Volume 9, Issue 2, February – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

A Survey on Image Denoising Techniques

1.
Kalyani Akhade. 2.Sakshi Ghodekar 3.Vaishnavi Kapse 4.Anuja Raykar 5.Sonal Wadhvane
Department of Computer Engineering
Sinhgad College of Technology and Science , Narhe, Pune.
Savitribai Phule Pune University, Pune.

Abstract:- In the digital era of the world images are vital part autoencoders. These models are designed to take the input
of life and media. This survey explores a wide array of image, convert it into a hidden representation, and then
image denoising methods, spanning traditional and recreate the original image. The goal is to minimize any
contemporary approaches. The review encompasses errors in the recreated image. By learning the underlying
classical filters, statistical methods, and modern machine structure of clean images during training, these models can
learning-based algorithms, with a focus on their principles, effectively denoise new, unseen images.
advantages and limitations. Through a systematic
examination of the literature, we categorize the denoising Another noteworthy technique is the use of generative
techniques based on their underlying methodologies and adversarial networks (GANs) for image denoising. GANs
applications. Insights are drawn from comparative analyses, includes a generator and a discriminator, involved in a
highlighting the trade-offs and performance variations competitive learning process. The generator tends to
across different approaches. Additionally, emerging trends produce denoised images, while the discriminator
and future directions in image denoising research are discriminate between clean and denoised images. This
discussed. This comprehensive survey serves as a valuable hostile training leads to the generation of realistic and high-
resource for researchers, practitioners, and enthusiasts in quality denoised images.
understanding of the different image denoising techniques.
II. LITERATURE SURVEY
Keywords:- Wavelet Transformer, Image Denoising,
Machine Learning. In [1], a novel approach called "unprocessing" for raw
image denoising which aims to undo the processing steps
I. INTRODUCTION applied to raw images, allowing for better denoising using
learned models.The unprocessing step could lead to better
Image denoising techniques play a vital role in the field utilization of learned models for raw image denoising.
of digital image processing, aiming to enhance the grade of
images by reducing unwanted noise and artifacts. Noise, In [2], deep raw image denoising specifically designed
often introduced during image acquisition or transmission, for mobile devices which presents a solution tailored for
can degrade the visual quality and affect the accuracy of mobile devices, considering computational constraints. It
subsequent image analysis tasks. Image denoising also addresses the challenges associated with deploying
techniques employ various algorithms and methods to deep denoising models on resource-constrained devices.
effectively remove or minimize noise, revealing clearer and
more informative images. In [3], BM3D-AMP, a novel image restoration
algorithm based on the BM3D denoising method which
Traditional image denoising methods typically involve integrates the BM3D denoising algorithm into the image
mathematical operations, statistical analyses, and filter- recovery process. Demonstrates the effecticacy of the
based approaches. These techniques focus on identifying discussed algorithm.Combines denoising techniques with
and suppressing noise while preserving essential image image recovery, potentially improving overall performance.
details. Common filters include Gaussian filters, median
filters, and bilateral filters, each with its own advantages and In [4], a method for learning a deep convolutional
limitations. neural network (CNN) as a denoiser prior for image
restoration having the efficacy of using a deep CNN as a
With the advancements in machine learning and deep denoiser prior in image restoration tasks. Addresses
learning, there has been a paradigm shift towards data- challenges in learning denoiser priors from data.
driven image denoising techniques. Convolutional Neural
Networks (CNNs) and other deep learning architectures In [5], a residual learning framework for deep
have demonstrated remarkable performance in denoising convolutional neural networks (CNN) to enhance image
tasks. These models learn complex representations from denoising beyond traditional Gaussian models .Proposes a
large datasets, enabling them to adaptively filter out noise residual learning approach, where the network learns the
patterns and generalize well to diverse image types. residual denoising functions instead of the direct mapping.
Also,it is noticed that residual learning can improve the
One popular approach in deep learning-based image performance of deep CNNs for image denoising tasks.
denoising is the use of autoencoders and variational

IJISRT24FEB673 www.ijisrt.com 688

Volume 9, Issue 2, February – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
In [6], addresses the challenge of blind image approaches, along with standardized benchmarking, will
denoising, where the noise level is not known in advance, by contribute to the continuous advancement of image
proposing a convolutional blind denoising model.It denoising technologies.
Introduces a convolutional blind denoising framework
capable of handling real photographs with unknown noise IV. CONCLUSION
characteristics. Addresses the challenge of real-world
denoising where noise characteristics may vary across Subsequently, an analysis was conducted on 10
images. It enlighten applicability of CNN-based denoising scholarly articles, and the salient findings were encapsulated
methods to real-world scenarios with varying and unknown in the literature review. The research delved into identifying
noise levels. gaps within the extant knowledge, considering the
articulation of the problem statement and its corresponding
In [7], the SURE-LET (Stein's unbiased risk estimate objectives. Additionally, a meticulously crafted activity
learned experts in transform domain) strategy behind image regimen was delineated. The developed system is
denoising based on learned experts in the transform engineered to cater to the requirements of the
domain, utilizing Stein's unbiased risk estimate. It focuses ultimate end user.
on denoising images while preserving important structural
information. REFERENCES

In [8], focuses on image denoising using sparse and [1]. T. Brooks, B. Mildenhall, T. Xue, J. Chen, D. Sharlet,
redundant representations over learned dictionaries. and J. T. Barron, “Unprocessing images for learned
Demonstrates a method based on sparse coding and learned raw denoising,” in Proc.IEEE/CVF Conf. Comput. Vis.
dictionaries for image denoising.Emphasizes the importance Pattern Recognit. (CVPR), Jun. 2019,pp. 11036–
of sparse representations in capturing essential image 11045.
features during denoising.Pioneers the use of learned [2]. Y. Wang, H. Huang, Q. Xu, J. Liu, Y. Liu, and J. Wang,
dictionaries and sparse representations for image denoising, “Practical deep raw image denoising on mobile
contributing to the broader field of sparse signal processing. devices,” in Proc. Eur. Conf. Comput.Vis. (ECCV),
Aug. 2020, pp. 1–16.
In [9], a sparsity-based approach to image denoising by [3]. C. A. Metzler, A. Maleki, and R. G. Baraniuk, “BM3D-
combining dictionary learning and structural clustering. AMP: A new image recovery algorithm based on
Demonstrates a method that leverages both dictionary BM3D denoising,” in Proc. IEEE Int. Conf. Image
learning and structural clustering to enhance the denoising Process. (ICIP), Sep. 2015, pp. 3116–3120.
process. [4]. K. Zhang, W. Zuo, S. Gu, and L. Zhang, “Learning
deep CNN denoiser prior for image restoration,” in
In [10], WINNet, an invertible neural network for Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
image denoising. Draws inspiration from wavelet transforms (CVPR), Jul. 2017, pp. 3929–3938.
to enhance denoising performance.Incorporates multi- [5]. K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang,
resolution analysis and sparse representations inspired by “Beyond a Gaussian denoiser: Residual learning of
wavelet techniques.Ensures invertibility in the neural deep CNN for image denoising,”IEEE Trans. Image
network, allowing accurate reconstruction of the original Process., vol. 26, no. 7, pp. 3142–3155, Jul. 2017.
image.Advances image denoising technology by combining [6]. S. Guo, Z. Yan, K. Zhang, W. Zuo, and L. Zhang,
invertible neural networks with wavelet-inspired “Toward convolutional blind denoising of real
insights.Potential applications in medical imaging, remote photographs,” in Proc. IEEE/CVF Conf. Comput.Vis.
sensing, and computer vision. Pattern Recognit. (CVPR), Jun. 2019, pp. 1712–1722.
[7]. T. Blu and F. Luisier, “The SURE-LET approach to
III. FUTURE SCOPE image denoising,”IEEE Trans. Image Process., vol. 16,
no. 11, pp. 2778–2786, Nov. 2007.
The future scope for image denoising is promising and [8]. M. Elad and M. Aharon, “Image denoising via sparse
involves several key areas of exploration. This includes the and redundant representations over learned
integration of deep learning techniques, such as CNNs and dictionaries,” IEEE Trans. Image Process.,vol. 15, no.
GANs, for more sophisticated denoising models. The 12, pp. 3736–3745, Dec. 2006.
development of end-to-end learning systems and the [9]. W. Dong, X. Li, L. Zhang, and G. Shi, “Sparsity-based
exploration of adversarial training aim to enhance model image denoising via dictionary learning and structural
robustness. Future research may focus on dynamic noise clustering,” in Proc. CVPR,Jun. 2011, pp. 457–464.
modeling, adaptive to different noise patterns in real-time, [10]. WINNet: Wavelet-Inspired Invertible Network for
and the extension of denoising techniques to handle Image Denoising Jun-Jie Huang , Member, IEEE, and
multimodal and multispectral imaging data. Real-time and Pier Luigi Dragotti , Fellow, IEEE. IEEE
hardware implementation for applications like video TRANSACTIONS ON IMAGE
streaming and surveillance, as well as domain-specific PROCESSING, VOL. 31, 2022.
denoising models for various imaging domains, are also
areas of interest. Exploring explainable AI, self-supervised
learning, and the fusion of classical and deep learning

IJISRT24FEB673 www.ijisrt.com 689