P-ISSN 2586-9000

E-ISSN 2586-9027

Homepage : https://tci-thaijo.org/index.php/SciTechAsia Science & Technology Asia

Vol.27 No.2 April - June 2022 Page: [137-147]

Original research article

Reconstruction of an Improved High-

Resolution Image from Low-Resolution
Image Using Image Enlargement and
Enhancement Algorithm
Loveleen Kumar*, Manish Jain

Department of Computer Science and Engineering, JECRC University, Jaipur 303905, India

Received 18 September 2020; Received in revised form 27 January 2021

Accepted 19 April 2021; Available online 29 June 2022

ABSTRACT
The resolution of images in the medical field plays an important role in taking decisions
to identify diseases, but if the image resolution is low, then disease identification may go
wrong. For the correct decision, the image should be of high-resolution. High-resolution im-
ages can be taken using efficient hardware devices also, but then it is a costly proposition.
Thus, we propose a model to reconstruct a low-resolution image into a high-resolution image
using four image enlarging algorithms and an image enhancement algorithm. Initially, images
are enlarged using enlarging algorithms to generate four images, followed by an image en-
hancement algorithm to generate four enhanced quality images. Thereafter, one high-
resolution composite image is constructed using these four images. Experimental results were
compared with the results of existing image enlarging approaches. Significant improvements
in important image parameters like Peak Signal to Noise Ratio (PSNR), Signal to Noise Ratio
(SNR), Structural similarity (SSIM), and Mean Squared Error (MSE) could be achieved.
Keywords: Composite image; High-resolution image; Image enhancement algorithm; Image
enlargement algorithm

1. Introduction largement algorithms, the image size is in-

High-resolution images are required creased by increasing the number of pixels
in the medical field so that diseases can be that are present in the original image. If we
adequately diagnosed. Image enlarging is a are increasing the size of the image by twice
process that reconstructs the image in high- the actual size, then we need to add the
resolution from its low resolution. During same number of pixels already present in
this process of increasing the image size, the actual image. New pixels can be gener-
image quality is taken care of. In image en- ated using various algorithms. In our

*Corresponding author: [email protected] doi: 10.14456/scitechasia.2022.31

 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

work, we have used the Nearest Image enhancement is a process to

Neighbour Interpolation Algorithm, Bilinear improve the quality and content of the im-
Interpolation Algorithm, High-Quality Scale age. The iterative back-projection (IBP) al-
(Hqx) and Scale by rules (xBR) [1]. gorithm generates a super-resolution image
The Nearest Neighbour Interpolation using enhancing the pixel information.
Algorithm is the most straightforward algo- When this algorithm is applied to a single
rithm to enlarge the image with the help of image, blurring can be removed without
the neighbour of the pixels available in the increasing the sampling rate. This algorithm
actual image. It checks the value of the is suitable for both computer generated im-
nearest neighbour, ignoring the other pixels. ages and real-time images. In the IBP algo-
An image can be enlarged by 2x, 3x, and so rithm, pixels are reconstructed by compar-
on. If we are enlarging the image by 2x, ing the past values of the pixel and the best-
then it will increase the pixels in double; suited value will be assigned to enhance the
one row and column will be added just after image. The algorithm divides the whole
each row and column in the original image process into two parts. The first part builds
[1]. the imaging process so that the old values of
The Bilinear Interpolation Algorithm the pixel can be compared. The second part
is the updated version of neighbour interpo- assesses new values of the pixels through
lation to interpolate the pixel. This algo- the image enrollment approach [6-8].
rithm works in two directions instead of a The 4 enlarging algorithms described
single direction to get the pixel value. This above have certain limitations when applied
algorithm takes an average of the four near- independently.
est pixels to find the value of the unknown For example, the enlarged image ob-
pixel [2]. tained using the Nearest Neighbour Interpo-
High-Quality Scale (Hqx) comes un- lation Algorithm gives jaggies at boundaries
der the category of pixel art scaling algo- that increase the noise, affecting the quality
rithms that usually work better on a very of the output image [1]. Likewise, the Bilin-
low-resolution image. This technique gives ear Interpolation Algorithm produces imag-
undistorted and sharp images. Three filters es with blurring and edge halos. Due to this
are used to enlarge the image, defined as issue, the image produced is of low quality
hqx2x, hqx3x and hqx4x, which enlarges [2]. Similarly, Hqx and xBR are pixel-art
the image by 2, 3 and 4 factors, respective- enlarging algorithms that produce artifacts
ly. Single pixels can be enlarged into 2x2, in the output image [3-4].
3x3, or 4x4 pixel arrangements. The algo- In the proposed methodology, we
rithm checks similar or dissimilar neigh- have used the IBP algorithm to enhance the
bours by comparing the surrounding pixels images enlarged using the Nearest Neigh-
[3-4]. bour Interpolation Algorithm, Bilinear In-
Scale by rules (xBR) is another type terpolation Algorithm, High-Quality Scale
of pixel art scaling algorithm. It has six fil- (Hqx) and Scale by rules (xBR).
ters in the family: xBR, xBRZ, xBR-Hybrid, The use of the IBP algorithm over-
Super xBR, xBR+3D and Super xBR+3D. comes the aforesaid limitations of the en-
xBR can handle complex and zig-zagged larging algorithms and improves the quality
images and generates a smooth image. No- of the output image by enhancing while
blend xBR is more popular than another smoothing the boundaries, removing the
variant that is with a blend. No-blend xBR blur effect and artifacts [6-8]. Further com-
sometimes gives jaggy outputs, but it main- posite images are generated to improve the
tains image sharpness [3, 5]. overall quality of the output image received
from the IBP algorithm.

138
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

We have identified that high- using a revised regularization method to

resolution images play a crucial role in the resolve the ill-posed problem and enhance
identification of any object in the medical the edges also. We have used an ensemble
field. In this paper, we have proposed a technique to enhance the edges of the im-
model by which images can be reconstruct- age.
ed in high-resolution from low-resolution Tanaka et al. [13] proposed an algo-
with improvement to the quality of the im- rithm to reconstruct a high-resolution image
age. In our experiments, we have focused on using a MAP-based technique. The algo-
the reconstruction of MRI images from low- rithm optimizes cost function in the fre-
resolution images. This model has improved quency domain of the high-resolution im-
the quality of the image after converting it age. We use four different algorithms that
to high-resolution from low-resolution. collectively optimize the cost function of the
This paper has the following sections. enlarged image.
In section 2, the literature review is present- Li et al. [14] have proposed a method
ed. In section 3, the methodology is de- for image deblurring by relaxed initializa-
scribed. Results and Analysis are given in tion and pixel-wise updates of the iterative
section 4. Section 5 describes the conclusion method. Li et al. [15] proposed an improved
and future work. Finally, a list of references Iteration Back Projection method to con-
is provided. struct super-resolution images using a low-
resolution image and applied it to Advanced
2. Literature Review Land Observing Satellite (ALOS) imagery.
Zhu et al. [9] proposed a new tech- Irani et al. [16] proposed an approach to
nique for super-resolution image reconstruc- improve resolution by an image registration
tion that uses Surveying Adjustment in technique. They determined that while im-
which they used a sequence of low- proving the resolution the resulting image
resolution images as observations and then may be blurry. Purkait et al. [17] proposed a
equations for the observations are consid- regularization method to reduce the blurring
ered for the construction of a super- effect and also reduced the noise generated
resolution image. We have constructed four during low-resolution image formation
images using different algorithms that will while preserving edge information. We have
work as a sequence of enlarged images. used the Iterative Back Projection algorithm
Sundar et al. [10] proposed an algo- with improved parameters to enhance the
rithm to reconstruct super-resolution. They image quality of the medical image by re-
used the Discrete Wavelet Transform and duction of noise and blur.
combined multiple low-resolution images to
improve image quality. We have used itera- 3. Methodology
tive back-projection algorithms and con- We have used four enlarging algo-
structed a composite image to enhance the rithms, namely the Nearest Neighbour In-
overall quality of the image. terpolation Algorithm, Bilinear Interpolation
Yu et al. [11] proposed an algorithm Algorithm, High-Quality Scale (Hqx) and
named adaptive inverse hyperbolic tangent Scale by rules (xBR) to enlarge the input
(AIHT) to consider the image feature during low-resolution image. Then the generated
the enlargement of the image. They en- low-resolution images get enhanced and
hanced the image before enlarging. We have finally, these enhanced images are grouped
enhanced the image after enlarging so that together to generate enhanced high-
their features should also be enhanced. resolution images. Fig. 1 depicts the overall
Liu et al. [12] proposed a methodolo- process of reconstruction of a high-
gy to construct a super-resolution image resolution image.

139
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

The whole process was done in the Step 2: Enhancement of enlarged images
following three phases: using back-projection algorithm
Step 1: Enlarging the low-resolution images Step 3: Finally, enhanced images are
using four different enlargement algorithms. grouped to generate a composite enhanced
final high-resolution image.

Fig. 1. The proposed model to reconstruct a high-resolution image from a low-resolution image.

For the purpose of reconstruction of The Bilinear Interpolation algorithm

enhanced high-resolution image (512x512), interpolates the pixel value as per the near-
we have taken low-resolution (256x256) est pixel. So we have enlarged the same
Magnetic Resonance Imaging (MRI) images low-resolution image to generate one more
of the human brain [18-19]. Four enlarging image using a bilinear interpolation algo-
algorithms, namely the Nearest Neighbour rithm that interpolates the pixel value by
Interpolation algorithm, Bilinear Interpola- averaging the most adjacent pixels, which
tion algorithm, High-Quality scale (Hqx), also preserves the edge.
Scale by rules (xBR) are used to enlarge the Two other algorithms, Hqx and xBR,
image and images I1, I2, I3 and I4 are gen- are used to enlarge the low-resolution im-
erated using these enlargement algorithms, age. Both algorithms are a type of pixel art
respectively. scaling algorithms. While enlargement im-
All the enlarging algorithms have ages may produce distortion, Hqx handles it
their significance. The Nearest Neighbour and gives a sharp image. Hqx2x is used for
Interpolation algorithm is the most straight- the enlargement. The xBR algorithm is used
forward algorithm and takes very little time to remove the zig-zagged effect of the im-
to generate an enlarged image. As the num- age.
ber of pixels doubles, it needs to interpolate These four image enlargement algo-
the same number of pixels present in the rithms convert a 256x256 size image into a
image and the nearest Neighbour property is 512x512 size image. I1, I2, I3 and I4 are the
used to find the value of the pixels. 512x512 images generated by the image
enlargement algorithms. Image enlargement
140
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

algorithms generate images by increasing

the number of pixels. We have increased
twice the number of pixels to enlarge the
image.
In the next step, we reduce the noise
and the blur effect produced during the en-
largement of the image.
Iterative back-projection (IBP) was
applied to generate enhanced images from
images I1, I2, I3 and I4. The image en-
hancement algorithm was applied to the four
images received from the previous step to
generate enhanced images I5, I6, I7 and I8
with less noise and blurring. The function of
IBP in MATLAB is shown in Listing 1 and
flow chart is shown in Fig. 2

List 1. Basic function of IBP to reduce the

noise and blur.
numImages = 4;
blurSigma = 1;
[ images offsets croppedOriginal ] = SynthDataset(im,
numImages, blurSigma);
%% for enhanced image
[ lhs rhs ] = SREquations(images, offsets, blurSig-
ma);
K = sparse(1 : size(lhs, 2), 1 : size(lhs, 2), sum(lhs,
1));
initialGuess = K \ lhs' * rhs; % average image pro- Fig. 2. Flow Chart of the basic function of IBP
duced from the LR images. to reduce the noise and blur.

Next, two images are grouped togeth- List 2. MATLAB code to construct compo-
er to construct a composite image. This site image
grouping of images was done using the im- image-sr-nn = imread(sr_nearest.png');
fuse function of MATLAB. We are group- image-sr-bi = imread('sr_bi_output.png');
ing I5 with I6 and I7 with I8 to generate two image-sr-xbr = imread('sr_xbr.png');
composite images I9 and I10, respectively. image-sr-hqx = imread('sr_hqx.png');
Finally, these two images are grouped to- c=imfuse(image-sr-nn, image-sr-bi);
gether to construct a final enhanced high- d=imfuse(image-sr-xbr, image-sr-hqx);
imwrite(composite-1,'bilinear_plus_nearest.png')
resolution image of 512 x 512. MATLAB
imwrite(composite-2,'xbr_plus_hqx.png')
code for composite image construction is hr-image=imfuse(composite-1, composite-2);
shown in Listing 2 and flow chart is shown imwrite(hr-image, 'final_composite.png')
in Fig. 3

141
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

cates that the predicted pixels are very near

identical to the original pixels. We have
compared the reconstructed image with the
original image of size 512x512 to calculate
the values of image parameters.
MATLAB’s online environment
was used to generate all the intermediate
images and the final high-resolution image.
We have developed the functions to enlarge
the image using the Nearest Neighbour In-
terpolation algorithm, Bilinear Interpolation
algorithm, High-Quality scale (Hqx), and
Scale by rules (xBR). These functions take a
low-resolution image of size 256x256 as an
input and give an enlarged image of
512x512. All input and output images are of
PNG format. Enlarged images are passed to
the IBP algorithm to reduce the image blur-
ring.

4. Results and Analysis

The proposed methodology produced
Fig. 3. Flow Chart to construct composite image. a high-resolution image from a low-
resolution image of the human brain MRI
In this paper, we have used the pa- [18-19].
rameters of Peak Signal to Noise Ratio Fig. 4 shows the input and output im-
(PSNR), Signal to Noise Ratio (SNR), ages after applying the image enlargement
Structural similarity (SSIM) and Mean algorithm and values of image parameters
Squared Error (MSE) to compare the image are shown in Table 1. It is evident from Ta-
quality. PSNR checks the noise in the im- ble 1 that different enlargement algorithms
age. The value of noise should be minimum improve different parameter values of the
to increase the value of PSNR. A better im- input image. The Bilinear Interpolation im-
age always has a high PSNR. SNR indicates age enlargement algorithm gives us the best
a signal to noise ratio, the signal should be PSNR value. Hqx gives the lowest PSNR
higher than the noise. A high value with and the highest structural similarity. MSE of
threshold will have a good quality image. the image is minimum with Nearest Neigh-
SSIM shows the similarity of structure be- bour Interpolation. Similarly, we are getting
tween the images. A higher value of SSIM maximum SNR with xBR image enlarge-
produces images of similar structure. MSE ment algorithm.
is a mean squared error between the corre-
sponding pixels, and the lower value indi-

142
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

(a) (b) (c) (d)

(e)
Fig. 4. Input and output images after applying image enlargement algorithms, (a) input image (b) Near-
est Neighbour Interpolation (c) Bilinear Interpolation d) Hqx (e) xBR.

Table 1. Image parameters after applying Image Enlarging Algorithms.

Nearest Neighbour Inter-
Bilinear Interpolation HQX xBR
polation
PSNR 21.1468 21.2471 21.0172 21.1422
SNR 11.4698 11.5701 11.3403 11.5952
SSIM 0.7040 0.6969 0.7165 0.7058
MSE 483.3434 487.9484 514.4660 499.8716

We have used the IBP algorithm to values are shown in Table 2. As compared
improve image quality. The improved imag- to Table 1, the values of parameters have
es are shown in Fig. 5, and new parameter been significantly improved.

(a) (b) (c) (d)

Fig. 5. Images after applying iterative back-projection (IBP) algorithm on the four images received
from the previous step, enhanced images of the output of (a) Nearest Neighbour Interpolation (b) Bilin-
ear Interpolation (c) Hqx (d) xBR.

143
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

Table 2. Image parameters after applying Iterative Back Projection Algorithm on enlarged
images.
Enhanced-Nearest Enhanced-Bilinear Enhanced-Hqx Enhanced-xBR
Neighbour Interpolation Interpolation
PSNR 21.1759 21.5021 21.2734 21.2586
SNR 11.4989 11.8252 11.5964 11.6217
SSIM 0.6975 0.7232 0.7310 0.7115
MSE 456.0142 460.1153 484.9981 486.6495

Figs. 6a and 6b show the results of The final enhanced composite image,
the composite images. Two composite im- shown in Fig. 6c, is generated using the im-
ages are generated: the first using En- ages received from Fig. 6a and 6b. As is
hanced-Nearest Neighbour Interpolation and evident from Table 3, the PSNR, SNR,
Enhanced-Bilinear Interpolation and the SSIM and MSE values of the final image
second using Enhanced-HQX and En- have better parameter values as compared to
hanced-xBR. Parameter values of composite images obtained after applying the image
images are shown in Table 3. These compo- enlargement algorithm and further after us-
site images have further improved PSNR, ing the IBP algorithm.
SNR, SSIM and MSE.

Table 3. Image parameters of composite images and final high-resolution image.

Composite of Enhanced-Nearest Composite of Enhanced-HQX Final Image
Neighbour Interpolation and and Enhanced-xBR
Enhanced-Bilinear Interpola-
tion
PSNR 21.7207 21.7970 22.0599
SNR 12.4337 12.1200 14.3829
SSIM 0.7587 0.7374 0.8420
MSE 437.5337 429.9160 302.6586

(a) (b) (c)

Fig. 6. Composite images of, (a) enhanced images of Nearest Neighbour Interpolation and Bilinear In-
terpolation (b) Hqx and xBR (c) final enhanced high-resolution image, composite of previous two im-
ages (a) and (b).

Likewise, we applied the proposed algorithms for image enhancement, namely

methodology to the image shown in Fig. MIEBF and KBIE. The comparison is
7(a). Fig. 7(b), 7(c) and 7(d) show enhanced shown in Table 4. It is evident from Table 4
images from MIEBF[21], KBIE[22], and that for 3 of the parameters, namely PSNR,
the proposed approach, respectively. SNR and SSIM, values of the resultant im-
We have also compared the results age obtained using our methodology are
obtained from our approach with two other better than the other two approaches.

144
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

(a) (b) (c) (d)

Fig. 7. Input and outout image of Knee MRI image dataset MRNet [20], (a) Input image (b) MIEBF
[21] (c) KBIE [22] (d) Proposed Methodology.

Table 4. Comparison of parameters of the image shown in Fig. 7 with methodologies [21] &
[22] and proposed methodology.
Proposed
MIEBF[21] KBIE[22]
Methodology
PSNR 27.0021 27.9658 28.4514
SNR 18.9887 19.5521 19.8822
SSIM 0.8778 0.8696 0.8852
MSE 178.0215 152.2374 200.8541

Furthermore, we applied our method- MSE parameters were improved for most of
ology on 177 knee MRI images of the the images as depicted in Fig. 8.
MRNet [20] and the PSNR, SNR, SSIM and

Fig. 8. Number of images for which PSNR, SNR, SSIM and MSE parameters were improved.

5. Conclusion and Future Work ity of enlarged images, the IBP algorithm
We proposed a methodology to re- has been applied to the enlarged images,
construct an enhanced high-resolution im- followed by the construction of composite
age from the low-resolution medical image. images. Experimental results have shown
We have used four image enlargement algo- that the proposed method has reduced the
rithms. The enlargement algorithms are also noise and enriched the details. The proposed
able to generate a high-resolution image but methodology has resulted in improved im-
may have less detail, so to improve the qual- age parameters, namely PSNR, SNR, SSIM,

145
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

and MSE, in each step. The final high- rithm based on ringing artifacts suppres-
resolution image has significantly improved sion. Neurocomputing. 2015;162:171-9.
the values of the image parameters under
consideration. Our methodology evaluated [8] Yoo J, Kim J. Noise-Robust Iterative
Back-Projection. IEEE Transactions on
the performance of the composite image
Image Processing. 2020;29:1219-32.
before and after the reduction of noise and
blur. In the future, we intend to improve the [9] Zhu J, Zhou C, Fan D, Zhou J. A New
image quality by creating a composite im- Method for Superresolution Image Re-
age using other algorithms so as to experi- construction Based on Surveying Ad-
ment and observe how finer details of the justment. Journal of Nanomaterials.
images can be preserved. Different image 2014;2014:1-6.
enlargement algorithms can also be applied
for experimentation. [10] Sundar K, Vaithiyanathan V. A novel
method for super resolution image recon-
struction. International Journal of Ad-
References vanced Intelligence Paradigms.
[1] Lin H, Lin C, Lin C, Yang S, Yu C. A 2019;14(1/2):46.
Study of Digital Image Enlargement and
Enhancement. Mathematical Problems in [11] Yu C, Lin H, Lin R. Eight-Scale Image
Engineering. 2014; 2014:1-7. Contrast Enhancement Based on Adap-
tive Inverse Hyperbolic Tangent Algo-
[2] Danahy E, Agaian S, Panetta K. Algo- rithm. 2014 International Symposium on
rithms for the resizing of binary and Computer, Consumer and Control. 2014.
grayscale images using a logical trans-
form. Image Processing: Algorithms and [12] Liu M, Huang J, Gao M, Qin S. High
Systems V. 2007. Performance Super-Resolution Recon-
struction of Multiple Images Based on
[3] Kreuzer F, Kopf J, Wimmer M. Depix- Fast Registration and Edge Enhancement.
elizing pixel art in real-time. Proceedings Lecture Notes in Computer Science.
of the 19th Symposium on Interactive 3D 2013;:649-57.
Graphics and Games - i3D '15. 2015.
[13] Tanaka M, Okutomi M. A fast MAP-
[4] Kerautret B, Lachaud J. Geometric Total based super-resolution algorithm for gen-
Variation for Image Vectorization, Zoom- eral motion. Computational Imaging IV.
ing and Pixel Art Depixelizing. Lecture 2006.
Notes in Computer Science. 2020;:391-
405. [14] Li T, Chen H, Zhang M, Liu S, Xia S,
Cao X et al. A new design in iterative im-
[5] Stasik P, Balcerek J. Extensible Imple- age deblurring for improved robustness
mentation of Reliable Pixel Art Interpola- and performance. Pattern Recognition.
tion. Foundations of Computing and De- 2019;90:134-46.
cision Sciences. 2019;44(2):213-39.
[15] Li F, Fraser D, Jia X. Efficient IBP with
[6] Fuentes J, Ruiz J. A deblurring model for super resolution for ALOS imagery.
super-resolution MRI interpolated imag- Geoinformatics 2007: Remotely Sensed
es. 15th International Symposium on Data and Information. 2007.
Medical Information Processing and
Analysis. 2020. [16] Irani M, Peleg S. Improving resolution by
image registration. CVGIP: Graphical
[7] Yang X, Zhang Y, Zhou D, Yang R. An Models and Image Processing.
improved iterative back projection algo- 1991;53(3):231-9.

146
 L. Kumar and M. Jain | Science & Technology Asia | Vol.27 No.2 April - June 2022

[17] Purkait P, Chanda B. Super Resolution [20] Bien N, Rajpurkar P, Ball R, Irvin J, Park
Image Reconstruction Through Bregman A, Jones E et al. Deep-learning-assisted
Iteration Using Morphologic Regulariza- diagnosis for knee magnetic resonance
tion. IEEE Transactions on Image Pro- imaging: Development and retrospective
cessing. 2012;21(9):4029-39. validation of MRNet. PLOS Medicine.
2018;15(11):e1002699.
[18] Marcus D, Wang T, Parker J, Csernansky
J, Morris J, Buckner R. Open Access Se- [21] Anoop V, Bipin PR. Medical Image En-
ries of Imaging Studies (OASIS): Cross- hancement by a Bilateral Filter Using Op-
sectional MRI Data in Young, Middle timization Technique. J Med Syst. 2019;
Aged, Nondemented, and Demented Old- 43(8).
er Adults. Journal of Cognitive Neurosci-
ence. 2007;19(9):1498-507. [22] Muslim H, Khan S, Hussain S, Jamal A,
Qasim H.A knowledge-based image en-
[19] Yigit A, Isik Z. Application of artificial hancement and denoising approach.
neural networks in dementia and alzhei- Computational and Mathematical Organ-
mer's diagnosis. 2018 26th Signal Pro- ization Theory. 2018;25(2):108-21.
cessing and Communications Applica-
tions Conference (SIU). 2018.

147