Analysis of Different Image Compression Techniques: A Review

Garima Garga*, Raman Kumarb

a
Department of Computer Science and Engineering, IKGPTU, Kapurthala, Punjab, India, garimagarghce@gmail.com
b
Department of Computer Science and Engineering, IKGPTU, Kapurthala, Punjab, India, er.ramankumar@aol.in

Abstract: The availability of images in a wide variety of applications has expanded due to technological developments that have not to influence the variety
of image operations, the availability of advanced image modification software, or image management. Despite technological breakthroughs in storage and
transmission, demand for storage capacity and communication bandwidth exceeds available capacity. As a result, image compression has proven to be a
helpful technique. When it comes to image compression, we don't just focus on lowering size; we also focus without sacrificing image quality or information.
The survey outlines the primary image compression algorithms, both lossy and lossless, and their benefits, drawbacks, and research opportunities. This
examination of several compression techniques aids in the identification of advantageous qualities and the selection of the proper compression method. We
suggested some general criteria for choosing the optimum compression algorithm for an image based on the review.

Keywords: Image Compression, types of images, performance assessment metrics, compression techniques.
vector quantization for image compression. VQ is a multi-dimensional
version of Scalar Quantization [7].
1. Introduction
Zhang et al. (2020) proposed a multi-scale progressive statistical model-
An image is a two-dimensional communication processed by the human based lossless image compression system. The suggested statistical model
visual system. The impulses that depict images are usually analog. effectively balances pixel-wise model accuracy and multi-scale model
Computer applications convert them from analog to digital for processing, speed [8]. Mohammed and Abou-Chadi (2011) investigated picture
storage, and transmission [1]. A digital image is a 2D pixel array. Image compression techniques based on block truncation coding. For comparison
compression reduces the amount of storage space required for photos and purposes, the original block truncation coding (BTC) and the Absolute
movies, hence improving storage and transmission performance. Lossy or Moment block truncation coding (AMBTC) were used [9]. BTC breaks the
lossless image compression is possible. Lossless compression entails original image into n × n sub-blocks, reducing the number of grey levels
compressing data so that it may be decompressed into an identical within each block. Sarkar et al. (2018) proposed a hybrid lossy image
reproduction of the original [2-4]. However, in lossy compression compression model using run-length encoding and Huffman coding by
techniques, some of the image's finer details can be sacrificed in order to taking an example of two images i.e., clock.tiff and man.tiff. The other
save a little more bandwidth or storage space. parameters like PSNR, MSE, and structural similarity remain almost the
same while the storage size is reduced [10]. Table 1 summarizes the
Working procedure of image compression techniques: numerous image compression approaches that researchers in the past have
used.
The most common processes in compressing an image are [2]:
Table 1: Detailed Summary of various image compression techniques used
by researchers
▪ Specifying the Rate (available bits) and Distortion (tolerable Author Year Title of Paper Compression Reference
error) parameters for the target image. Name Technique
▪ Classifying the visual data according to their relevance. used
▪ Distributing the available bit budget across these classes in such Gharavi and 1988 “Sub-band coding Sub-band [5]
a way that distortion is minimized. Tabatabai of monochrome Coding (SBC)
▪ Using the bit allocation information acquired in step 3, quantify and color images”
each class independently. Patel et al. 2016 “A fast and Huffman [6]
▪ Using an entropy coder, encode each class independently and improved Image Encoding
Compression
save to a file. It is frequently faster to reconstruct an image from
technique using
compressed data than it is to compress it. The procedures are as
Huffman coding”
follows: Kekre et al. 2016 “Color image Vector [7]
▪ Using the entropy decoder, read the quantized data from the file. compression using Quantization
(Step 5 is reversed.) vector (VQ)
▪ Reduce the number of variables in the data. (Step 4 is reversed.) quantization and
▪ Re-create the image. (Step 2 is reversed.) hybrid wavelet
transform”
The following sections comprise this paper: section II explains the related
Zhang et al. 2020 “Lossless image Statistical [8]
work of researchers, section III defines the types of the different image
compression using Coding
compression techniques, and section IV concludes the paper by a multi-scale
summarizing the conclusion. progressive
statistical model”
Mohammed 2011 “Image Block [9]
2. Literature Review and Abou- compression using Truncation
Chadi block truncation Coding (BTC)
Gharavi and Tabatabai (1988) proposed utilizing QMF to encode digital coding
images. Using a 2-D separable QMF bank, the input signal spectrum is Sarkar et al. 2018 "Novel Hybrid Run Length [10]
decomposed into numerous narrowband images [5]. To reduce redundant Lossy Image Encoding
bits in data or images, Patel et al. (2016) introduced the Huffman coding Compression (RLE)
technique, which analyses multiple features or specifications such as “Peak Model using Run-
Signal to Noise Ratio (PSNR), Mean Square Error (MSE), Bits Per Pixel Length Coding
(BPP), and Compression Ratio (CR)” [6]. Kekre et al. (2016) proposed and Huffman
Coding"

Electronic copy available at: https://ssrn.com/abstract=4031725

 Li et al. 2018 “Efficient Arithmetic [11] were obtained in the same sequence using classification and decomposition.
trimmed Encoding
To reduce data storage, the mask image was then concealed in these
convolutional sequences. LZW was then used to encode the sequences. The algorithm was
arithmetic
simple and had a higher compression ratio than previous approaches like
encoding for
lossless image
LZW [17]. Fractal image compression was defined by Liu et al. (2016). The
compression” fractal compression technique works by comparing sections of an image to
DeVore et 1992 “Image Transform [12] other parts of a similar image. Fractal algorithms decode images by
al. compression Coding converting geometric forms into mathematical data known as "fractal
through wavelet codes" [18].
transform coding”
Cooper and 2006 “Image Singular Value [13]
Lorenc compression using Decomposition 3. Image Compression Techniques:
singular value (SVD)
decomposition” There are two types of image compression techniques: lossless and lossy
Telagarapu 2011 “Image Discrete Cosine [14] compression. Decompression of lossless data results in an image that is
et al. compression using Transform identical to the original. These processes obliterate data, resulting in an
DCT and wavelet (DCT) image that is not identical to the original.
transformations”
Leon-Salas 2007 “A CMOS imager Predictive [15] 3.1 Lossless Compression:
et al. with focal plane Coding
compression using The reconstructed image in lossless compression techniques is
predictive coding” quantitatively identical to the original image. Only a limited amount of
Chowdhury 2012 “Image Discrete [16] compression may be achieved with lossless compression. Lossless
and Khatun compression using Wavelet compression can cut the size of a file in half depending on the type of data
discrete wavelet Transform being compressed. As a result, lossless compression is advantageous for
transform” (DWT) delivering files over the Internet, as smaller files move more quickly.
Sun et al. 2008 “Image LZW Encoding [17] Generally, techniques for lossless image compression treat images as a
compression collection of pixels ordered in row-major order. Each pixel is processed
based on using two different procedures. The first step establishes a prediction for
classification row the next pixel's numeric value. This is often combined with an edge
by row and LZW detection technique that attempts to account for intensity discontinuities.
encoding” The difference between the anticipated and actual pixel values in the second
Liu et al. 2016 “A fractal image Fractal [18]
phase is coded using an entropy coder and a probability distribution. Figure
encoding method Compression
1 shows the block diagram of the lossless compression technique. There are
various types of lossless compression techniques, which are described in
based on statistical
detail one by one.
loss used in
agricultural image
compression” 3.1.1 Run Length Encoding
An alternative to lossy compression, RLE utilises a pair of (length, value)
A 3D binary cuboid was encoded using a trimmed convolutional network values to replace the original data. The value is unique and the length
(i.e., TCAE) by Li et al. (2018). To represent a vast context while keeping reflects the number of times it is repeated. This basic data compression
high computational efficiency, the fully convolutional network design may technique uses runs to store compressed data in smaller chunks for easier
benefit from reduced convolution and execute probability prediction to all retrieval later. Runs are patterns in which the same data value appears in
bits in a single forward pass [11]. DeVore et al. (1992) explained the multiple data items in a row. Rather of saving the original run, these
concept of image compression through transform coding technique with the sequences are preserved as a single data value and count. Consider the
help of mathematical expressions in detail [12]. Cooper and Lorenc (2006) following illustration: a screen with solid white writing on a basic black
focused on image compression through Singular Value Decomposition background In the unoccupied zone, there will be many long lines of white
(SVD). It is a valuable device for reducing data storage and transport. SVD pixels and many short runs of black pixels.
is a matrix factorization that allows you to extract algebraic and geometric
information from an image in a new way. Many fields have adopted SVD,
including data compression, signal processing, and pattern analysis [13].
On the other hand, Telagarapu et al. (2011) compressed the three images by
applying Discrete Cosine Transform (DCT). DCT represents the input data
as a sum of cosine functions with variable frequencies and magnitudes. The
most widely used DCTs are one-dimensional and two-dimensional [14].

To solve the problem of focal plane compression, Leon-Salas et al. (2007)

developed a single-chip approach. In this study, a predictive coding
technique was applied in the analog domain, followed by a compact entropy
coder on-chip. Because of its lossless compression capability and ease of Fig. 1: Block Diagram for Lossless Compression
use, predictive coding was chosen [15]. Chowdhury and Khatun (2012)
introduced the Discrete Wavelet Transform as a new approach (DWT). 3.1.2 Statistical Coding
DWT is a multi-resolution image compression transform. This method uses The statistical coding technique incorporates the following techniques such
sub-band coding to describe the signal's time-frequency [16]. Sun et al. as Huffman Encoding, .Arithmetic Encoding, and LZW Encoding [8].
(2008) suggested a near-lossless image compression algorithm that
combined categorization, information hiding, and LZW. Similar pixels (i) Huffman Encoding

Electronic copy available at: https://ssrn.com/abstract=4031725

In Huffman coding, a data item's frequency is used (pixel in images). The
goal is to encode data more frequently with fewer bits. The codes are kept 3.2.1. Transform Coding
in a Code Book for each image or collection of images. To enable decoding, Frequently, the original picture is divided into smaller sub-images (blocks)
the codebook and encoded data must be transmitted in all circumstances (usually 8 x 8). It transforms the eight pixel values into an array of
[6]. To encrypt images, follow these steps: First, divide the image up into 8 coefficients closer to the top-left corner, which typically includes the bulk
x 8 blocks, then each block is a symbol to be coded at the second step. of the information required to quantify and encode a picture with little
Compute the Huffman codes for a set of the block at the next step and perceptual distortion. The quantized coefficients are then utilized for
finally encode the blocks accordingly at the last step. encoding the image's bitstream. The decoder uses the same process, except
that the 'dequantization' stage only approximates the real coefficient values.
(ii) Arithmetic Coding In other words, the quantizer's loss in the encoder step is irreversible.
Rather than coding each symbol separately, this approach codes the entire
visual sequence with a single code. As a result, nearby pixels' association 3.2.2. Block Truncation Coding
is utilized. The following principle governs arithmetic coding [11]. The BTC is a grayscale picture compression technique [9]. It employs a
symbol alphabet is finite, as are all potential symbol sequences of a given quantizer to decrease the number of grey levels inside each block while
length; all conceivable sequences are countable infinite; and we may assign maintaining the mean and standard deviation. Although BTC compression
a unique subinterval to each given input since the range [0,1] includes an was used to compress color long before DXTC, it is an early ancestor of the
unlimited number of real values (sequence of symbols). popular hardware DirectX Texture Compression (DXTC) technology. If 8-
bit integer values are employed during transmission or storage, sub-blocks
(iii) LZW Coding of 4 × 4 pixels allow for around 25% compression. Larger blocks allow for
The LZW algorithm [17] is based on the number of character sequences in more compression, but due to the method's nature, quality suffers as block
the encoded string. Its theory is to gradually develop a dictionary by size grows.
substituting patterns with an index code. The ASCII table's 256 values are
used to populate the dictionary. For monochromatic images (coded in 1 bit), 3.2.3. Sub-band Coding
each string is compared to the dictionary and added if it is not found. That It divides the signal's frequency range into sub-bands and then codes each
is, if a string is never shorter than the dictionary's greatest word, it is sub-band with a coder and bit rate that fit the band's statistics [5]. This is
conveyed. The algorithm rebuilds the dictionary while decoding; thus it because SBC allows for variable bit assignment between sub-bands and
doesn't need to be stored. coding error confinement within sub-bands. The decoded sub-band signals
are un-sampled and routed through a bank of synthesis filters before being
3.1.3 Predictive Coding appropriately summed at the decoder.
Another example of inter-pixel redundancy research is the Predictive
Coding Technique [15], which uses the basic notion of encoding only new 3.2.4. Vector Quantization
information in each pixel. The difference between the pixel's actual and Vector quantization (VQ) techniques apply scalar quantization ideas to
intended value is frequently utilised to define this additional information. several dimensions. This method entails creating a code vector dictionary,
The prediction error is computed by rounding the predictor's output to the which is a collection of fixed-size vectors. After that, the image is separated
nearest integer and comparing it to the actual pixel value. Variable Length into image vectors, or non-overlapping pieces [7]. The dictionary's closest
Coding can be used to encode this error (VLC). The paradigm utilized to match vector is then discovered for each picture vector, and its index in the
describe the visuals is the method's distinguishing feature. The images are dictionary is used as the original image vector's encoding. VQ-based coding
represented as non-causal random fields, in which the intensity of each schemes are particularly well-suited for multimedia applications due to
pixel is determined by the intensities of sites in all directions around it. their decoder-side search capabilities.

3.2 Lossy Compression: 3.2.5. Fractal Compression

The fractal compression technique works by comparing regions of photos
Figure 2 shows how a lossy compression system may analyze colour data to other images. Algorithms convert these pieces, or geometric shapes, into
for a range of pixels and detect subtle variances in pixel colour values that "fractal codes" that may be used to replicate the encoded image. When
the human eye/brain couldn't distinguish differently. The computer may converted to fractal coding, an image loses its resolution-dependent
replace the others with smaller pixels whose color value disparities are relationship. The image can be reconstructed to fit any screen size without
within human vision. The pixels that have been finely graded are pixel-based compression artifacts or loss of quality [18].
subsequently discarded. This type of compression can result in large file
size savings, but the complexity of the algorithm dictates the image 3.2.6. Singular Value Decomposition (SVD)
superiority of the final product. Linear algebra is used extensively in data compression. In today's society,
the necessity to reduce the amount of digital data saved and communicated
is becoming increasingly important. Singular Value Decomposition [16] is
a valuable technique for minimizing data storage and transport. SVD is a
matrix factorization that allows you to extract algebraic and geometric
information from an image in a new way. Many fields have adopted SVD,
including data compression, signal processing, and pattern analysis [4]. The
goal of SVD is to find the best approximation of the original data points in
the smallest number of dimensions. This can be accomplished by
identifying regions with the highest degree of variability. SVD is used to
reduce a big, highly variable set of data points to a lower-dimensional space
that more clearly displays the original data's substructure and ranks it from
Fig. 2: Block Diagram for Lossy Compression highest to lowest variance. Using the SVD approach, this strategy locates

Electronic copy available at: https://ssrn.com/abstract=4031725

the most variable region and decreases its size. To put it another way, SVD improved Image Compression technique using Huffman coding. In 2016
is a data reduction approach. International Conference on Wireless Communications, Signal Processing
Numerous eminent mathematicians regard SVD as a crucial topic in linear and Networking (WiSPNET) (pp. 2283-2286). IEEE.
Kekre, H. B., Natu, P., & Sarode, T. (2016). Color image compression using
algebra. Apart from image reduction, SVD is useful in a wide variety of
vector quantization and hybrid wavelet transform. Procedia Computer
practical and theoretical contexts. The advantage of SVD is that it may be Science, 89, 778-784.
used to any real (m,n) matrix. It divides A into three matrices, U, S, and V, Zhang, H., Cricri, F., Tavakoli, H. R., Zou, N., Aksu, E., & Hannuksela, M.
in such a way that M. (2020). Lossless image compression using a multi-scale progressive
A = USVT statistical model. In Proceedings of the Asian Conference on Computer
S is a diagonal matrix, whereas U and V are orthogonal matrices. Vision.
Mohammed, D., & Abou-Chadi, F. (2011). Image compression using block
truncation coding. Cyber Journals: Multidisciplinary Journals in Science
3.2.7. Discrete Cosine Transform (DCT) and Technology, Journal of Selected Areas in Telecommunications (JSAT),
A DCT represents the input data points as a sum of cosine functions with February Edition.
various frequencies and magnitudes. The most prevalent DCTs are one- Sarkar, J. B., Poolakkachalil, T. K., & Chandran, S. (2018). Novel Hybrid
dimensional and two-dimensional. The Joint Photographic Expert Group Lossy Image Compression Model using Run-Length Coding and Huffman
(JPEG) was founded in 1992 on DCT [14]. It's a common compression Coding. International Journal of Computer Science and Information
method. This is the JPEG method: The initial block size is 8x8. Second, Security (IJCSIS), 16(10), 103-107.
Li, M., Gu, S., Zhang, D., & Zuo, W. (2018). Efficient trimmed
DCT is applied in a left-to-right and top-to-bottom orientation to each
convolutional arithmetic encoding for lossless image compression. arXiv
block. Then, to limit the amount of data in the memory, quantization is preprint arXiv:1801.04662, 107-120.
utilized, and data is stored in a precise manner. DeVore, R. A., Jawerth, B., & Lucier, B. J. (1992). Image compression
through wavelet transform coding. IEEE Transactions on information
3.2.8. Discrete Wavelet Transform (DWT) theory, 38(2), 719-746.
The DWT approach is a multi-resolution transform technique that is Cooper, I., & Lorenc, C. (2006). Image compression using singular value
decomposition. College of the Redwoods, 1-22.
frequently used to increase the compression ratio of images. This technique
Telagarapu, P., Naveen, V. J., Prasanthi, A. L., & Santhi, G. V. (2011).
uses sub-band coding to represent the signal's time and frequency Image compression using DCT and wavelet transformations. International
components. In DWT, a picture is represented by a collection of wavelet Journal of Signal Processing, Image Processing and Pattern
functions (also called wavelets) with varying locations and scale [16]. Recognition, 4(3), 61-74
Transform's high pass (detail) and low pass (approximate) coefficients Leon-Salas, W. D., Balkir, S., Sayood, K., Schemm, N., & Hoffman, M. W.
reflect discrete data. 3232 blocks are first partitioned. It divides the data (2007). A CMOS imager with focal plane compression using predictive
into approximation and detail coefficients. In the rectified matrices, the coding. IEEE Journal of Solid-State Circuits, 42(11), 2555-2572.
Chowdhury, M. M. H., & Khatun, A. (2012). Image compression using
coefficients are labelled. Then LL is translated to the second level. The
discrete wavelet transform. International Journal of Computer Science
compression ratio is obtained by multiplying the coefficients by a Issues (IJCSI), 9(4), 327.
predetermined scaling factor. Sun, M. Y., Xie, Y. H., & Tang, X. A. (2008, May). Image compression
based on classification row by row and LZW encoding. In 2008 Congress
on Image and Signal Processing (Vol. 1, pp. 617-621). IEEE.
4. Conclusion: Liu, S., Zhang, Z., Qi, L., & Ma, M. (2016). A fractal image encoding
method based on statistical loss used in agricultural image
As there is a trade-off between compression ratio and peak SNR in image compression. Multimedia Tools and Applications, 75(23), 15525-15536.
compression, developing a more effective compression-decompression
technique remains a difficult task in the area. Though considerable research
has been conducted in this field, with the ever-increasing demand for low-
bit-rate compression methods, there is still room for new approaches and
the evolution of more efficient algorithms within existing methods. The
review demonstrates that the topic will continue to pique academics' interest
in the coming years. We have discussed various types of image
compression methods in this article. As a result, we discover that lossy
compression provides a higher compression ratio than lossless
compression. Compression without loss of data is optimal for text
compression. When images have a bit depth greater than 0.5 bpp, all lossy
compression algorithms have a high compression ratio. Additionally, image
compression is highly dependent on the image's quality.

References

Sindhu, M., & Rajkamal, R. (2009). Images and its compression

techniques-A Review. International Journal of Recent Trends in
Engineering, 2(4), 71.
Arora, K., & Shukla, M. (2014). A comprehensive review of image
compression techniques. International Journal of Computer Science and
Information Technologies, 5(2), 1169-1172.
Pancholi, B., Shah, R., & Modi, M. (2014). Tutorial review on existing
image compression techniques. Int. J. Eng. Comput. Sci, 3(8), 7882-7889.
Singh, A. P., Potnis, A., & Kumar, A. (2016). A Review on Latest
Techniques of Image Compression. International Research Journal of
Engineering and Technology (IRJET), 3(7), 2395-0056.
Gharavi, H., & Tabatabai, A. (1988). Sub-band coding of monochrome and
color images. IEEE Transactions on Circuits and Systems, 35(2), 207-214.
Patel, R., Kumar, V., Tyagi, V., & Asthana, V. (2016, March). A fast and

Electronic copy available at: https://ssrn.com/abstract=4031725