2018 Third International Conference on Electrical, Electronics, Communication, Computer Technologies and Optimization Techniques (ICEECCOT) 14-

15, December 2018

A Survey on Performance of Various Image

Compression Techniques
Gireesh H R Rangaiah L
VTU,Belagavi, India Department of ECE
[email protected] RRCE, Bangalore, India
[email protected]

internal structures of the body and their functionalities. MRI

ABSRACT--Image compression enables quite exciting solutions images also provide the 3D image of the internal organisms
in many fields, such as image analysis, bio-medical image
of human body which further assist the doctors to analyze
processing, wireless systems and seems to be a key application
in today’s digital and smart world. Image compression seems and cure the diseases as compared to the traditional
to be a powerful tool in case of transmission and storage of techniques such as x-ray imaging which provide the 2D
large data images in various applications such as big data, images.
medical etc. However due to exclusive and quality soft tissue
contrast, Dynamic Magnetic Resonance Imaging (MRI) has From past few decades, Dynamic Magnetic
been a field of attraction with increasing attention in recent Resonance Imaging (MRI) has been employed in order to
decades. Moreover, MRI is considered as one of the most extract all the related functional information corresponding
effective and strongest diagnosis system making the extensive to the peripheral vascular organism [1-3]. Here, multiphase
usage of magnetic and radio waves in order to diagnose the MRI scanned images are in use along with the extensive use
human organs. This diagnosis is capable of generating 3D of intravenous injection of a contrast agent in order to
images with detailed anatomical features without any X-ray perform the clinical practices. However, a lot of researchers
radiations. The prime purpose of this survey is to provide a
have worked in the domain of the utility of the dynamic
comprehensive report of different image compression schemes
in order to design an efficient compression scheme for dynamic contrast enhancement (DCE) MRI in order to perform the
MRI images. In this paper, author has surveyed different various diagnosis schemes for salivary gland tumors [4].
image compression schemes which are either sole There are lots of techniques which are used in order to fulfill
implementations or hybrid of two or more algorithms. The these requirements. However, there are two important
author has also presented a comparative analysis for the schemes which are used more frequently in order to measure
surveyed compression schemes. This survey paper finally the perfusion and energy consumption. These are known as
makes inroads for further researches in the domain of image arterial spin labeling (ASL) MRI and Blood Oxygen Level-
compression schemes for dynamic MRI images. Dependent (BOLD) MRI. While these techniques are
limited to the stated criterion, they are further incapable to
Keywords–-Dynamic MRI, Image Compression, Compression
Ratio, Redundancy, PSNR, SSIM. discover and provide the information for volume functions
or vessel permeability [5-6].
Since the medical imaging technology is getting
I. INTRODUCTION developed so rapidly along with the insertion of a variety of
imaging modalities, the field of bio-medical image
Magnetic Resonance Imaging (MRI) includes both processing has further been renovated exceptionally.
the features of non-invasive for superior flexible tissue However, the image acquisition is performed with the
contrast and non-exposing nature with respect to the high extensive use of variety of devices using various types of
ionizing radiation. Nowadays, MRI are becoming one of the modalities. Among all the modalities MRI was found to be
key area of study in order to carry out computer aided the most dominant and probable tool in order to visualize all
diagnosis and clinical routine since they are the sources of the comprehensive internal organisms to enable the precise
prime and inevitable information about structure of tissue, and extraordinary measurement of human organ anatomy.
such as size, shape, and localization. This is also one of the The MRI medical imaging further presents a superior
reasons for selection of domain by researcher too. contrast among the various soft tissues of human body.
Furthermore, medical imaging enables us to obtain the Hence, this is extensively utilized in order to diagnosis and
images of the inner organs, bones, tissues etc. in order to hence treatment of the diseases. However while these
perform research and analysis work in medical domain. images are broadcasted, they are going to take up a great
These images can further be utilized in order to perform the storage space, with inferior speed and larger bandwidth.
study of functions of few organs. These images are also Also, these images cannot be compressed with the use of
used in radiology in order to visualize common compression techniques [7].

978-1-5386-5130-8/18/$31.00 ©2018 IEEE 1233

 The rest of the paper is planned as follows. Section threshold value can be achieved to perform desired level of
2 presents the different image compression schemes. compression to obtain the required compression ratio.
Section 3 presents the different performance parameters
used for checking the quality and efficiency of compression The implementation of multi-wavelet transform in
schemes. Section 4 presents a comparison of different image order to perform lossless image compression was done in
compression schemes discussed in section 2 based on [9]. Here, the performance of the IMWT (Integer Multi-
different performance parameters discussed in section 3. Wavelet Transform) was analyzed for lossless image
Section 5 presents the conclusion of the paper. compression. The IMWT enables to achieve good results
during the reconstruction phase too. However, the
II. DIFFERENT IMAGE COMPRESSION performance of the Integer Multi-Wavelet Transform during
TECHNIQUES the lossless compression was further improved with
extensive use of magnitude set coding. In this work, the
Image compression is a technique which deals with author coded the transform coefficient with the use of both
the issues related to the reduction of the amount of data or magnitude set of coding and run length encoding scheme.
information required in order to represent any digital image. Here, the bit rate of about 2.1 bpp (bits per pixel) was
This image processing methodology is anticipated in order achieved when use of MS-VLI (Magnitude Set-Variable
to capitulate an optimized and compact demonstration of Length Integer Representation) with RLE scheme was taken
any digital image while tumbling the huge requirements into consideration. While the bit rate was further improved
related to the image transmission and storage. There are two to about to 3.1 bpp was achieved when use of MS-VLI
different classifications of compression schemes based on (Magnitude Set-Variable Length Integer Representation)
the data available after the reconstruction. These are lossy without RLE scheme was implemented.
and lossless compressions. In lossy compression schemes,
the image reconstructed after compression is composed of In [10] author proposed a hybrid technique for
degradations with respect to the original image. Since these image compression. This technique used unique type of
techniques enable us to obtain extremely high embedded Wavelet based image coding along with
compressibility of data, the redundant data is completely Huffman-encoder. Here, author implemented both EZW and
rejected. However, in lossless compression techniques, the SPIHT coding schemes along with the standard Huffman
image reconstructed after compression is exactly similar to encoding while making the extensive use of various types of
the original image. But these schemes are reserved in terms available wavelet families. For implanted various types of
of compressibility and hence don’t provide appreciable wavelet families author compared both bit rates and PSNR
compression. Furthermore, the compression techniques with values for various types of images and proved to be efficient
predictive coding are facilitated with prediction of the future in term of image quality, compression ratio as compared to
values for transmitted data and hence the difference is going the existing schemes of image compression. A novel scheme
to be coded in order to achieve the compression. Here, the of image compression along with the encryption was
entire work is carried out in spatial domain and is easy to proposed in [11]. In this technique, the image compression
implement. Alternatively, the compression techniques with was done using SPIHT scheme in order to achieve better
transformations are going to enhance the data compression ratios but before that the images were
compressibility while using various types of transforms in encrypted using stream cipher technique to obtain the better
order to perform compression at the cost of increased encryption before transmission. This enables us to obtain a
complexity and computations. In this section we are going highly confidential encryption along with the best
to discuss different researches conducted in domain of compression rate enhance the security, which was the key
image compression in past decade. objective of research.
A novel image compression technique based on the A hybrid image compression (HIC) scheme based
utilization of the reference points coding along with the on the combination of discrete wavelet transform (DWT)
computation of threshold values was presented in [8]. The and discrete cosine transform (DCT) was proposed in [12].
key objective of this paper was to implement and lighten up Here, author used the images of 256 × 256 resolution. The
the image compression technique which is proficient in author employed two levels DWT on each sub band images
order to perform both lossy and lossless types of image of size 32 × 32. However out of four available sub bands
compression. In addition to this, a threshold value was also (LL, LH, HL, HH), author used only LL sub band images.
incorporated in the presented compression process. The This in turn provides the 75% compression. Furthermore,
variation of this threshold value enables to achieve various the standard DWT was applied which avails 8x8 blocks.
compression ratios. Hence, in order to perform lossless Other quantization schemes along with Huffman encoding
image compression, the threshold value is set to zero. technique may be employed in order to obtain higher
However, in order to perform lossy image compression, the compression ratios. These hybrid techniques provide better
threshold value is set to positive value. Hence, the optimal compression ratios along with superior values of bit rates
and PSNR.

1234
 However, in preference to make use of spatial values. This technique uses DWT for compression. Here,
based, the techniques based on the histogram approaches the image is first sub divided into sub bands of 16. Further,
can be used in order to compress the digital image. In this the entropy of each sub band is calculated and priority is
scheme, first the original image is sub-divided into brittle assigned to every sub band based on these entropy values.
clusters of probabilistic divisions (threshold) base on the Based on this priority the compression is performed on the
image histogram. Furthermore, for each division the selective sub bands only instead of all sub bands. The sub
threshold values are calculated using the famous Shannon's bands with lower entropy values are highly intended for
entropy theorem. This theorem enables us to measure the compression in order to make the technique energy efficient.
randomness of the brittle clusters. As we keep on increasing However, as the distance increases the compression of more
the number of thresholds, the image quality and sub bands is needed. This technique was found to save the
compression ratios keep on improving. This occurs mainly energy up to 28.26% as compared to the fully compressed
because of precise accuracy of histogram values and hence scheme. The only limitation of this scheme is the
the diminished compression errors at the cost of increased requirement of ideal condition as after more distance, the
complexity and computations. In order to achieve the fully compressed schemes are proficient.
optimal results, the Shannon’s entropy function was
maximized and incorporated with new optimization scheme In [17] author proposed a new cost effective,
of differential evolution in [13]. simple iterative threshold based scheme in order to perform
segmentation of brain tumor and to isolation of tumor area
Discrete Anamorphic Stretch Transform (DAST) from non region of interest in MRI images. Since MRI
based technique was proposed for image compression in images are very large in size, hence storage is the critical
order to achieve higher image space-bandwidth compression task. Hence, non region of interest of image are going to be
in [14]. This technique does not use feature detection which compressed using simple enhanced differential pulse code
means the algorithm is carried out without the prior modulation transform (EDPCMT) along with the standard
knowledge of the image. However, the Discrete Huffman coding. Furthermore, the uncompressed tumor
Anamorphic transform is going to reshape the input image images along with the compressed non region of interest
before performing the uniform sampling. This sampling is images are transmitted across the network which are going
performed in such a manner that the sharp and fine features to reduce the storage space and hence bandwidth efficient.
incorporate a higher sampling density than the coarse or However, the complexity of the implemented compression
generic features. The extensive use of this feature-selective scheme is found to be more than lossless JPEG.
stretching enables us to allocate the additional samples to Furthermore, the implementation of iterative threshold
available sharp features where they are mostly required and scheme with the proposed scheme provides the improved
very few to the coarse features where they are redundant or segmentation results.
not required. This scheme is non-iterative in nature and can
be employed by itself. This scheme can further be combined Along with these discussed compression
with other types of compression schemes in order to techniques there are few more schemes which have been
improve the compression ratios further. proposed and proven to be efficient in different aspects.
These are hiding and compression with image inpainting
In [15] author presents a fast zonal DCT based [18], Side match vector quantization (SMVQ) [19],
technique in order to conserve energy in wireless image hardware compression scheme based on low complexity
sensor networks. The algorithm was developed using arithmetic encoding [20], DCT SVD and RLE based hybrid
combination of cardiac based DCT and zonal DCT. The image compression [21], binary discrete cosine and Hartley
cardiac based DCT is a standard multiplier less algorithm transforms based compression [22] etc.
which is used in 1D-DCT domain. In cardiac based DCT, 8
× 8 blocks are obtained after performing the block level III. PERFORMANCE PARAMETERS
DCT coefficients. This operation is carried out by first The performance parameters are used in order to
applying algorithm eight times over the rows and then eight justify the selection of the implemented image compression
times similarly over the columns. However, cardiac DCT algorithm to be efficient and prolific. The selection of these
with 8x8 processing requires less number of shift and parameters depends on the accuracy and types of
addition operations as compared to standalone cardiac - compression schemes. In this section we will discuss
DCT. Hence, by reducing the number of operation with the various types of performance parameters which can be used
extensive use of fewer DCT coefficients enables the accordingly in order to justify the efficiency of the
technique for reduction in energy consumption too. The implement compression technique. These parameters are as
only disadvantage of this technique is inferior image quality per following –
and hence less PSNR value.
In [16] author proposed new scheme for image
compression based on wavelet compression using entropy

1235
A. Compression Ratio (CR) image and the output or final image. Superior values of
PSNR result in the superior quality of the output image. For
The compression ratio is defined as shown in any image this can be calculated as shown in equation (4) or
equation (1). This is generally used in order to measure the (5).
capability of data compression. Here, comparison of size of
compressed image is performed with respect to the original ( )
image. As compression ratio increases, superiority of = 10 log (4)
compression technique enhances.
Where, = 255 i.e. the maximum pixel value
for any image. Hence equation (4) can be rewritten as –
= (1)
(255) 255
B. Mean Square Error (MSE) = 10 log = 20 log (5)

Mean Square Error (MSE) is defined as the E. Weighted Signal to Noise Ratio (WSNR)
cumulative squared error between the input image and the
compressed image. It is calculated as shown in equation (2). This performance parameter is going to use a
frequency domain transform function known as Contrast
1 Sensitivity Function (CSF). This is generally used in order
= ‖ ( , ) to filter spatially all irrelevant or inappropriate frequencies
∗ from the perspective of the human vision system. The
− ( , )‖ (2) computation of this parameter enables us to visualize and
quantify the effects of image dimensions, their printing
Where, Where M and N are the number of rows resolutions, screening or visual distance and the ambient
(height) and number of columns (width) and hence M x N is lighting and visualization. This parameter is also generally
the size of image and total number of pixels in image are expressed in terms of the logarithmic scale or decibel (dB)
M*N. ( , ) and ( , ) are the respective pixel values scale. In order to compute WSNR, first the error image is
for initial uncompressed image and final compressed image evaluated by taking the differences between original input
respectively. If this mean square error is equal to zero then image and the output noisy image. Next, the obtained error
both the input uncompressed image and the compressed image is weighted with the extensive use of a linear and
image are unerringly similar or identical. spatially invariant approximation to the respective
frequency response of the human visual system given by
C. Root Mean Square Error (RMSE)
CSF transformation. In the final step, the weighted signal to
The Root Mean Square Error (RMSE) is evaluated noise ratio parameter is going to be calculated as shown in
as shown in equation (3). This parameter is generally used equation (6).
in order to measure the differentiations between the sample
∑ | ( , ) ( , )|
and population data. On the whole, the RMSE is going to = 10 log (6)
represent the sample standard deviation of the differences ∑ | ( , ) ( , )|
between the original input uncompressed image and the
resulted output compressed image. Root mean square error F. Structural Similarity Index Measure (SSIM)
parameter is generally a superior quantification of accuracy. The structural similarity measure is used as an
The only limitation of this parameter is that it is used to image quality parameter in order to evaluate the output
compare only the forecasting errors of several models compressed image. This parameter is a measure in two
presented for any special type of variable rather than among separate signal vectors in x and y dimensions respectively.
the various types of variables due to its scale dependency. This parameter attempts to measure the similarities between
input and output images. This parameter is also known as
= √ (3)
full reference metric since this enables the measurement of
D. Peak Signal to Noise Ratio (PSNR) image quality which is generally based on an initial input
uncompressed or noise free image as reference image. This
This parameter is used to measure the effects of the parameter is incorporated in order to improvise the
noise on the quality of the signal. It is defined as the ratio of conventional performance parameters such as MSE and
the maximum possible value (power) of the signal with PSNR too due to their inconsistencies in case of human
respect to the power of interfering noise which is going to visual system. This key concept behind this parameter is
affect the quality of signal representation. This parameter is based on the inter-pixel dependencies in close spatial
generally expressed in terms of the logarithmic scale or proximities or neighborhood. These inter-pixel
decibel (dB) scale. This parameter is further used to dependencies are going to bring out the significant amount
measure the differences in quality between the original input of information about the structure of the objects from visual

1236
perspectives. This parameter is generally being calculated as Mohammad H.
Very less MSE
Increased coding
shown in equation (7) Superior PSNR
Asghari et. al. complexity
Less execution time
[14] Less CR
(2 Superior SSIM
+ )(2 + )
= (7) Low computational complexity
Low CR
( + + )( + + ) B. Heyne et. al.
High energy compaction
Low PSNR
[15] High MSE
Low processing power
G. Quality Index (QI) Low SSIM
High
Quality Index (QI) parameter is also defined very computational
complexity
similar to SSIM. This can be calculated as shown in A. Mittal et. al. High CR
Moderated PSNR
equation (8) [16] Low processing power
Moderated SSIM
Large memory
(4 ) requirements
= (8)
( + )( + )

H. Other Parameters V. CONCLUSION

In addition to the parameter discussed above few This paper has presented a widespread and up to
other parameters such as computational complexity, coding date survey on different image compression schemes. Since,
efficiency, power, entropy or randomness, bit rate, storage quality of reconstructed image is the most challenging issue
space, bandwidth, execution time etc. during the designing of compression techniques and
especially in case of bio-medical images such as CT and
IV. COMPARISON OF DIFFERENT MRI images. Moreover, the compression ratio and memory
COMPRESSION SCHEMES requirements are also equally significant in order to design
an efficient and prolific compression scheme. In order to
Table-I presents a comparative analysis of various
perform compression on dynamic MRI images different
compression schemes discussed and surveyed in this paper.
performance metrics are significant upon which the
The different compression techniques along with their pros
discussion has been carried out in respective sections. The
and cons are listed in this table. They are arranged in
summarized table presented in section 4 covers the different
chronological order in this table. The selection of particular
compression techniques along with their pros and cons. The
compression scheme depends on the choice of performance
enhanced values of PSNR, SSIM, compression ratio (CR)
parameters such as MSE, RMSE, PSNR, WSNR, CR,
also create the issues like increased complexity, overhead,
SSIM, QI, entropy etc.
coding complexity and cost etc. This survey will enable the
TABLE I COMPARISON OF DIFFERENT COMPRESSION author to design an optimized and more efficient
SCHEMES compression scheme for dynamic MRI imaging system.
Compression REFRENCES
Advantages Disadvantages
Technique
Extra calculation of [1] Jiji RS, Pollak AW, Epstein FH, et al. Reproducibility of rest and
Good CR threshold increases exercise stress contrast-enhanced calf perfusion magnetic resonance
Yi-Fei Tan et. al. Applicable for both lossy and complexity imaging in peripheral arterial disease. J Cardiovasc Magn Reson
[8] lossless compression techniques Optimal threshold 2013; 15:14.
Good PSNR criterion limits the [2] Versluis B, Backes WH, van Eupen MG, et al. Magnetic resonance
quality imaging in peripheral arterial disease: reproducibility of the
Increased assessment of morphological and functional vascular status. Invest
Good CR
complexity Radiol 2011; 46:11-24.
K. Rajakumar et. Optimized bit rates in case of
Calculation of sub [3] Isbell DC, Epstein FH, Zhong X, et al. Calf muscle perfusion at peak
al. [9] both with and without Run
bands increases the exercise in peripheral arterial disease: measurement by first-pass
Length Encoding
coding overhead contrast-enhanced magnetic resonance imaging. J Magn Reson
High complexity Imaging 2007; 25:1013-20.
Better CR
PSNR is not good [4] M. Hisatomi, J. I. Asaumi, Y. Yanagi et al., “Diagnostic value of
C. Better security due to
Single bit error of dynamic contrast-enhanced MRI in the salivary gland tumors,” Oral
Rengarajaswamy encryption
SPIHT Oncology, vol. 43, no. 9, pp. 940–947, 2007.
et. al. [11] Very compact output bit stream
More execution [5] Partovi S, Schulte AC, Jacobi B, et al. Blood oxygenation level-
with large bit variability
time dependent (BOLD) MRI of human skeletal muscle at 1.5 and 3 T. J
Moderated Magn Reson Imaging 2012;35:1227-32.
Better PSNR
Execution time [6] Partovi S, Schulte AC, Aschwanden M, et al. Impaired skeletal
Sujoy Paul et. al. Better SSIM
Moderated Entropy muscle microcirculation in systemic sclerosis. Arthritis Res Ther
[13] Better CR
Increased coding 2012; 14:R209.
Less MSE
complexity [7] Shen, S., Sandham, W., Granat, M., Sterr, A., “MRI fuzzy
segmentation of brain tissue using neighborhood attraction with
neural network optimization” IEEE transaction on Information
Technology in Biomedicine, vol. 9, 2005, pp. 459-467.

1237
[8] Yi-Fei Tan and Wooi-Nee Tan, “Image Compression Technique
Utilizing Reference Points Coding with Threshold Values,” IEEE, pp.
74-77, 2012.
[9] K. Rajakumar and T. Arivoli, “Implementation of Multiwavelet
Transform coding for lossless image compression,” IEEE, pp. 634-
637, 2013.
[10] S.Srikanth and Sukadev Meher, “Compression Efficiency for
Combining Different Embedded Image Compression Techniques with
Huffman Encoding,” IEEE, pp. 816-820, 2013.
[11] C. Rengarajaswamy and S. Imaculate Rosaline, “SPIHT Compression
of Encrypted Images,” IEEE, pp. 336-341, 2013.
[12] K. N. Bharath, G. Padmajadevi and Kiran, “Hybrid compression
using DWT-DCT and Huffman encoding techniques for biomedical
image and video applications,” International Journal of Computer
Science and Mobile Computing (IJCSMC), vol. 2, no. 5, pp. 255 –
261, 2013.
[13] Sujoy Paul and Bitan Bandyopadhyay, "A Novel Approach for Image
Compression Based on Multi-level Image Thresholding using
Shannon Entropy and Differential Evolution", Proceedings of the
IEEE Students Technology Symposium, IIT Kharagpur, West Bengal,
India, pp.56-61, Feb 2014
[14] Mohammad H. Asghari and Bahram Jalali , " Discrete Anamorphic
Transform for Image Compression", IEEE Signal Processing Letters,
Vol.2l, No.7, July 2014.
[15] B. Heyne, C. Sun and J. Goetze, “A computationally efficient high
quality cordic based DCT ,” IEEE 14th European Signal Processing
Conference, 2015, pp. 1 – 5.
[16] A. Mittal, C. Kundu, R. Bose and R. Shevgaonkar, “Entropy based
image segmentation with wavelet compression for energy efficient
LTE systems,” IEEE 23rd International Conference on
Telecommunications (ICT), 2016, pp. 1-6.
[17] H. R. Vilas, S. N. Kulkarni, H. Chiranth and M. Bhille,
"Segmentation and compression of 2D brain MRI images for efficient
teleradiological applications," 2016 International Conference on
Electrical, Electronics, and Optimization Techniques (ICEEOT),
Chennai, 2016, pp. 1426-1431.
[18] Chuan Qin, Chin-Chen Hiding and Compression Image Inpainting”
IEEE IMAGE PROCESSING, 2014
[19] S. M. Varghese, A. Johny and J. Job, "A survey on joint data-hiding
and compression techniques based on SMVQ and image
inpainting," 2015 International Conference on Soft-Computing and
Networks Security (ICSNS), Coimbatore, 2015, pp. 1-4.
[20] Ahmed Chefi, Adel Soudani and Gilles Sicard, "Hardware
Compression Scheme Based on Low Complexity Arithmetic
Encoding for Low Power Image Transmission Over WSNs",
international Journal of Electronics and Communications(A EU),
pp.193-200, August 2013.
[21] Raghavendra.M.J, Prasantha.H.S and S.Sandya, “Image Compression
Using Hybrid Combinations of DCT SVD and RLE”, International
Journal of Computer Techniques, Volume 2 Issue 5-2015.
[22] S. Bouguezel and O. Ahmad, “Binary discrete cosine and Hartley
transforms,” IEEE Trans. Circuits. Syst, pp. 1 – 14, 2013.

1238