Indonesian Journal of Electrical Engineering and Informatics (IJEEI) Vol.
No.
September 2025, pp.
ISSN: 2089-3272.
DOI: 10.
52549/ijeei.
Blind Image Quality Metric for Color Images Based on Human Vision System and Deep CNN Ali Erdem Altnba1.
Yldray Yalman2 1Department of Electronics and Communication Engineering.
Kocaeli University.
Turkey 2Department of Information Systems Engineering.
Piri Reis University.
Turkey
Article Info
ABSTRACT
Article history:
Abstract: This article introduces a novel blind image quality metric (BIQM) for color images which is designed taking into account human visual system The BIQM has a four-stage framework: RGB to YUV transformation, denoising with convolutional neural network , quality evaluation, and weighting to make it compatible with the human visual Experimental results, including Spearman's rank-order correlation coefficient, confirm BIQM's effectiveness, particularly in scenarios involving white noise and its compatibility with the human visual system.
Furthermore, a survey involving 100 participants ranks images based on three distinct qualities, validating the method's alignment with the human visual system.
The comparative analysis reveals that the proposed BIQM can compete with commonly used non-referenced quality measures and is more accurate than some of them.
The MATLAB codes for the development of the BIQM are made available through the provided link: https://bit.
ly/49MrbFX Received Feb 21, 2024 Revised Aug 10, 2025 Accepted Aug 27, 2025 Keyword:
blind image quality metric deep CNN no-reference image quality human vision system Copyright A 2025 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Ali Erdem Altnba.
Department of Electronics and Communication Engineering.
Kocaeli University.
Kabaolu.
Baki Komsuolu Bulvar No:515.
Umuttepe, 41001 zmit/Kocaeli Email: alierdemaltinbas@gmail.
INTRODUCTION
The determination of perceived image quality by the human eye holds significant importance in visual communication and image processing.
In communication systems, digital images undergo a series of processes including digitization, compression, and transmission before reaching the end-user .
Moreover, image quality plays a crucial role in the concept of object identification within biomedical imaging, which represents one of the most prevalent applications of image processing .
Quality metrics for digital images encompass three distinct measurement methodologies: fullreference (FR), no-reference (NR), and reduced-reference (RR) .
RR studies employ partial image information or specific features to assess image quality.
FR metrics, widely documented in the literature .
, rely on numerical comparisons but face challenges in integration with the human visual system (HVS).
Additionally, a significant limitation of FR studies is their dependence on a reference image for comparison.
Prominent examples of FR metrics include Peak Signal to Noise Ratio (PSNR).
Structural Similarity (SSIM) .
, and Universal Quality Index (UQI) .
In the literature, the frequency of proposals for no-reference (NR) metrics is comparatively lower.
Mathematical comparison methods are dysfunctional because there is no reference image in the NR metrics.
From this point of view, in the field of NR quality metrics, various studies have been carried out with pattern recognition algorithms and feature extraction methods.
For example.
Blind Image Quality Index (BIQI) .
Perception based Image Quality Evaluator (PIQUE) .
, and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) .
have been developed in recent years.
One approach utilized in IQAs is the Journal homepage: http://section.
com/index.
php/IJEEI/index A ISSN: 2089-3272 incorporation of features into the assessment architecture .
This approach is supported by existing literature highlighting the significant impact of local or structural features on the determination of image quality .
, .
For clarity, the working principles and the NR-IQA categories of some methods used in this study are summarized in Table 1.
Table 1.
Working Principles of NR-IQA Metrics Included in the Evaluation Metric Category Working Principle BLIINDS-II Transform Domain-based Uses statistical features of DCT coefficients.
Distortions alter DCT statistics, and these changes are used to predict
BRISQUE
Traditional ML (NSS-base.
Extracts Natural Scene Statistics (NSS) features in the spatial domain.
Quality scores are predicted using Support Vector Regression (SVR).
BIQI
Early NSS-based Divides the image into sub-bands, extracts NSS features from each band, and maps them to quality scores via DIIVINE Traditional ML (NSS-base.
Extracts NSS-based features, classifies the distortion type first, and then predicts quality using a regression model trained for that distortion category.
NIQE
Traditional ML (NSS-base.
A completely opinion-unaware model.
Learns a multivariate Gaussian model of NSS features from natural images and predicts quality without training on human opinion scores.
IL-NIQE
Traditional ML (NSS-base.
An improved version of NIQE with an extended set of NSS features and multiple image categories for better In the presented study, the deep convolutional neural network is used.
By employing the residual learning method, the proposed approach effectively estimates and removes various types of noise in the image, including gaussian, sharpen, salt and pepper, poisson, and JPEG compression .
The image obtained as a result of denoising is accepted as a reference image.
Thus, the quality of the image can be determined by using the image and its estimated version.
In addition, during the comparison of two images for quality determination.
YUV color space is used in order to be more compatible with the HVS.
The novelty of the proposed BIQM lies in the integration of a DnCNN-based denoising stage with HVS-driven feature weighting in the YUV domain.
Unlike existing NR-IQA metrics that rely solely on NSSbased or handcrafted features, proposed BIQM first reconstructs a pseudo-reference image through residual learning, and then incorporates the biological sensitivity of the human eye in the final quality estimation.
This dual design bridges the gap between deep learning-based denoising and perceptual modeling, providing a more reliable no-reference image quality metric.
The second section of the paper outlines the key components of the proposed blind quality metric, including Deep convolutional neural networks (CNN).
HVS, and SSIM.
The third section provides a detailed examination of how the quality metric operates.
In the fourth section, experimental results are presented, along with comparisons to other methods.
The paper concludes with the final section, which offers conclusive findings and perspectives.
FUNDAMENTAL THEMES
Human Vision System (HVS) Due to certain characteristics of the human visual system (HVS), measuring image quality using the RGB color space may hinder the attainment of accurate results.
Figure 1 illustrates the two types of photoreceptors in the human eye, known as rods and cones, which are responsible for detecting light and color, respectively.
IJEEI.
Vol.
No.
September 2025: 585 Ae 600
IJEEI
ISSN: 2089-3272
Figure 1.
View of the human eye, rod, and cone cells The retina consists of approximately 120 million rod cells sensitive to luminance, while the number of color-sensitive cone cells is around 7 million.
This distribution indicates that approximately 95% of the total cells in the retina are primarily sensitive to luminance, with the remaining 5% dedicated to color perception .
Consequently, the impact of luminance on visual perception is more pronounced than that of By conducting quality measurements in the YUV color space, which allows for the separation of light and color information, the measurement results can be aligned with the characteristics of the HVS .
Deep CNN In the proposed BIQM framework, the denoiser is implemented using the DnCNN (Denoising Convolutional Neural Networ.
architecture introduced by Zhang et al.
In this model.
DnCNN employs a deep convolutional structure with 17Ae20 layers, which makes it capable of capturing more complex statistical dependencies in image structures (Figure .
The network design is inspired by the VGG (Visual Geometry Grou.
architecture and is specifically tailored for image denoising tasks .
Figure 2.
Residue Learning by using DnCNN The input detoriated image is defined as:
yc=ycu yc where yc is the distorted image, ycu is the original clean image, and yc denotes the additive noise.
Traditional discriminative denoising models aim to learn the mapping function ya.
= ycu, directly predicting the clean image.
Instead.
DnCNN adopts a residual learning strategy, where the network is trained to approximate the residual mapping (Figure .
OO yc yc = yc Oe ycI.
Blind Image Quality Metric for Color Images Based on Human Vision SystemA (Ali Erdem Altnba et a.
A ISSN: 2089-3272 Figure 3.
Denoising with DnCNN This formulation simplifies the optimization task, as learning the noise component is generally easier than directly predicting the clean image.
To further improve convergence speed and stability, batch normalization (BN) has been integrated into each convolutional block .
The combination of residual learning and batch normalization accelerates training and enhances denoising performance.
The objective function is defined as the mean squared error (MSE) between the estimated and the ground-truth residuals:
ycA ya.
= Oc Ou ycI.
Oe .
cycn Oe ycuycn ) Ou2ya ycn=1 where o represents the trainable parameters of the CNN, and ycA is the number of training samples.
This loss guides the network to minimize the discrepancy between the predicted noise and the actual noise, thereby indirectly enhancing the reconstruction quality of the clean image .
The success of DnCNN lies in its ability to generalize to various distortion levels, including blind Gaussian denoising, thanks to its deep residual learning design.
Although training deep CNNs from scratch can be computationally demanding .
, the use of pre-trained networks significantly reduces the training cost and makes such models practical for real-world applications .
Furthermore, utilizing fine-tuning in conjunction with pre-trained networks has been shown in some studies to lead to more favorable outcomes .
While DnCNN itself is not the primary contribution of this work, it plays a critical role as a preprocessing stage.
By producing denoised images, it ensures that the proposed BIQM can extract human visual system (HVS)-based features more reliably.
Consequently, the originality of this study is not in developing a new denoiser, but in integrating the denoised outputs into a novel HVS-driven blind image quality metric.
Peak Signal to Noise Ratio PSNR is a ratio obtained by comparing two numerical sequences.
During PSNR calculation, the Mean Squared Error (MSE) is first calculated (Eq.
MSE has a long history predating PSNR and has been widely used as a quality metric in signal processing for many years.
Originally used for signal fidelity.
MSE has also been applied to digital signals with the advent of the computer age .
The calculation of MSE involves determining the mean squared differences between two images of size M y N, as defined by:
ycAycIya = ycA Oc Oc.
cuycnyc Oe ycycnyc ) , ycAycA ycn=1 yc=1 x and y can be two different images.
The farther these two images are numerically from each other, the MSE will be higher.
Based on this, the PSNR can be defined as:
ycEycIycAycI.
cu, y.
= 10ycoycuyci10 ( ycAycIya IJEEI.
Vol.
No.
September 2025: 585 Ae 600
IJEEI
ISSN: 2089-3272
The reason 255 is used here is that the bit depth of the image is assumed to be 8, corresponding to 8 bits per pixel.
Moreover, it is important to note that higher MSE values result in lower PSNR values .
However.
PSNR has been subject to scrutiny and ongoing research due to its limited similarity to the human visual system.
Consequently, considerable efforts in quality metric research in recent decades have been focused on addressing this limitation .
Structural Similarity Over the past decades, researchers have explored alternatives to the purely numerical comparison offered by MSE to better align with the characteristics of the human visual system (HVS).
In addition to the transmitter, channel, and receiver variables associated with numerical comparison, the HVS, as described in Section 2.
2, incorporates visual perception phenomena studied in psychology.
Notably.
Bovik et al.
conducted significant research on this aspect, resulting in the development of SSIM .
The SSIM can be expressed in the following formula:
yu yu yu ycIycIyaycA.
cu, y.
= .
cu, y.
) .
cu, y.
) .
cu, y.
) , .
In this expression, l, c, and s correspond to luminance, contrast, and structural components, respectively.
When these statements are examined one by one, the luminance is expressed as:
cu, y.
= 2yuNycu yuNyc ya1 yuNycu2 yuNyc2 ya1 yuNycu and yuNyc are the mean of the image x and y, respectively.
The constant ya1 is added to avoid instability when the 2yuNycu yuNyc expression is too close to zero.
The contrast in the second component of Eq.
is expressed as:
cu, y.
= 2yuaycu yuayc ya2 yuaycu2 yuayc2 ya2 Covariances of x and y images are specified as yuaycu and yuayc .
Finally, the structural component is expressed as:
cu, y.
= yuaycuyc ya3 yuaycu yuayc ya3 As it draws attention, the numerator of the contrast element and the denominator of the structural component are shared by the yuaycu yuayc expression.
Therefore, if yu = yu = yu = 1 and ya3 = ya2 /2 are considered to simplify Eq.
, the result is .
ycIycIyaycA.
cu, y.
= .
yuNycu yuNyc ya1 ).
yuaycuyc ya2 ) .
uNycu2 yuNyc2 ya1 ).
uaycu2 yuayc2 ya2 ) .
SSIM gained prominence upon its initial proposal due to several advantages it offered over PSNR.
However, it is important to note that SSIM, as a full-reference image quality metric, requires two images for It assesses the structural similarity between the two images to determine the quality relationship between them.
THE PROPOSED METRIC: BIQM
The proposed BIQM comprises four distinct phases, which are outlined below and visualized in Figure 4:
A Phase 1: Image which quality is to be measured is transformed into YUV space.
A Phase 2: CNN is determined according to the file type (.
jpeg or .
of the image.
A Phase 3: Denoising is performed by using CNN specified for each color channel.
The obtained denoised color channels .
cUycc , ycOycc , ycOycc ) are assumed as reference image color channels.
A Phase 4: Pre-quality results are calculated by comparing the original Y.
V channels and the ycUycc , ycOycc , ycOycc channels via the SSIM.
Considering the biological properties of the HVS, the pre-quality results are weighted and the final BIQM quality result is obtained.
Blind Image Quality Metric for Color Images Based on Human Vision SystemA (Ali Erdem Altnba et a.
A ISSN: 2089-3272 Figure 4.
Flowchart of the proposed BIQM Phase 1: YUV Transformation Process In this phase, the image whose quality is to be measured is transformed into YUV space.
The decision to transform the image into the YUV color space is based on the understanding that the Y channel, which represents the luminance information, holds greater importance in terms of human visual perception .
The equation used to separate .
r make independen.
luminance information from color information is expressed as:
ycU = 0.
257ycI 0.
098yaA 16 ycO = Oe 0.
148ycI Oe 0.
439yaA 128 ycO = 0.
439ycI Oe 0.
386ya Oe 0.
071yaA 128 The process of returning from YUV space to RGB space involves utilizing specific transformation These formulas, which are given in .
, facilitate the conversion of the YUV representation back to the RGB color space:
ya = 1.
cU Oe .
cO Oe .
R = 1.
cU Oe .
Oe 0.
cO Oe .
Oe 0.
cO Oe .
B = 1.
cU Oe .
cO Oe .
The constants used in Eq.
and Eq.
correspond to the standard coefficients of the RGB to YUV color space conversion as defined in widely adopted video and image processing standards ITU-R BT.
In this study, the transformation was performed to prioritize the luminance information, as it plays a crucial role in visual perception.
The purpose of this conversion was to ensure that the luminance channel IJEEI.
Vol.
No.
September 2025: 585 Ae 600
IJEEI
ISSN: 2089-3272
receives appropriate weighting compared to the color channels, taking into account its greater significance in determining image quality.
Phase 2: Choosing Network It has been shown in previous studies that the blocking effect caused by JPEG should be examined with a separate concept considering the denoising process .
The blocking effect because of compression, especially at low bit rates, is sometimes examined separately .
Although most of the commonly used types of generated noise can be classified on a statistical basis, the blocking effect, a kind of compression problem, must be considered from a different perspective .
Because it is a noise caused by the quantification of DCT coefficients during compression .
From this point of view, two different pretrained CNNs have been used in this study.
Then, separate pre-trained CNNs would be used for JPEG and other noise types.
The use of pre-trained CNN has a direct effect on the result in most cases.
The use of a properly trained network is often of great importance in structures that require forecasting.
In addition to this, the use of multiple pre-trained CNNs is an issue that has been investigated in recent years .
Because training a CNN is a challenging process in terms of time and calculation costs.
In order to overcome this problem, the use of the GPU, which is very successful in matrix operations, is being examined with the parallel programming concept, and research on performance analysis continues .
The main purpose of using pre-trained CNN in this study is to avoid training cost.
The training and test data of the CNN used are described below.
Training and Testing the CNN The pre-trained network used in the proposed model had been trained using the Berkeley segmentation dataset .
For the training, 400 images of 180y180 size were used.
It has been noted that using more images for training did not provide a worthwhile improvement to the training cost.
In that study, three noise levels had been used to train DnCNN for Gaussian denoising with known noise levels: E = 15, 25 and 50.
The patch size as 40 y 40, and crop 128 y 1,600 patches were set to train the model.
In addition, to train a single DnCNN model for blind Gaussian denoising.
Zhang et al.
set the range of the noise levels as E OO .
The patch size as 50y50 and the crop 128y3,000 patches were set to train the model .
Two different datasets were used for test images in the BIQM.
The first is the Berkeley segmentation dataset (BSD.
which is depicted above, the second is the 12 images which are detailed in the study of Zhang et al.
In addition, test images are not included in the training dataset.
As mentioned above, a separate model is trained for JPEG.
To generate the JPEG deblocking input, the image is compressed with a quality factor ranging from 5 to 99 using the MATLAB JPEG encoder.
128y8000 image patch .
ize 50y.
pairs are generated for training.
Operations on the patch pairs based on rotation/flip are used during mini-batch learning.
Phase 3: Denoising with Deep CNN The more successful this denoising process, which is used to develop the intended quality metric, the more successful it will be to predict the quality.
For this reason, the denoising method used is extremely The method proposed by Zhang et al.
has proven to be more successful than previous studies .
In addition, the feedforward denoising CNN developed by them gives a fast result if it is operated with a pre-trained network.
Even if it is not pre-trained, the developed network is fully compatible with GPU parallel programming.
The utilization of pre-trained sets in MATLAB enables efficient and effective denoising processes.
The selection of the denoising level is predefined and not customizable by the user.
These functions are relatively new and do not offer input parameters for fine-tuning denoising applications according to desired To enhance the quality prediction performance, one can consider increasing the number of denoising iterations or retraining the model using transfer learning techniques.
While there are multiple methods to enhance denoising performance, it is important to note that the primary objective of this study is to propose an image quality metric rather than focusing on denoising techniques.
Phase 4: Weightening by the HVS In the fourth phase, an FR quality metric SSIM is used for quality calculation.
However, when using the SSIM, the denoised image is assumed as a reference image in the BIQM.
This measurement is made in the YUV color space to suit the HVS.
In addition, as mentioned in Section 2.
2, the ratio between light and colorsensitive cells in the human eye should also be reflected in the Y and UV channels .
For this reason, the quality of the Y channel in the measurement result is 95% effective compared to the human eye, while only 5% of the average of the U and V channels are effective .
As a result, the image quality to be determined using the presented method expressed as:
Blind Image Quality Metric for Color Images Based on Human Vision SystemA (Ali Erdem Altnba et a.
A ISSN: 2089-3272
yaAyaycEycA = ycIycIyaycA.
cU, ycUA) y 0.
95 [.
cIycIyaycA.
cO, ycOA) ycIycIyaycA.
cO, ycOA))/.
With this calculation, a color image, it is possible to determine the quality without a reference image.
EXPERIMENTAL RESULTS
During the evaluation of the experimental outcomes, it was observed that the proposed method achieved notable efficacy specifically in handling white noise.
To further validate the performance of the method, correlation calculations were conducted not only on the widely utilized LIVE .
CSIQ .
, and TID2013 .
databases but also on additional databases TID2008 .
and VCL@FER .
This broader selection of databases allowed for a more comprehensive assessment of the method's effectiveness across different noise levels.
These databases encompass a range of characteristics such as varying numbers of observers contributing to the determination of mean opinion scores (MOS), different quantities of reference and noisy images, diverse image sizes, and distinct levels of distortion given in Table 2.
These variations in the databases contribute to the diversity and richness of the data available for evaluation and enable a comprehensive analysis of the proposed method's performance across a wide range of conditions.
Table 2.
Characteristics of databases used in the study Database Reference Image
Distorted Image
Image Size
Distortion Level Observers LIVE
CSIQ
TID2013
TID2008
VCL@FER
512y512 To assess the degree of compatibility of the proposed study with the human visual system, the Spearman rank-ordered correlation coefficient (SROCC) and Pearson linear correlation coefficient (PLCC) coefficients were computed.
These coefficients measure the degree of correlation between the proposed method's output and human perceptual judgments.
A value closer to 1 for these coefficients indicates a stronger correlation between the proposed method and the human visual system.
Table 3 provides the ground truth values for the study at hand on the LIVE database .
The numeric suffixes "1" and "3" represent the number of times the algorithm was executed on the Y channel.
Additionally, the columns labeled as "JPEG" denote the usage of alternative pre-trained convolutional neural networks for JPEG compression.
Table 3.
Spearman rank-ordered correlation coefficient (SROCC) and Pearson linear correlation coefficient (PLCC) comparison of proposed BIQM on individual distortions types on the LIVE database.
SROCC-1
SROCC-3
SROCC-1 JPEG
SROCC-3 JPEG
PLCC-1 JPEG
PLCC-3 JPEG
JP2K
PLCC-1 PLCC-3
JPEG
White Noise Gaussian Blur Fast Fading Average Upon initial examination, the table illustrates the efficacy of the denoising method for both white noise and compression noise.
However, the observed correlation for Gaussian blur and fast fading is comparatively lower due to the inherent blurring effect of the denoising process, particularly in repetitive However, it is evident that using a CNN specifically trained for JPEG compression yields improved results for images that have undergone compression, such as those in the JP2k and JPEG formats.
The experimental studies have demonstrated the notable success of the proposed method, particularly in white noise.
To thoroughly investigate this aspect, an in-depth analysis was conducted on the state-of-the-art LIVE.
CSIQ, and TID2013 databases, focusing specifically on white noise.
In order to ensure a fair evaluation, fifteen different IQAs were assessed, and the SROCC results are given in Table 4.
It is noted that the proposed BIQM produces scores ranging from 0 to 1, where higher values indicate superior perceived quality and lower values correspond to more significant distortions.
IJEEI.
Vol.
No.
September 2025: 585 Ae 600
IJEEI
ISSN: 2089-3272
Table 4.
Comparative Analysis of SROCC Values for IQAs Across Different Databases IQA METHODS
LIVE
CSIQ
TID2013
DIIVINE .
BLIINDS-II .
NIQE .
IL-NIQE .
CORNIA .
BRISQUE .
GMLOG .
NFERM .
FRIQUEE .
BJLC .
HOSA .
BIECON .
WaDIQAm .
DB-CNN .
Proposed BIQM Table 5 presents the SROCC results obtained from various IQAs for white noise in the TID2008 and VCL@FER databases.
These databases, although less commonly used nowadays, contain a variety of image characteristics and noise types that can provide valuable insights into the performance of IQAs.
The performance of the proposed BIQM was compared against 14 cutting-edge NR-IQA methods across a range of benchmark datasets.
As demonstrated in Table 4.
BIQM exhibits robust performance on the LIVE.
CSIQ, and TID2013 databases, achieving SROCC values of 0.
981, 0.
951, and 0.
919, respectively.
The values obtained are consistently competitive with, or in some cases superior to, leading NR metrics such as BRISQUE.
BLIINDS-II.
NIQE, and CORNIA.
It is noteworthy that BIQM demonstrates a high degree of robustness for white noise distortions, aligning closely with the human visual system (HVS) perception.
Furthermore, the results presented in Table 5 offer additional verification of the proposed approach's generalisation ability, utilising the TID2008 and VCL@FER datasets.
In this study, the BIQM model demonstrated its capability to achieve correlation values of 0.
869 (TID2.
882 (VCL@FER), which exceed the performance of conventional NR metrics such as BIQI and NIQE.
Furthermore, the BIQM model exhibited a comparable or superior performance to BRISQUE in these datasets.
These comparisons demonstrate that the proposed BIQM not only competes with but also surpasses several well-established NR-IQA models across diverse datasets.
The consistent results observed across five distinct benchmark databases underscore the robustness, generalisation, and practical applicability of BIQM.
This comprehensive evaluation directly addresses the concern regarding performance comparison with other methods and strengthens the claim that BIQM offers a novel and effective solution for blind image quality In addition to the ground truth evaluation, peer-based PSNR tests were conducted to compare the image quality.
The test set consisted of images with various types of commonly used noise, including Gaussian.
Poisson, salt & pepper, sharpen, and JPEG.
While the PSNR metric described in Section 2.
4 failed to sort the images based on their quality, the presented metric allowed for accurate ranking of the images from the highest to the lowest quality.
To enhance the visibility of noise effects, selected parts of the images were enlarged during the comparison.
Table 5.
Comparative Analysis of SROCC Values for IQAs on TID2008 and VCL@FER Databases.
BLIINDS-II .
DIIVINE .
BRISQUE .
NIQE .
BIQI .
QAC .
Proposed BIQM
TID2008
VCL@FER
Blind Image Quality Metric for Color Images Based on Human Vision SystemA (Ali Erdem Altnba et a.
A ISSN: 2089-3272 As stated above, according to the PSNR, which is an FR metric, all images in Fig.
5 have the same PSNR value .
29 dB).
However, this is not a suitable situation for the HVS, and quality difference between images is perceived.
In addition, the original image may not be available, so the use of the NR metrics is It is also shown in Table 5 that other NR quality metrics mentioned in the introduction section do not give satisfactory results.
According to the BIQI, the PIQUE, and the BRISQUE metrics used in comparisons, when the image quality result approaches 0, it is assumed that the visual quality of the image increases.
On the contrary, when the BIQM result approaches 0, the visual quality of the image decreases.
Figure 5.
Comparison of AuMonarchAy images with different types of distortions, all with PSNR = 28.
29 dB:
Original image.
Sharpen image.
Salt & Pepper image.
Poisson image.
JPEG image.
Gaussian image.
Figure 5 and Table 6 show that the PIQUE and the BRISQUE metrics exhibit limitations in accurately determining the impact of salt & pepper.
JPEG, and gaussian noises, as well as differentiating between salt & pepper and Gaussian noise.
Despite the fact that the HVS perceives the Gaussian noisy image as the worst in terms of visual quality, the these metrics suggest that it has the best image quality among the different types of noise.
Table 6.
Performance comparisons of NR image quality metrics for Monarch image.
BIQI
PIQUE
BRISQUE
Proposed BIQM
Sharpen
S&P
Poisson
JPEG
Gaussian Another example is presented in Figure 6 with images prepared in the same PSNR value .
30 dB).
Comparison results are given in Table 7.
Table 7.
Performance comparisons of NR image quality metrics for Parrot image.
BIQI
PIQUE
BRISQUE
Proposed BIQM
Sharpen
Poisson
S&P
JPEG
Gaussian IJEEI.
Vol.
No.
September 2025: 585 Ae 600
IJEEI
ISSN: 2089-3272
Figure 6.
Comparison of AuParrotAy images with different types of distortions, all with PSNR = 29.
30 dB: .
Original image.
Sharpen image.
Poisson image.
Salt & Pepper image.
JPEG image.
Gaussian image.
As shown in Table 7, while the BIQI struggled to distinguish between sharpen and salt & pepper noise, the BIQM yielded a successful result under HVS in this regard.
Similarly, the PIQUE failed to separate JPEG and Gaussian noises, and the BRISQUE failed to separate Poisson and salt & pepper noises.
Even if the quality measurement scenario is FR, the PSNR values are equal .
, 24.
15 dB), on the contrary, the BIQM can distinguish digital images by considering the HVS as seen in Figure 7.
Table 8 gives performance comparisons of NR methods and the BIQM method .
ll noisy images have the same PSNR .
15 dB) value.
Figure 7.
Comparison of AuWallAy images with different types of distortions, all with PSNR = 24.
15 dB: .
Original image.
Sharpen image.
Poisson image.
Salt & Pepper image.
JPEG image.
Gaussian image.
Table 8.
Performance comparisons of NR image quality metrics for Wall image.
Sharpen
Poisson
S&P
JPEG
Gaussian BIQI
PIQUE
BRISQUE
Proposed BIQM The results obtained from the BIQI and PIQUE methods indicate that the quality values of sharpened and Gaussian images are similar, while the quality values of sharpened and Poisson images are close according to the BRISQUE method.
However, these results do not align with the HVS.
In contrast, visual examination reveals that the Gaussian and sharpened images exhibit significant differences.
The BIQM metric supports this perceptually noticeable result with its measurement outcomes.
Blind Image Quality Metric for Color Images Based on Human Vision SystemA (Ali Erdem Altnba et a.
A ISSN: 2089-3272 Table 9 presents the comparative performance of various NR image quality metrics, including the BIQM.
BIQI.
PIQUE, and BRISQUE, on different noisy images.
The noisy variations within each image have the same PSNR values, namely Flowers .
87 dB).
House .
22 dB), and Whitewater .
73 dB).
While sharpened noise produces less quality deterioration at the same PSNR than other types of noise, some NR image quality metrics fail to accurately distinguish between them.
Similarly, other NR image quality metrics exhibit instability in differentiating salt & pepper and gaussian noise.
However, visual examination of the noisy images in Fig.
Fig.
6, and Fig.
7, in the context of the HVS, shows that salt & pepper and gaussian images differ significantly in quality at the same PSNR value, and the BIQM confirms these results.
The corresponding images used in the assessments of Table 9 are shown in Figure 8.
Table 9.
Performance comparison of NR image quality metrics for different noisy images.
I1.
I2,
and I3 as Flowers.
House, and Whitewater respectively.
BIQI
PIQUE
BRISQUE
Proposed BIQM
BIQI
PIQUE
BRISQUE
Proposed BIQM
BIQI
PIQUE
BRISQUE
Proposed BIQM
Sharpen
Poisson
S&P
Gaussian JPEG Figure 8.
Images compared in Table 8.
I1.
I2, and I3 are Flowers.
House, and Whitewater, respectively.
A survey was conducted to understand the effect of the presented method and measure its compatibility with the HVS.
In this survey conducted with 100 participants, random participants were asked to rank 3 different images given in Figure 9 according to their quality.
Figure 9.
Images on which the survey was conducted.
Lena.
Peppers.
Baboon.
IJEEI.
Vol.
No.
September 2025: 585 Ae 600
IJEEI
ISSN: 2089-3272
The sorting of these three images, each of which added a different amount of noise, was made with a strict distinction among the participants.
Based on the survey results presented in Figure 10, a clear majority of the participants ranked the image in Figure 9.
as the worst quality, followed by .
and then .
as the best quality.
Lena .
Peppers Worst Better .
Baboon Best Figure 10.
Result of the survey.
According to the evaluations of the participants, a majority of respondents .
out of .
rated Lena as the image with the lowest perceived quality.
Similarly, 46 participants considered Peppers to be the median-quality image, while 56 participants ranked Baboon as the image with the highest perceived quality.
These findings provide insights into the subjective assessment of the images in Figure 9.
Additionally, the quality measurements of these images using NR-IQAs are presented in Table 10.
Table 10.
NR-IQA results of survey images.
Images BIQI
PIQUE
BRISQUE
Proposed BIQM
Lena
Peppers Baboon
IQA
When considering the collective findings from Table 9 and Figure 10 it can be observed that the BIQM demonstrates strong compatibility with the human visual system (HVS).
Conversely, the results obtained from the application of the BIQI.
PIQUE, and BRISQUE metrics to the Peppers image do not align well with the HVS.
Furthermore, the survey conducted indicates that utilizing the proposed BIQM as a noreference quality metric is more preferable in terms of user satisfaction and its alignment with visual
CONCLUSION
This study proposes a new deep CNN-based NR image quality metric (BIQM) consisting of four phases, including RGB to YUV transform.
CNN determination considering the file type, denoising with deep-CNN, and quality value calculation.
The experimental results demonstrate the effectiveness of the presented method, particularly for white noise.
The SROCC values indicate that BIQM provides more consistent results in terms of HVS compared to its counterparts, particularly for images corrupted with white Moreover, the survey results suggest that BIQM could potentially enhance user satisfaction when used for image quality measurement.
Another important finding of this survey, consistent with the results of some studies in recent years, is the challenge of distinguishing between low-level and high-level noise images .
, .
For future research, investigating the relationship between the amount of noise and the human visual system presents an area of growing interest.
Furthermore, with the advancements in CNN models, developing new high-accuracy metrics that do not require human scoring is becoming an increasingly important research area.
On the other hand, when performance evaluation is made, using pre-trained CNN ensures that it operates in a reasonable time even for low-medium level hardware.
Performance improvement is possible with the use of advanced GPUs and CPUs.
In this way, a fit-for-purpose deep CNN can be trained using a larger and more suitable data set.
Finally, the method of working in YUV space and weighting these channels will go further with the development of the concepts of vision and visual perception for future works.
Blind Image Quality Metric for Color Images Based on Human Vision SystemA (Ali Erdem Altnba et a.
A ISSN: 2089-3272 In brief, the main contribution of this work is the integration of DnCNN residual learning with HVSbased channel weighting.
This combination improves the robustness of BIQM against various distortions.
Unlike existing NR-IQA models, which usually rely only on handcrafted features or lack perceptual weighting, our BIQM bridges deep denoising and perceptual modeling.
As a result, it provides a more perceptually consistent and generalizable solution for blind image quality assessment.
The computational complexity of the presented algorithm has been tested only when MATLAB R2022b is running.
Technicial Specifications.
IntelA Alder Lake CoreE i5-12500H 12C/16T.
18MB L3.
E-CORE
Max 3.
30GHZ E-CORE Max 4.
5GHZ.
10nm, nVIDIAA GeForceA RTX3050 TI Max-Performance 4GB GDDR6 128-Bit DX12, 32GB .
x32GB) DDR4 1.
2V 3200MHz SODIMM, 500GB PCIe M.
2 2280 3.
x4 (R: 2050 MB/s - W: 940 MB/.
Under these conditions, the presented work is completed in an average of 252 seconds.
With the parallel pool, this time is reduced to 1.
103 seconds.
REFERENCES