https://scientific-jl.com/luch/
Часть-44_ Том-4_ Май-2025
334
IMPROVING IMAGE QUALITY BASED ON THE SRCNN
ALGORITHM
Turdali Jumayev Saminjonovich
Associate Professor of the Department of
“Modern information and communication technologies”, PhD
e-mail:
Abstract. Improving the quality of digital images, in particular, restoring
high-resolution versions of low-resolution images, is one of the most relevant
research areas in artificial intelligence today. This article considers an image
quality improvement algorithm based on the Super-Resolution Convolutional
Neural Network (SRCNN) model. SRCNN is the first super-resolution model
developed based on a convolutional neural network, which has high efficiency in
converting low-resolution images into high-quality ones. The article analyzes the
architecture, operating principle, advantages and disadvantages of the model, and
highlights areas of practical application. The results show that using the SRCNN
model provides much higher quality in restoring image clarity and detail compared
to traditional interpolation methods.
Keywords. Information security, biometric systems, face recognition, crash
control, super-resolution, SRCNN, image enhancement, artificial intelligence
INTRODUCTION
Digital images are widely used in almost all areas of our lives: medical
diagnostics, artificial vision systems, security cameras, multimedia, artificial
intelligence, etc. The accuracy and quality of images used in these areas directly
affect the reliability of the analysis results. However, due to various reasons -
technical limitations, signal transmission losses or the use of low-quality devices -
in most cases, low-resolution images are obtained. Therefore, the problem of
https://scientific-jl.com/luch/
Часть-44_ Том-4_ Май-2025
335
improving image quality is one of the most relevant areas of modern digital image
processing [1-2].
In recent years, the rapid development of deep learning methods has opened
up new opportunities for solving this problem. In particular, the SRCNN (Super-
Resolution Convolutional Neural Network) model has been recognized as one of
the effective solutions for restoring high-quality images from low-resolution
images. SRCNN is able to restore high-resolution images by learning the main
features of the image using a simple convolutional neural network architecture.
This article discusses in detail the working principle, architecture,
advantages, and practical applications of the SRCNN model.
MAIN PART
1. General idea of the SRCNN algorithm.
SRCNN is a super-resolution
(SR) model based on a convolutional neural network, which aims to restore a high-
resolution (HR) image from a low-resolution (LR) image. Unlike traditional
interpolation methods, SRCNN is able to restore lost image details based on learned
features [3-5].
2. SRCNN algorithm workflow
Stage 1: Pre-processing.
In the pre-processing stage of the SRCNN model,
the low-resolution image (LR) is first upscaled to the desired size using bicubic
interpolation. This stage provides the model with a high-resolution but low-quality
image as input.
Stage 2: Processing via a three-layer CNN
The model consists of three main convolutional layers:
3. Layers of the SRCNN architecture. Layer 1: Feature Extraction. This
layer detects the main structures (edges, textures, and contours) in the image.
✓
Convolution: Conv2D;
✓
Number of filters: 64;
✓
Filter size: 9×9;
https://scientific-jl.com/luch/
Часть-44_ Том-4_ Май-2025
336
✓
Activation function: ReLU.
Mathematical expression:
𝐹
1
(𝑌) = max(0, 𝑊
1
∗ 𝑌 + 𝐵
1
)
where,
✓
𝑌
: input interpolated image;
✓
*: convolution operation;
✓
𝑊
1
, 𝐵
1
: 1
- layer weight and shift parameters.
Layer 2: Non-linear Mapping.
In this stage, the features in the first layer
are transformed into higher-level abstractions.
✓
Number of filters: 32;
✓
Filter size: 1×1;
✓
Activation function: ReLU.
𝐹
2
(𝑌) = max(0, 𝑊
2
∗ 𝐹
1
(𝑌) + 𝐵
2
)
This layer is the main "intellectual" part of the neural network, preparing the
image for reconstruction.
Layer 3: Reconstruction.
This layer reconstructs the final HR image. It
produces the final result based on the features in the previous layer.
✓
Number of filters: 1;
✓
Filter size: 5×5;
✓
Activation function: none (linear).
𝐹
3
(𝑌) = 𝑊
3
∗ 𝐹
2
(𝑌) + 𝐵
3
4. Loss function.
The model is trained using the MSE (Mean Squared Error)
loss function:
𝑀𝑆𝐸 =
1
𝑛
∑(𝑌
𝑖
̂ − 𝑌
𝑖
)
2
𝑛
𝑖=1
where:
✓
𝑌
𝑖
̂
: model output;
✓
𝑌
𝑖
: original HR image;
https://scientific-jl.com/luch/
Часть-44_ Том-4_ Май-2025
337
✓
𝑛
: number of pixels.
5. Model training.
The model is trained on datasets such as ImageNet, Set5,
Set14, BSD500. For each input image, a low-resolution version is created and fed
to the model. The model is trained to produce results that are closer to the original
HR
image [6].
CONCLUSION
The SRCNN model is the first CNN-based super-resolution model that has
made a significant breakthrough in improving image quality. Its simple and
efficient architecture, feature-based approach, and high-quality results have made
it a model widely used in scientific and practical projects. Currently, many
advanced models (FSRCNN, VDSR, EDSR, ESRGAN) have been created inspired
by it.
REFERENCES
1. Pratt W.K. Digital image processing. New York: JOHN WILEY & SONG,
2001. - 792 p. Burge M. and Burger W. (1998a) Ear Biometrics. BIOMETRICS:
Personal Identification in a Networked Society, - P. 273-286.
2. Zhumaev, T. S., Mirzaev, N. S., & Makhkamov, A. S. (2015). Algorithms
for segmentation of color images based on the selection of strongly coupled
elements. Studies of technical sciences,(4), 22(27), 4.
3. Jumayev T.S. Shaxsni tanib olish masalasida tasvirdagi obyekt belgilarini
ajratish algoritmlari //O‘zbekiston jurnali “Islom tafakkuri”. – Toshkent, 1-son,
2021, -B.172-175.
4. Mahkamov, A. A., Jumayev, T. S., Tuhtanazarov, D. S., &
Dadamuxamedov, A. I. (2023). Using AdaBoost to improve the performance of
simple classifiers. In Artificial Intelligence, Blockchain, Computing and Security
Volume 2 (pp. 755-760). CRC Press.
https://scientific-jl.com/luch/
Часть-44_ Том-4_ Май-2025
338
5. Saminjonovich, T. J. (2022). Algorithm for extraction of identification
features in ear recognition. ACADEMICIA: An International Multidisciplinary
Research Journal, 12(1), 232-237.
6. Saminjonovich, J. T., & Abdujabborovich, M. A. (2024). ALGORITHM
FOR FEATURE EXTRACTION FROM EAR IMAGES ON THE BASE OF
DISCRETE
COSINE
TRANSFORMATIONS.
Web
of
Technology:
Multidimensional Research Journal, 2(2), 48-53.
