to autonomously learn hierarchical features image quality. [18] This ability to capture and reconstruct
from raw image data without explicit human intervention. relevant image features facilitates the generation of visually
Through multiple layers comprising convolutions, pooling, improved outputs, significantly benefiting applications
and activation functions, CNNs progressively extract and reliant on high-quality image data. [19]
encode complex features, allowing for the understanding Furthermore, CNNs facilitate an end-to-end learning
and representation of vital elements within images. [12] approach, allowing the network to learn the mapping
A pivotal aspect contributing to CNNs' efficacy in image directly from low-quality input images to high-quality
enhancement is their adaptability and capacity for output images. This streamlined learning process eliminates
generalization. These networks can learn from diverse the need for intermediate manual steps, enabling the
datasets with varying image qualities, lighting conditions, network to generate enhanced images based on learned
and contexts. [13] Consequently, CNNs can generalize their patterns and features. [20] This aspect streamlines the
learned representations, enabling them to enhance images in enhancement pipeline, potentially improving efficiency and
scenarios not explicitly encountered during training. This effectiveness in producing high-quality image outputs.
adaptability is crucial in real-world applications where

~ 11 ~
International Journal of Advances in Electrical Engineering https://www.electricaltechjournal.com

the high-level features extracted from intermediate layers of

a pre-trained neural network, allowing the model to focus
on perceptually relevant aspects of image quality, such as
texture, structure, and content. [25]

Methodology
Image enhancement using Convolutional Neural Networks
(cnns) encompasses a systematic sequence of steps designed
to leverage deep learning for improving image quality.
Initially, the process involves gathering and preprocessing a
diverse dataset comprising low-quality images and their
corresponding high-quality references. These images
undergo normalization and resizing to create a clean dataset.
Subsequently, the selection of an appropriate CNN
architecture suited for image enhancement tasks, like U-Net
or resnet, becomes pivotal. Determining the architecture's
complexity, including layer numbers, filter sizes, and
activation functions, aligns with the task's intricacy and
computational resources available. [26]
The dataset is partitioned into training, validation, and test
sets after the architecture is chosen, guaranteeing balanced
distributions under varied lighting conditions or picture
Fig 2: Role of deep learning quality. The training data is the creation of input-output
pairs that associate low-quality photos with their high-
Literature Survey quality equivalents. [33] Following the initialization of the
Traditional image enhancement methods often encounter CNN with appropriate weights and an iterative optimization
limitations when dealing with real-world images that exhibit process, his data drives the training phase. Stochastic
diverse lighting conditions and complex content. gradient descent or Adam algorithms are used for parameter
Techniques like histogram equalization, contrast stretching, optimization, and methods like Mean Squared Error (MSE)
and spatial filtering have been used extensively, but they or perceptual loss are used as guides. [27]
lack the adaptability and learning capabilities that are Following the training phase, hyper parameters such as
inherent to CNNs. One of the primary drawbacks of learning rates and batch sizes are fine-tuned with the goal of
traditional methods is their reliance on fixed, handcrafted maximizing network performance while minimizing over
rules and parameters, which may not effectively capture the fitting. When evaluating the trained CNN, quantitative
nuanced characteristics of the images they aim to enhance. metrics like as PSNR, SSIM, or MSE are used to evaluate
In contrast, the introduction of CNNs into the realm of improvements in picture quality. These metrics are
image enhancement has brought about a paradigm shift. [21] supplemented with qualitative evaluations made by
CNN-based image enhancement methods adopt a data- comparing the improved images to their originals visually.
driven approach, leveraging the power of extensive datasets The improved photos are further refined by post-processing
containing pairs of input images and their corresponding methods like denoising filters or histogram equalization,
enhanced versions. These datasets are integral in training and the model is optimized for use in real-time applications.
CNN models to learn intricate mappings from raw input
images to their desired enhanced outputs. By doing so,
CNNs are capable of capturing complex and nonlinear
relationships between image features and the desired
enhancement effects. [22]
An essential feature of CNN-based image enhancement is
the concept of end-to-end learning. [31] In this context, end-
to-end learning implies that CNNs are trained to perform
the entire image enhancement process seamlessly, from the
initial raw input image to the final enhanced output. [22] This
eliminates the need for a multitude of handcrafted pre-
processing steps or the reliance on domain-specific
knowledge. CNNs take on the role of feature extractors,
Fig 4: Image enhancement using CNN
learners, and generators, allowing for a more holistic and
adaptive approach to image enhancement. [23, 32]
Conclusion
The training process of CNNs for image enhancement In conclusion, the application of Convolutional Neural
involves the use of loss functions that quantify the disparity Networks (CNNs) for image enhancement represents a
between the enhanced output and the ground truth, or groundbreaking advancement in computer vision [28]. These
reference, images. Common loss functions include mean techniques leverage deep learning to intelligently improve
squared error (MSE), which measures the pixel-wise image quality, addressing challenges such as noise
difference between the enhanced and reference images [24]. reduction and detail preservation. [29] The success of CNNs
In addition to traditional loss functions, perceptual loss in learning intricate image patterns ensures robust
metrics have gained prominence. Perceptual loss considers performance across diverse datasets, making them

~ 12 ~
International Journal of Advances in Electrical Engineering https://www.electricaltechjournal.com

accessible and applicable in various real-world scenarios. [30] networks for video recognition. In Proceedings of the
However, it is crucial to prioritize ethical considerations and IEEE/CVF International Conference on Computer
responsible deployment. [34] Overall, CNN-based image Vision. 2019:6202-6211.
enhancement signifies a transformative shift, promising 15. Chopra Y, Kaushik P, Rathore SPS, Kaur P.
superior visual content and paving the way for future Uncovering Semantic Inconsistencies and Deceptive
innovations in computer vision. [35] Language in False News Using Deep Learning and
NLP Techniques for Effective Management.
References International Journal on Recent and Innovation Trends
1. Chen C, Chen Q, Xu J, Koltun V. Learning to See in in Computing and Communication. 2023;11(8s):681-
the Dark. arXiv preprint arXiv:1805.01934; c2018. 692. DOI: 10.17762/ijritcc.v11i8s.7256.
2. Lai WS, Huang JB, Ahuja N, Yang MH. Deep Retinex 16. Kaushik P. Role and Application of Artificial
Decomposition for Low-Light Enhancement. In Intelligence in Business Analytics: A Critical
Proceedings of the IEEE Conference on Computer Evaluation. International Journal for Global Academic
Vision and Pattern Recognition (CVPR); c2018. & Scientific Research. 2022;1(3):01-11.
3. Fu X, Huang J, Ding X, Liao J, Paisley J. A Deep DOI: 10.55938/ijgasr.v1i3.15.
Learning Approach for Real-Time 3D Object 17. Kaushik P. Deep Learning Unveils Hidden Insights:
Recognition and Reconstruction in Low-Light Advancing Brain Tumor Diagnosis. International
Environments. In Proceedings of the IEEE International Journal for Global Academic & Scientific Research.
Conference on Computer Vision (ICCV); c2016. 2023;2(2):01-22. DOI: 10.55938/ijgasr.v2i2.45.
4. Sharma, S., Tyagi, A., Kumar, S., & Kaushik, P. 18. Kaushik P. Unleashing the Power of Multi-Agent Deep
(2022). Additive manufacturing process based EOQ Learning: Cyber-Attack Detection in IoT. International
model under the effect of pandemic COVID-19 on non- Journal for Global Academic & Scientific Research.
instantaneous deteriorating items with price dependent 2023;2(2):23-45. DOI: 10.55938/ijgasr.v2i2.46.
demand. In A. Editor & B. Editor (Eds.), Additive 19. Kaushik P, Rathore SPS. Deep Learning Multi-Agent
Manufacturing in Industry 4.0 (1st ed.). CRC Press. Model for Phishing Cyber-attack Detection.
5. Balamurugan, A., Krishna, M.V., Bhattacharya, R., International Journal on Recent and Innovation Trends
Mohammed, S., Haralayya, B. & Kaushik, P. (2022). in Computing and Communication. 2023;11(9s):680-
Robotic Process Automation (RPA) in Accounting and 686. DOI: 10.17762/ijritcc.v11i9s.7674.
Auditing of Business and Financial Information. The 20. Kaushik P, Miglani S, Shandilya I, Singh A, Saini D,
British Journal of Administrative Management, 58 Singh A. HR Functions Productivity Boost by using AI.
(157), 127-142. International Journal on Recent and Innovation Trends
6. Park S, Son J, Cho S, Hong S. SRFeat: Single Image in Computing and Communication. 2023;11(8s):701-
Super-Resolution with Feature Discrimination. In 713. DOI: 10.17762/ijritcc.v11i8s.7672.
Proceedings of the European Conference on Computer 21. Kaushik P, Singh Rathore SPS, Kaur P, Kumar H,
Vision (ECCV); c2018. Tyagi N. Leveraging Multiscale Adaptive Object
7. Zhang K, Van Gool L, Timofte R, Yang MH. Deep Detection and Contrastive Feature Learning for
Plug-and-Play Super-Resolution for Arbitrary Blur Customer Behavior Analysis in Retail Settings.
Kernels. In Proceedings of the European Conference on International Journal on Recent and Innovation Trends
Computer Vision (ECCV); c2018. in Computing and Communication. 2023;11(6s):326-
8. Huang X, Belongie S. Arbitrary Style Transfer in Real- 343. DOI: 10.17762/ijritcc.v11i6s.6938.
Time with Adaptive Instance Normalization. In 22. Kaushik P, Yadav R. Reliability design protocol and
Proceedings of the IEEE International Conference on block chain locating technique for mobile agent.
Computer Vision (ICCV); c2017. Journal of Advances in Science and Technology
9. Lim B, Son S, Kim H, Nah S, Lee KM. Enhanced Deep (JAST). 2017;14(1):136-141. DOI: 10.29070/JAST.
Residual Networks for Single Image Super-Resolution. 23. Kaushik P, Yadav R. Deployment of Location
In Proceedings of the IEEE Conference on Computer Management Protocol and Fault Tolerant Technique for
Vision and Pattern Recognition (CVPR); c2017. Mobile Agents. Journal of Advances and Scholarly
10. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image Researches in Allied Education [JASRAE].
Quality Assessment: From Error Visibility to Structural 2018;15(6):590-595. DOI: 10.29070/JASRAE.
Similarity. IEEE Transactions on Image Processing. 24. Kaushik P, Yadav R. Mobile Image Vision and Image
2004;13(4):600-612. Processing Reliability Design for Fault-Free Tolerance
11. Ma K, Duanmu Z, Wu Q, Yong H, Wang Z. Waterloo in Traffic Jam. Journal of Advances and Scholarly
Exploration Database: New Challenges for Image Researches in Allied Education (JASRAE).
Quality Assessment Models. IEEE Transactions on 2018;15(6):606-611. DOI: 10.29070/JASRAE.
Image Processing. 2018;27(2):1004-1016. 25. Kaushik P, Yadav R. Reliability Design Protocol and
12. Hore A, Ziou D. Image Quality Metrics: PSNR vs. Blockchain Locating Technique for Mobile Agents.
SSIM. In 2010 20th International Conference on Journal of Advances and Scholarly Researches in
Pattern Recognition (ICPR). IEEE; c2010. p. 2366- Allied Education [JASRAE]. 2018;15(6):590-595.
2369. DOI: 10.29070/JASRAE.
13. Wang Z, Simoncelli EP, Bovik AC. Multiscale 26. Kaushik P, Yadav R. Traffic Congestion Articulation
Structural Similarity for Image Quality Assessment. In Control Using Mobile Cloud Computing. Journal of
The Thrity-Seventh Asilomar Conference on Signals, Advances and Scholarly Researches in Allied
Systems & Computers, 2003. IEEE. 2003;2:1398-1402. Education (JASRAE). 2018;15(1):1439-1442.
14. Fan H, Malik J, He K, Feichtenhofer C. Slowfast DOI: 10.29070/JASRAE.

~ 13 ~
International Journal of Advances in Electrical Engineering https://www.electricaltechjournal.com

27. Rathore SPS. Analysing the efficacy of training

strategies in enhancing productivity and advancement
in profession: theoretical analysis in Indian context.
International Journal for Global Academic & Scientific
Research. 2023;2(2):56-77.
DOI: 10.55938/ijgasr.v2i2.49.
28. Rathore SPS. The Impact of AI on Recruitment and
Selection Processes: Analysing the role of AI in
automating and enhancing recruitment and selection
procedures. International Journal for Global Academic
& Scientific Research. 2023;2(2):78-93.
DOI: 10.55938/ijgasr.v2i2.50.
29. Rathore R. Application of Assignment Problem and
Traffic Intensity in Minimization of Traffic Congestion.
IJRST. 2021 Jul-Sep;11(3):25-34.
DOI: http://doi.org/10.37648/ijrst.v11i03.003.
30. Rathore R. A Review on Study of application of
queueing models in Hospital sector. International
Journal for Global Academic & Scientific Research.
2022;1(2):01-05. DOI: 10.55938/ijgasr.v1i2.11.
31. Rathore R. A Study on Application of Stochastic
Queuing Models for Control of Congestion and
Crowding. International Journal for Global Academic
& Scientific Research. 2022;1(1):01-07. DOI:
10.55938/ijgasr.v1i1.6.
32. Rathore R. A Study Of Bed Occupancy Management In
The Healthcare System Using The M/M/C Queue And
Probability. International Journal for Global Academic
& Scientific Research. 2023;2(1):01-09.
DOI: 10.55938/ijgasr.v2i1.36.
33. Sharma T, Kaushik P. Leveraging Sentiment Analysis
for Twitter Data to Uncover User Opinions and
Emotions. International Journal on Recent and
Innovation Trends in Computing and Communication.
2023;11(8s):162-169.
DOI: 10.17762/ijritcc.v11i8s.7186.
34. Sharma V. A Study on Data Scaling Methods for
Machine Learning. International Journal for Global
Academic & Scientific Research. 2022;1(1):23-33.
DOI: 10.55938/ijgasr.v1i1.4.
35. Yadav M, Kakkar M, Kaushik P. Harnessing Artificial
Intelligence to Empower HR Processes and Drive
Enhanced Efficiency in the Workplace to Boost
Productivity. International Journal on Recent and
Innovation Trends in Computing and Communication.
2023;11(8s):381-390.
DOI: 10.17762/ijritcc.v11i8s.7218.

~ 14 ~

View publication stats