2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS)

Hybrid LWT and DCT based High-Quality

Image Compression
S Narasimhulu T Ramashri
Department of ECE, Professor and Head of the Department of ECE,
S V University College of Engineering, S V University College of Engineering,
Sri Venkateswara University, Tirupati-517502, India. Sri Venkateswara University,
[email protected] Tirupati-517502, India.
2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS) | 978-1-6654-0816-5/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICACCS54159.2022.9785197

Abstract— Research in the field of image compression has compressor [6-7]. Before encoding, the image is pre-
grown significantly with the increasing need to publish images processed for encoding and does not contain any application-
in computer and mobile environments. Image abstraction plays specific functions. After decoding the compressed file, a
an important role in digital image processing, which is crucial subsequent processing can be performed to remove some
for efficient image transmission and storage. Without
degrading the image quality, the image size in byte is minimized
unwanted artifacts obtained due to the compression process.
by using image compression technique. Many different ways The first step in pre-processing is data reduction. Gray level
are used in this process to compress the input image. Therefore, and/or spatial scaling can reduce image data or improve a
this research study proposes hybrid compression technique desired image (eg., remove noise). The mapping process is
includes Lifting Wavelet Transform (LWT) with discrete cosine the second step in IC technique, where mathematical space is
transform (DCT) to provide the best properties for an image mapped from the original data and it is easy to compress the
that is performed in JPEG compression method. To get the input data [8].
high quality compression ratio, lossy and lossless methods are
used by this technique and also used to maintain the The quantisation stage is the next part of encoding
reconstructed image quality. A high compression ratio is process, where mapping stage provides the continuous data
produced by lossy compression, where reconstructed images potentially and discrete form is obtained. The coding of
high quality is brought by lossless compression and later it will resultant data is presented in the final stage of encoding. [9].
be decompressed with same results. MATLAB software tool is
used to implement the proposed compression process and to
The summary algorithm may include all levels or only one
verify its performance in terms of various parameters such as or two levels. There are two stages presents in the decoding
PSNR, MSE and SNR. process. The decoder converts the original code by taking
the zip file in the first step and assigning the codes to the
Keywords— Image Compression; Lifting Wavelet Transform; original and measured values. Then, In order to change the
Discrete Cosine Transform; Sub-bands; High Quality actual mapping process, these values are processed through a
Compression ratio, lossless compression, Quality Image. step-by-step reverse mapping [10]. Finally, the image can be
I. INTRODUCTION processed later to improve the final image’s appearance.
Creating an abstract algorithm is a very specific application.
For lossy image compression (IC), one of the most During the compression pre-treatment phase, processing
common compression techniques is JPEG, where lossy and such as upgrading, denoising or calibration is used. The
lossless compression [1] are the two different types of IC purpose of pre-processing is to remove irrelevant
technique. A high compression ratio is achieved by lossy IC information identified by the application and prepare the
techniques, which is non-reversible and the best visual image for the encryption process. Due to high relevant of
experience is provided by lossless methods. A quality trade- input image, the mapping process is highly recommended
off between decomposed image and file size is achieved by [11]. If we know the value of one pixel, then the value of the
lossy compression. Due to high compression ratio, users nearest pixel will be the same. This kind of data redundancy
prefers lossy compression schemes on consumer devices [2]. can be eliminated by finding the mapping equation
associated with the data. It should be noted that the
In the field of multimedia processing, researchers measurement process cannot be undone, so some
considered digital image compression as a popular topic. information may be lost during the measurement.
Among neighboring pixels or samples [3-4], redundancy or
significant correlation is exists within a single image and The rest of the paper includes: Section 2 provides a
therefore, an image is compressed. Hence, spatial correlation review of existing methodologies, while Section 3 introduces
is also known as significant correlation. Among samples the new IC technique. Section 4 details the process of testing
from multiple sensors for data acquisition, the proposed model, and Section 5 summarises the findings
redundancy/spectral correlation is exists in those images. of the study as a whole..
Thus abstraction is achieved through repetition of films [5].
The most effective contraction can be achieved if repetition II. RELATED WORKS
is eliminated prior to contraction. The image compression Song and et al. (2018) [12] developed an efficient
system consists of two parts: the compressor and the de- algorithm used to compress prognostic clinical images based
compressor. There are two stages presents in the compressor on least square (LS). In this work, adaptive geometric
that includes pre-processing and encoding stage, where partitioning and tetragonal partitioning achieve adaptive
decoding and post-processing stage are presented in the de-

978-1-6654-0816-5/22/$31.00 ©2022 IEEE

1620
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 10,2023 at 06:02:28 UTC from IEEE Xplore. Restrictions apply.
 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS)
stochastic partitioning. The images are segmented by using III. PROPOSED SYSTEM
these segmentation techniques in an effective way. With
high lock size, non-overlapping square blocks are obtained Figure 1 and 2 shows the workflow of image
by dividing this input images. The each block's segmentation compression and decompression process. Initially, input
mode decision is decided by cost function of Lagrange. Once image are processed with LWT technique, which is divided
the block segment is complete, LS-based calculators are used into four sub-bands. The suitable components of images are
to predict segmented areas. Each pixel's predictor is updated categorized into block size of 8×8 and then forward DCT
by using LS-based prediction, however implementing this transformation is applied in that for effective compression.
predictor is a complexity and time-consuming process. Finally, compressed bit streams are obtained by using
Huffman code. This same process is reversed by applying
Song and et al. (2019) [13] Developed a higher ILWT to get the decompressed image.
resolution to reduce high-frequency data distortion in the
image. This volume was created based on the X-Golomb
code to achieve an adaptive volume. The infinite alphabet of
exponential symbol is used to encode the symbols, where
this code is one of the variable length coding type. Instead of
standard scaling, features of this code are used to achieve
adaptive scaling. Here, the code is used only for non-
negative values to obtain the remaining image before
calibration. Therefore, positive values are not taken into
account in the measurement process.
Devadas and Shankara Gomati (2019) [14] provided
helpful summaries of the medical film. In order to segregate
the region of ROI and non ROI as NROI, morphological
segmentation is used, where Burrows–Wheeler Transform
(BWT) are cascaded with move to front transform (MTF) to
compress the ROI images. BWT is a reversible transformer Fig. 1. Proposed Compression block diagram.
that groups the input data into groups of similar continuous
components. The BWT version contains the same
ingredients over and over again. MTF is a global
architectural change that transforms the local environment
into a global one. The NROI region is compressed by the
hybrid fractal technique. To reduce the time taken by BWT,
size separation is performed on the ROI image. In addition,
increasing the volume reduces the percentage of space saved.
Mei and others (2018) [15] proposed a hybrid algorithm
called Folded-PCA with JPEG2000. The spectral vector is
folded into matrix by calculating the covariance matrix, the
principal components are obtained by eigenvectors to
represent the entire image' features. JPEG2000 will be used
to further compress the thumbnail. The algorithm was
extended to re-weight PCA (WPCA) using the adaptive
cosine estimator algorithm for target detection. Later, after
converting to a 2D matrix, PCA was applied to HSI, but Fig. 2. Proposed De-Compression block diagram.
some changes were made to the weight matrix. Around the
identified target, pixels gained more weights by calculating Lifting wavelet transform (LWT)
the covariance and mean pixel matrix.
In many applications such as compression and processing
Guerra and others. (2018) [16] proposed a loss extraction of images are effectively handled by the signal processing
algorithm called hyper LCA using the transformation tool called LWT. When compared with classical transforms,
process to obtain an optimal CR at a reasonably calculated floating point values or integers are highly supported by
complex cost. In order to remove the spectral redundancy, LWT, which is considered as one of its major properties.
three steps are used in this work such as spectral But, due to round-off operations, information loss may occur
transformation function, pre-processing and lossless in image processing, when the floating point presentation is
encoding. During pre-processing stage, the selection of used. In order to describe the pixels, LWT uses 8-bit integers
various pixels are carried out that will be an independent of for processing the image. The input data is mapped into
spatial alignment. Golomb rice encoding is used to store this integers by using IWT, which is reversible and provided that
information until the image is finally compressed. integers without quantization errors. IWT has three
processes such as updating, predicting and splitting, which is
more computationally efficient, even though it has similar

1621
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 10,2023 at 06:02:28 UTC from IEEE Xplore. Restrictions apply.
 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS)
processes to LWT. Four subdomains are used to develop
IWT scale of Lena image, which is shown in Figure 3.

(a) (b) (c)

Fig. 3. IWT sub-bands of Lena

DCT is mostly used to separate pictures into discrete
frequency components. For example, in decompression, it is
(d) (e) (f)
possible to use only the most important frequencies to
extract an image, while discarding the less important ones. A
Fig. 4. Input compression images.
resolution of bits is used to quantize the blocks in
Quantization. There are a lot of zero sequences because of
Compression Quality Measurement
the DCT coefficients vector thresholding. It is then replaced
by a code of bits for each sequence 1 is the new minimum SNR (signal to noise ratio), MSE, and PSNR are three of
value. the most often used measures for objective picture
measurement. Image quality is measured by the amount of
A. Discrete cosine transformation (DCT)
quality lost when images are compressed. The image can be
DCT is mostly used to separate pictures into discrete viewed as a matrix, with the image's elements acting as the
frequency components. A method of decompression in matrix's nodes. Analyzing differences between input and
which only the most relevant frequencies are used to extract output matrixes can be simplified by using this method. For
the image, while the less relevant frequencies are discarded. example, it is possible to compare the performance of
different compression methods and perform the same
B. Quantization compression procedure, but with varying quantities of
With a resolution of Q bits, the blocks of NZC are compression at the same time. The number of picture
quantized linearly. Following the formula [4], these elements in the portrait orientation is M, and the number of
coefficients NZC are represented as follows: image elements in the landscape orientation is N, and these
elements make up the a ij set at the pressure system's
(1) entrance. MxN is the total number of picture elements in the
image that was observed. A ij components are created at the
where: QNZC is an array of quantized DCT coefficients end of the pressing system. When comparing matrices A and
dispersed at regular intervals. Inverse quantization is carried A, it is clear that there has been a decline in quality. The
out in accordance with the following formula. bigger the amount, the more likely it is that a mistake may
occur. The amount of data required to display the
(2) compressed image and the quality of the rebuilt image can
be modified by the user.
The developed novel encoder that combines two of the Reconstruction error is defined as:
most often used approaches. There are four possible ways to
scan each quantized block: horizontal, vertical, Hilbert, and (3)
zigzag. This allows the coefficients to be ordered in a way
The difference among matrices A and A' can be shown using
that simplifies the coding. After the TRE coder scans the
the MSE:
quantized coefficients, the compressed data is then encoded.
(4)
IV. RESULT AND DISCUSSION
All the pixels in an image are processed to arrive at a
Detailed comparisons of our method to other methods total number of pixels known as MN.
are shown in this section so that we can demonstrate its As the number of bits needed to display an image's range
efficacy. MATLAB is used to evaluate the suggested model of elements increases, so does its amplitude.
on an Intel(TM) Core(TM) i3 CPU M 380 running at
2.53GHz with 5GB of RAM (R2019b). Figure shows some A PSNR was introduced because MSE failed to
examples of image compression. account for the size of image components.

(5)

1622
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 10,2023 at 06:02:28 UTC from IEEE Xplore. Restrictions apply.
 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS)
Here, the is the maximum pixel value of the image. TABLE III: COMPARATIVE ANALYSIS OF SUB-
When pixels are represented by B bits per sample, is BANDS IN TERMS OF PSNR VALUES.
2B-1.
Sub- Bit per pixel
If we assume that n = 8 bits per image element then it is: band
0.1 0.2 0.3 0.4 0.5
(6)
Typical PSNR values for lossy images are between 30 and LL 29.8 32.7 34.5 36.8 38.6
50 dB
A. Proposed System Evaluation LH 29.1 32.1 33.9 36.1 37.9
Table 1, Table 2, and Table 3 show the MSE, SNR, and
PSNR values for various band rates during JPEG HL 28.8 31.8 33.5 35.5 37.5
compression for the images used (shown in Figure 4)
HH 28.6 31.5 33.2 35.3 37.1
TABLE I: COMPARATIVE ANALYSIS OF SUB-
BANDS IN TERMS OF MSE EVALUATION
When the bpp values are increases, each and every sub-
Sub- Bit per pixel bands has higher PSNR value. The LL and LH bands
band achieved 29% of PSNR, where HL and HH achieved 28% of
0.1 0.2 0.3 0.4 0.5 PSNR for 0.1bpp. The LH, HL and HH achieved 33% of
LL 68.8 35.0 23.1 17.0 13.7 PSNR and LL achieved 34.5% of PSNR for 0.3bpp. The LL
and LH achieved 36% of PSNR and HL and HH achieved
LH 79.3 40.2 26.8 19.7 15.8 35% of PSNR for 0.4bpp. When the value is 0.5bpp, the LH,
HL 85.2 43.3 29.1 21.5 17.2 HL and HH achieved 37% of PSNR and LL achieved 38.6%
of PSNR. Table 4 presents the average performance of
HH 89.2 45.8 30.9 23.0 18.3 proposed model by using test images with 8× 8 block size
and Figure 5 shows the graphical representation of proposed
The sub-band LL achieved low MSE (68.8%), when model in terms of PSNR value.
compared with other bands for 0.1bpp. The other bands such
as LH, HL and HH achieved 79.3%, 85.2% and 89.2% for TABLE IV: THE AVERAGE PERFORMANCE OVER
the same bpp. When the value of bpp is increases, the MSE THE TEST IMAGES 8× 8 BLOCK SIZE AND FOR
value of every band decreases, for instance, LL has 13.7%, QUANTIZER WIDTH VALUES OF 8.
LH has 15.8%, HL has 17.2% and HH has 18.3% of MSE
for the 0.5bpp. However, the moderate MSE is achieved for Images 8 bits
all sub-bands, when they reaches 0.3bpp. The HH band PSNR bpp
achieved high MSE value for all different values of bpp.
A 33.88 1.04
TABLE II: COMPARATIVE ANALYSIS OF SUB- B 32.04 1.15
BANDS IN TERMS OF SNR. C 32.65 0.94
Sub- Bit per pixel D 35.69 0.74
band
0.1 0.2 0.3 0.4 0.5 E 34.14 0.95
F 34.32 1.09
LL 15.6 18.6 20.4 21.7 24.7
G 32.68 1.11
LH 15.5 18.4 20.2 22.5 24.2
H 33.62 1.00
HL 15.3 18.2 20.0 22.2 24.0
When compared with all test images, the image D
HH 15.1 18.0 19.7 22.0 23.8
achieved 35.69% of PSNR and 0.74bpp. The test images of
B, C and G achieved nearly 32.04% to 32.68% of PSNR,
In contrast with MSE, each and every sub-bands has high where B image achieved 1.15bpp, C image achieved 0.94bpp
SNR value for different bit per pixels. For instance, LL, LH, and G image achieved 1.11bpp. The image E and F achieved
HL and HH achieved nearly 18.0% to 18.6% of SNR for high PSNR than other images expect D image i.e. 34% of
0.2bpp. Likewise, every bands achieved nearly 15.1% to PSNR, where E has 0.95bpp and F has 1.09bpp. The first
15.6% of SNR for 0.1bpp. When the value is 0.3bpp, LL, LH image (A) and last image (H) achieved nearly 33% of PSNR
and HL achieved nearly 20% of SNR, but HH achieved and 1.00 to 1.04bpp.
19.7% of SNR. The HH band achieved 22% of SNR and
23.8% of SNR for 0.4bpp and 0.5bpp. The other bands such
as LH, HL and HH achieved 22% of SNR and 24% of SNR
for 0.4bpp and 0.5bpp.

1623
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 10,2023 at 06:02:28 UTC from IEEE Xplore. Restrictions apply.
 2022 8th International Conference on Advanced Computing and Communication Systems (ICACCS)
Theoretical and Applied Information Technology, vol/issue. 68(1),
pp. 215–223, 2014.
[9] V.V. Yap, and R. Comley, “A Hybrid Wavelet-Based Compression
System”, Proceedings of the Third International Conference on
Information Technology and Applications (ICITA’05), IEEE, 2005.
[10] R. Sharadha, and C. Bhanuprakash, “Image Compression Technique
for Color Images using Wavelets”, International Journal of Advanced
Research in Computer Science and Software Engineering, vol/issue:
3(11), pp.322–327, 2013.
[11] G.M. Padmaja, and P. Panirupama, “Analysis of Various Image
Compression Techniques”, ARPN Journal of Science and
Technology, vol/issue: 2(4), pp. 371–376, 2012
[12] Song X, Huang Q, Chang S, He J, Wang H (2018) Lossless medical
image compression using geometry-adaptive partitioning and least
square-based prediction. Med Biol Eng Comput 56:957–966
[13] Song X, Liu B, Huang Q, Hu R (2019) Design of high-resolution
quantization scheme with exp-Golomb code applied to compression
Fig. 5. Graphical Representation of Proposed Model on test of special images. J Vis Commun Image Represent 65:102684
[14] Devadoss CP, Sankaragomathi B (2019) Near lossless medical image
images in terms of PSNR. compression using block BWT–MTF and hybrid fractal compression
techniques. Clust Comput 22:12929–12937
V. CONCLUSION [15] S. Mei et al., “Low-complexity hyperspectral image compression
using folded PCA and JPEG2000,” in IEEE Int. Geosci. and Remote
In this proposed method, a sample called Lena is Sens. Symp., IEEE, pp. 4756–4759 (2018).
compressed using LWT that contains four sub-bands namely [16] R. Guerra et al., “A new algorithm for the on-board compression of
LL, LH, HH and HL. The components of images are hyperspectral images,” Remote Sens. 10(3), 428 (2018).
partitioned into elemental blocks and then, forward DCT is
applied to get the transform co-efficient values. Finally, the
compressed bit streams are achieved by applying Huffman
code and quantization, where this process is reversed to get
the decompressed images with high quality compression
ratio. In this experiments, 8 color test images are taken as
sample input and tested in terms of PSNR, SNR and MSE
for four LWT sub-bands. The results proved that HH band
has 89.2% of MSE, 15.1% of SNR and 28.6% of PSNR for
0.1bpp. The test image called aeroplane (D) achieved high
PSNR value i.e. 35.59% and 0.74bpp. In future work, an
efficient hybrid image compression model is developed to
obtain high quality compression ratio for both input and
reconstructed images.
REFERENCES
[1] Wang, Z., Bovik, A., & Lu, L. (2021). Why is image quality
assessment so difficult? Retrieved from https://
live.ece.utexas.edu/publications/ 2002/zw_icassp2002_whyqa.pdf.
[2] Kunwar, S. (2021). Strategies in JPEG compression using
Convolutional Neural Network (CNN). arXiv preprint
arXiv:2112.04500.
[3] K.K. Hasan, U.K. Ngah, and M.F. M. Salleh, “Low Complexity
Image Compression Architecture Based on Lifting Wavelet
Transform and Embedded Hierarchical Structures”, IEEE
International Conference on Control System, Computing and
Engineering, IEEE, pp. 305–309, 2013.
[4] S.Z. Xie, et al., “Image Compression using Wavelet Transform with
Lifting Scheme and SPIHT in Digital Cameras for Bayer CFA”,
Proceedings of the 2013 International Conference on Wavelet
Analysis and Pattern Recognition, Tianjin, pp. 163–167, 2013.
[5] S. Srikanth, and S.Meher, “Compression Efficiency for Combining
Different Embedded Image Compression Techniques with Huffman
Encoding”, International Conference on Communication and Signal
Processing, IEEE, pp. 816–820, 2013.
[6] K.P. Swetha, “Still Image Compression by using New Wavelet Bi-
Orthogonal Filter Coefficients”, Indian Journal of Applied Research,
vol/issue: 3(7), pp. 304–306, 2013.
[7] Y. Liu, and Z. Liu, “An Improved Image Compression Algorithm
Based on Embedded Zerotree Wavelets Transform”, Journal of
Theoretical and Applied Information Technology, vol/issue: 48(2),
pp.1097–1102, 2013.
[8] T. Karthikeyan, et al., “Embedded Zero Tree Wavelet and
Orthogonal Polynomial Based Transformation Coding”, Journal of

1624
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on May 10,2023 at 06:02:28 UTC from IEEE Xplore. Restrictions apply.