Analysis of Weiner Filter Approximation Value Based On Performance of Metrics of Image Restoration

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Proc.

of the 2nd International Conference on Electrical, Communication and Computer Engineering (ICECCE)
12-13 June 2020, Istanbul, Turkey

Analysis of Weiner Filter Approximation Value


Based on Performance of Metrics of Image
Restoration
Joseph Domingo Davood Pour Yousefian Barfeh
Anna Liza Ramos
Technological Institute of the Batangas State University
Institute of Computer Studies
Philippines Batangas City, Philippines
Saint Michael’s College of Laguna
Manila, Philippines davoodpb@gmail.com
Platero, Bina City, Philippines
Domingo.joseph @gmail.com
AnnaLiza.Ramos@smcl.edu.ph

Abstract—Image restoration is used to recover the image the frequency domain to lessen the computation effort [7]
quality by reducing or eliminating the noise to go back and it applied a cost function for the filter together with the
to its original image. This study aims to investigate the Radial Basic function to restore the image [8].
performance of the Wiener Filter approximation value On the other hand, it is also used in different context like in
based on Mean Squared Error (MSE), Structural image transmission where the Wiener filter and wavelet
Similarly Index Image (SSIM), Root Mean Squared implementation is applied during compression to eliminate
Error (RMSE) and Peak Signal Noise Ratio (PSNR) irrelevant noise [9]. The application of wavelet transform
using the three (3) sample images in different dimensions aided in the analysis of the background noise and original
and image quality with the integration of five (5) voice of the speech frequencies [10]. It is being used to
different Gaussian noise and K-filter approximation compute the average spectral values based on the recursive
values. Based on experiment result, the Gaussian Noise noise estimation techniques through smoothing the
value of 10 marked a good performance based on the parameters [11] and incorporating equalization technique for
MSE for all the sample images however for the RMSE it possible interferences in digital communication system.
performs very well in Image2 and SSIM for Image3 and These various applications of Wiener filter lead the study to
the PSNR is evident in Image2. Indeed, the application investigate the performance of the filter, specifically to
of the said filter depends on the quality of the given provide relevant insights on how the filter performs to
image basis for the application of suitable measurements different types of images which can be very useful in
to achieve optimal results. biomedical field and the like in order to decide what
performance metric is applicable in achieving accurate
Keywords—image processing, image restoration, wiener filter assessment of the images.
algorithm

I. INTRODUCTION II. METHODOLOGY


Wiener filter served as model filtering technique in A. Datasets Pre-processing
improving the resolution and the signal-to-noise (SNR) ratio
The study used three (3) sample images with different
of an image. Basically, the filter is used to estimate the
dimensions and quality composed of : (a) 742 x 960 captured
original image in a way that it is as close as possible from
in a selfie mode using smart phone with filter and good
the degraded version of the same image caused by the added lighting condition of a fluorescent light , (b) 722 x 960
white noise [1]. Several studies recorded that Wiener filter captured from a one meter distance using a DSLR camera
marked a good performance in reducing noise and very with professional photography lights and, (c) 1200 x 900
flexible and powerful as well in de-blurring images [2]. In captured in a distance using a smart phone in a dim light
fact, the said filter is used in high resolutions which applies environment.
the similarity domain [3], implements in a non-uniform
interpolation and iterative approach [4]. It also serves as a
M
baseline model in image diagnosing like in biomedical field
by applying an iterative system of the wavelet packet, partial
differential equation, non -local Euclidean means, and first
order local statistics [5]. Likewise, having this filter utilized,
it enhances the block matching and 3D filtering image
diagnosis where the maximization of structural similarity is
applied instead of having the mean squared error which GNoise GNoise GNoise GNoise GNoise
resulted to an improved diagnosis performance [6].
Moreover, the filter was also enhanced by implementing the
Local Gaussian Markov Random Field as an alternative for
Original Blurring Noise
Grayscale Image Image
978-1-7281-7116-6/20/$31.00 ©20XX IEEE

Authorized licensed use limited to: Technological Inst of the Philippines. Downloaded on March 28,2022 at 03:45:02 UTC from IEEE Xplore. Restrictions apply.
 
 
Sp(u, v) H *(u, v)
Fig.1 Image with Gaussian Noise F (u , v) =   G (u , v)
 Sp (u , v)  2 Sn (u , v )  
These images were converted from RGB to Grayscale  | H *(u , v) | + Sp (u , v)  
values. The gray images were processed with the application   
of Gaussian Noise values of 10, 20, 30, 40 and 50 and K- (2)
filter values of 5, 7, 13, 15 and 25 to further evaluate the
Wiener filter performance. The image frequency domain denoted by F(u,v) was
equal to 1 because equation 1 was the derivation of the
undegraded/original image power spectrum divided by itself
B. The Weiner Filter Image Restoration Process was cancelled to 1, multiplied by the degradation function,
and multiplied to the degraded function conjugate which
was then divided by the degraded function and square of
degraded function conjugate, plus the spectrum of the noise
over the spectrum of the undegraded image (3).
 
 
1 H (u, v) H *(u , v)
F (u , v) =   G (u , v)
 H (u , v)  2 Sn (u , v )  
 | H *(u , v) | + Sp (u , v)  
  

(3)
In this equation, 1 is multiplied by the squared value of
Fig.2 Weiner Filter Process the degraded function and degraded function conjugate
square, plus the result of the noise power spectrum and the
The original image was transformed by integrating the power spectrum of the undegraded image denoted by {K}
Gaussian noise values of 10, 20, 30, 40 and 50 to generate a which was used to assign values through trial and error until
degraded image. The degraded spatial image was then it finds the closest match of the undegraded image. If K = 0,
transformed into frequency image using statistical methods it means that there was no power spectrum (4).
by computing the mean value and the power spectra of the
noise. Then, the degraded image would be observed by  1 | H (u, v) |2 
minimizing the square error by computing the expected
F (u, v) =   G (u , v )
 H (u , v) | H *(u, v) | + K  
2
value of the original image versus the degraded image
through doing a trial and error process on the noise power (4)
spectrum and the undegraded image power spectrum to
achieve the closest match of the original image, thus attain This expression is the complex conjugate of the Fourier
the restoration. response of the degradation filter divided by the magnitude
square of the filter, plus the power spectrum of the noise
From the degraded image function G(u,v), the image divided by the power of the original image signal invert
frequency domain function F (u,v) was computed based on which was the inverse Fourier to get the estimation to
the degradation function conjugate – when it’s taken to the minimize the square error (5).
Fourier transform it was represented in real numbers and 1 | H (u, v) |2
imaginary values over the power spectrum of the original
image multiplied by the power spectrum of the degraded H (u, v) | H *(u , v) |2 + K 
image from the frequency domain over the square of the
degraded function plus the power spectrum of the original (5)
image into degraded image (1).
C. Mean Squared Error
 H *(u , v) SF (u, v)  Before the minimum square error would be computed,
F (u, v) =  2  G (u , v) the following assumptions should be considered:
 Sf (u , v) | H (u , v) | + Sf (u , v) 
(1) • Noise of the images are uncorrelated which means
that the noise function N(u,v) and degraded image
The image frequency denoted by F(u,v) is equal to the function G(u,v), are not correlated to the original
power spectrum of the image frequency coordinates into image function F(u,v).
complex conjugate divided by the power spectrum of the
original image and the complex conjugate square, plus the • Any one of them has a zero mean referring to
power spectrum of the noise divided by power spectrum of degradation function G(u,v) and Noise function
the undegraded image(2). N(u,v).

Authorized licensed use limited to: Technological Inst of the Philippines. Downloaded on March 28,2022 at 03:45:02 UTC from IEEE Xplore. Restrictions apply.
• Gray levels of the estimates are linear function of the
degraded image which transforms the degraded image K=5 K=7 K=13
G(x,y) to original image function f(x,y).
If the assumptions are true, then it would minimize the error
e2 = E{( f − F )}2 given in the frequency domain.
Expected value of the original image minus the degraded
image. The minimum square error was applied after the
degraded image function G(x,y) was computed with the K=15 K=25
purpose to make a Trial and Error until the desired closest
match from the original image has been met (6). Fig.3. Image1 Histogram

1 This shows that there was a transition of the signal


 [ A(a, b) − W (a, b)]
2
MSE = power spectra when the filter values were set.
M *N a b
Table 1. MSE Result of Image1
(6)
Gaussian Noise
D. Structural Similarity Index Image(SSIM) 10 20 30 40 50
Filter
The approach of the SSIM was to measure the group of Value 2406.61 3227.56 4445.34 5960.14 7790.79
pixels of the original image and degraded image with 2
partial indices. SL is the local luminance and SV is the local WF K =5 2458.95 2469.15 2539.23 2614.33 2754.86
covariance/ variance [12] (7). WF K =7 2456.60 2460.94 2529.61 2597.83 2735.65
WF K=13 2446.76 2452.10 2518.28 2578.70 2714.49
 2 μ μ   2σ σ   σ 
SSIM { f1 , f 2 } = S =  2 1 2 2  x  2 1 2 2  x  12  WF K=15 2445.61 2450.45 2516.81 2574.96 2711.51
 μ1 + μ 2   σ 1 + σ 2   σ 1σ 2  WF K=25
2441.89 2447.13 2512.43 2567.13 2702.90
 2 μ μ   2σ 
S =  2 1 2 2  x  2 12 2  = S L SV
 μ1 + μ2   σ 1 + σ 2 
(7)
The SSIM is applied in this study to determine the K- Results show that the Gaussian noise without the
approximation value performance. integration of the K filter value, the mean squared error
increases. Likewise, for the image with Gaussian noise and
E. Root Mean Square Error K filter value, the figure shows that when a Gaussian noise
value increases, the approximation k-value should be also
This measures the original images and degraded images’
set to a higher value to remove the noise, thus smoothing the
residual error value which measure the accuracy of the
image. In result, the mean squared error value decreases.
image closest to the original image (8).
Specifically, Image1 recorded a minimum mean squared
1 error for the K-filter 25 with a range variance of 0.693
RMSE =
N
 ( E
i j
ij − Oij ) 2 which means that the image has a lot of noise, considering
that the device used to capture the image have automatic
(8) filter features. In result, the performance of the filter was
F. Peak Signal Noise Ratio highly observed with K value of 13. Moreover, with regards
to the aesthetic visualization of the image, the maximum K-
This measures the power of the signal and the noise that filter value was 15 which means that the image was still
affects the quality of the image (9). visible on that value because the higher the value of the
 ( L − 1) 2   L −1  filter, the image becomes blurrier.
PSNR = 10 log10   = 20 log10  
 MSE   RMSE  Table 2. Root Mean Squared Error of Image1
(9)
Gaussian Noise
III. RESULTS & DISCUSSION 10 20 30 40 50
Filter 49.057 56.812 66.673 77.202 88.265
Value
K -5 49.588 49.691 50.391 51.130 52.487
K =7 49.564 49.608 50.295 50.969 52.303
K=13 49.465 49.519 50.182 50.781 52.101
K=15 49.453 49.502 50.168 50.744 52.072
Original Image Blurring Image Noise Image
K=25 49.415 49.468 50.124 50.667 51.989

Authorized licensed use limited to: Technological Inst of the Philippines. Downloaded on March 28,2022 at 03:45:02 UTC from IEEE Xplore. Restrictions apply.
This shows that the residual value of the error marked a This result shows a lot of similarity to Image1
good performance at K filter with a value of 25 and a range regardless of the image quality. When a Gaussian noise
variance of 0.35, 0.45, 0.53, 0.91 and 0.95 between K=5 and increases and the K filter increases, the mean squared error
K=25 from a Gaussian noise value of 10 to 50. decreases.
For this Image2, the minimum mean square error was
Table 3. Structural Similarity Index Result of Image1 recorded in a K filter value of 25 with a range variance of
0.803 which means that the noise was minimal because the
Gaussian Noise image was captured using a DLSR camera with no filter
10 20 30 40 50 mechanism set and thus, it resulted to having a good quality
Filter 0.37641 0.19715 0.12692 0.09165 0.07056 image. In result, the filter was highly observed at K-filter
Value value of 15. Moreover, for the aesthetic visualization of the
K=5 0.66735 0.59660 0.51885 0.45380 0.39408 image, the maximum K-filter value visibility of the image
was at K filter value of 13 because setting the higher value
K=7 0.66897 0.60136 0.52546 0.46115 0.40090 of the K-filter makes the image blurrier.
K=13 0.67092 0.60653 0.53320 0.46982 0.40933
Table 6. Root Mean Squared Error of Image2
K=15 0.67123 0.60749 0.53452 0.47130 0.41076
K=25 0.67198 0.60978 0.53764 0.47532 0.41438 Gaussian Noise

The result shows that the Gaussian noise value of 10 10 20 30 40 50


Filter
recorded the highest score similarity index of about 37% Value 26.568 39.054 52.584 66.057 79.191
without the application of the K-filter value and for the K=5 24.235 25.120 26.612 29.345 31.801
image with K-filter value, the highest similarity score was
K =7 24.200 25.045 26.511 29.098 31.412
67%, recorded at K-25 among other filter values. Therefore,
the higher the K-value, the closer it gets to the original K=13 24.155 24.960 26.390 28.819 30.947
image. K=15 24.149 24.950 26.368 28.778 30.880
K=25 24.137 24.921 26.310 28.666 30.709
Table 4. Peak Signal to Noise Ratio of Image1

Gaussian Noise This shows that the residual value of the error still
10 20 30 40 50 recorded at the K filter value of 25 resulted to a lowest score
Filter 14.317 13.042 11.652 10.378 9.215 with a range variance of 0.40, 3.21, 1.13, 2.13, 3.43 between
Value K 10 and 25 from a Gaussian noise value of 10 to 50.
K=5 14.223 14.205 14.084 13.957 13.730
Table 7. Similarity Index Result of Image2
K =7 14.227 14.220 14.100 13.985 13.760
K=13 14.245 14.235 14.120 14.017 13.794 Gaussian Noise
K=15 14.247 14.238 14.122 14.023 13.799 10 20 30 40 50
Filter
K=25 14.254 14.244 14.130 14.036 13.812 Value 0.381 0.174 0.102 0.069 0.049
K =5 0.793 0.692 0.585 0.491 0.419
These results show that with the higher value of the K =7 0.796 0.698 0.593 0.500 0.429
PSNR, the K filter value performed very well with a range K=13 0.798 0.705 0.603 0.512 0.440
variance of 0.21, 0.27,0.32,0.56 and 0.59 from a Gaussian K=15 0.799 0.706 0.605 0.514 0.442
noise value of 10 to 50. In fact, the Gaussian noise effect
was highly observed based on the aesthetic visibility of the K=25 0.800 0.709 0.609 0.518 0.447
image.
The results show that the Gaussian noise value of 10
Table 5. MSE Result of Image 2 recorded the highest score similarity index of about 38%
without the application of the K-filter value while an image
Gaussian Noise with the application K filter value recorded the highest
similarity score of 80% at K=25 among other filter value
10 20 30 40 50 which means that information of the image were close to the
Filter
Value 705.84 1525.19 2765.05 4363.56 6271.21 original image.
K=5 587.32 630.99 708.22 861.15 1011.29
K =7 585.66 627.25 702.86 846.67 986.72
K=13 583.45 623.01 696.45 830.53 957.74
K=15 583.18 622.52 695.28 828.16 953.58
K=25 582.60 621.08 692.20 821.76 943.01

Authorized licensed use limited to: Technological Inst of the Philippines. Downloaded on March 28,2022 at 03:45:02 UTC from IEEE Xplore. Restrictions apply.
Table 10. Root Mean Squared Error Result of the Image 3
Table 8. Peak Signal Noise Ratio of Image2
Gaussian Noise
Gaussian Noise 10 20 30 40 50
10 20 30 40 50 Filter 34.654 45.702 59.223 73.403 87.285
Filter 19.644 16.298 13.714 11.732 10.157 Value
Value K=5 38.020 38.712 39.822 40.849 42.566
K =5 20.442 20.131 19.629 18.780 18.082 K=7 38.152 38.663 39.718 40.718 42.383
K=7 20.454 20.156 19.662 18.854 18.189 K=13 38.121 38.580 39.595 40.564 42.166
K=13 20.471 20.186 19.702 18.937 18.318 K=15 38.111 38.565 39.575 40.541 42.129
K=15 20.473 20.189 19.709 18.950 18.337 K=25 38.088 38.533 39.522 40.482 42.039
K=25 20.477 20.199 19.729 18.983 18.386
This show that the residual value of the error still
This shows that the image with Gaussian Noise value of recorded for the K filter value of 25 which also marked the
10 got the highest signal ratio which means the filter was lowest score with a range variance of 0.17, 0.46, 0.75, 0.89,
more effective compared to other values of the noise. 0.90 between K 10 and 25 from a Gaussian noise value of
Specifically, the range variances were 0.17,0.37,1.07 and 10 to 50.
1.65 between K=5 and K=25 from a Gaussian noise value of
10 to 50. In fact, the Gaussian noise effect was highly Table 11. Structural Similarity Index Result of Image3
observed based on the aesthetic visibility of the image.
Gaussian Noise
Table 9. MSE Result of the Image 3
10 20 30 40 50
Filter
Value 0.381 0.174 0.102 0.069 0.049
Gaussian Noise
K -5 0.793 0.692 0.585 0.491 0.419
10 20 30 40 50
Filter K =7 0.796 0.698 0.593 0.500 0.429
Value 1200.88 2088.69 3507.36 5388.00 7618.65
K=13 0.798 0.705 0.603 0.512 0.440
K =5 1445.52 1498.62 1585.83 1668.67 1811.90
K=15 0.799 0.706 0.605 0.514 0.442
K =7 1455.55 1494.84 1577.53 1657.94 1796.29
K=25 0.800 0.709 0.609 0.518 0.447
K=13 1453.21 1488.39 1567.78 1645.47 1777.99
K=15 1452.47 1487.30 1566.20 1643.60 1774.85 The result shows that the Gaussian noise value of 10
K=25 1450.66 1484.77 1561.99 1638.82 1767.28 recorded the highest score similarity index of about 38%
while the image with the application K filter value, the
highest similarity score was 80% among other filter value
which means that information of the image were close to the
original image.

Table 12. Peak Signal Noise Ratio of Image3


The results show that when a Gaussian noise increases,
the suggested approximation k-value should be higher to Gaussian Noise
remove the noise to smoothen the image with a resulted 10 20 30 40 50
decrease on the squared error value. In this case, the Filter 17.336 14.932 12.681 10.817 9.312
minimum mean squared error was K-filter 25 with a range Value
variance of 0.354 which means that the image has a lot of K -5 16.531 16.374 16.128 15.907 15.549
noise because the image was captured by just using a smart
K =7 16.501 16.385 16.151 15.935 15.587
phone camera with dim light environment which resulted to
generate poor quality image. In result, the filter was highly K=13 16.508 16.404 16.178 15.968 15.631
observed at K value of 5. Moreover, with regards to the K=15 16.510 16.407 16.182 15.973 15.639
aesthetic visualization of the image, the maximum K-filter K=25
value was 7 which means that image was still visible on that 16.515 16.414 16.194 15.985 15.658
value because with a higher filter value, the image becomes
blurry. Among the results, Gaussian Noise 10 got the highest
signal ratio which means the filter is more effective
compared to other values of the noise with a range variance
of 0.09,0.24,0.41 0.49 and 0.69. In fact, it was highly
observed based on the aesthetic visibility of the image.

Authorized licensed use limited to: Technological Inst of the Philippines. Downloaded on March 28,2022 at 03:45:02 UTC from IEEE Xplore. Restrictions apply.
IV. CONCLUSION REFERENCES
Based on the results of the experiment using three (3)
sample images in different quality and dimension and with [1] I. Quinani, “Deblurring image and removing noise from medical
images for cancerous disease using weiner filter “. Internationa
application of different values of the Gaussian Noise and K Research Journal of Engineering and Technology , 2017
filter, the performance of the Wiener Filter shows consistent [2] J.M. Blackedget, “Digital signal processing mathematical and
result. In particular, the image with Gaussian Noise value of computation methods software development”, Woodhead Publishing
10 based on the four (4) performance metrics such as MSE, Series in Electronic and Optical , 2nd Edition, 2006
SSIM, RMSE and PSNR were applied in this study. [3] V.S. Chaudhary and M.Shandilya, “Channel equalization using
However, considering the mean squared error metrics, it weiner filter”, vol.612, no.2, p 607-612, 2013
shows consistent result for the three (3) sample images but [4] C. Cruz, R. Mehta, V.Katkovnik and K.O, Egiazarian, “Single image
super-resolution based on weiner filter in similarity domain “, IEEE
not in Root Mean Squared Error. It performed very well in Trans Image Process:vol 27, no.3 p 1376-1389,
Image2 as well as in the Peak Signal Ratio while the doi:10.1109/TIP.2017.2779265, 2016
Similarity Index marked the highest score in Image3. [5] S. Lahmiri, “An iterative denoising system based on weiner filter with
Therefore, the study concluded that the performance application to biomedical images,” Optics and Laser Technology,
vol.90 , p 128-132..
behavior of the filter will depend on the quality of the image
which greatly affects the measurements used in this study. [6] N. Husan and M, El-Sakka.”Improved BM3D image diagnoising
using SSIM optimize weiner filter “, EURAISP Journal on Image and
Moreover, with lesser Gaussian Noise value of the image, it Video Processing, 2018
makes a significant result on whatever type of image across [7] E. Baselice, G. Ferraoili, M. Ambrosanio, V. Pascazio and G.
the performance metrics used in this study. The insight of Schirinzi, Enchance weiner filter for ultrasound image restoration”,
this study can be very helpful and useful to those who would Computer Methods and Programs in Biomedicine. Vol. 153 p 71-8,
2017
like to use the filter in achieving desired and optimal results.
[8] A. Sultana, “Image restoration using combined adaptive weiner filter
For further improvement of the study, the researchers and radial basis function ANN with sub-block decomposition for
suggest to design a technique that will control or determine medica”, Applications, vol.4 no.3 p 389-395.
the maximum limit of the filter to be applied to specific type [9] S. Muthuselvi and A.S A, “An adaptive image enhancement using
of image. weiner filtering with compression and segmentation”, no. 1, p15-19.
[10] A.Mathematics. http://www.acadpubl.eu/hub/,” vol. 119, no. 16, pp.
ACKNOWLEDGMENT 731–743, 2018
The study would like to thank Dr. Melvin D. Ballera of [11] N. Upadhyay and R.K, Jaiswal, “Single channel speech enhancement
using weiner filtering with recursive noise estimation:, Procedia
Technological Institute of the Philippines – Manila for the Comput, Sci., vol. 84 , p 22-30, 2016
motivation to produce this scholarly work and building our [12] K. Larkin. “Structural similarity index SSIMplified” Is there a simpler
skills in research. concepts at the image of image quality measurement.

Authorized licensed use limited to: Technological Inst of the Philippines. Downloaded on March 28,2022 at 03:45:02 UTC from IEEE Xplore. Restrictions apply.

You might also like