Lossy compression of color images based on discrete

Tchebichef transform

N. Kouadria1, I. Mansri1, S. Harize1, N. Doghmane1, K. Mechouek1

1
Laboratory of Automatic and Signals of Annaba (LASA), Electronics Department, Badji Mokhtar University-ANNABA, Sidi
Amar, Annaba 23000, Algeria
nasreddine.kouadria@univ-annaba.org

Abstract—The Discrete cosine transform (DCT) plays an it leads to a high decorrelation of coefficients and (2) it
important role in image and video coding. It is used as the first guarantees good compaction of energy.
step in lossy / lossless image compression. Consequently, many If we compare the complexity of different transformations
algorithms have been developed to efficiently compute it. In the archived in the literature, we find that the Haar transform is the
image compression community, an alternative to the DCT, known
best choice. Nevertheless, this is achieved through a degradation
as the discrete Tchebichef transform (DTT), is increasingly used.
It possesses an excellent energy compaction property and a low of the quality. In this paper, we looked at the transformation
implementation complexity. However, the DTT has not been which gives the best trade-off between the low complexity and
investigated in the compression of color images. In this paper, the high quality. The energy compaction capability of a linear
theory behind the DTT and its pruned version (PDTT) are transformation used in a lossy compression scheme is measured
investigated and a lossy compression method of color images through the coding gain and represents for this purpose the most
based on the DTT is proposed. Experimental results demonstrate important property. Table 1 from [8] shows the Cg for several
that the proposed method exhibits almost the same performance famous transformations. From this table it is shown that the
as the DCT-based technique, in terms of objective measures such DTT is the best candidate for lossy image compression,
as the PSNR and SSIM while reducing the arithmetic complexity
compared to Walsh-Hadamard transform (WHT), Haar, Signed
by more than 37 %.
DCT (SDCT) and Binary DCT (BDCT). The considerable DTT
Cg (equal to 8.6756 for a correlation coefficient ρ = 0.95) will
I. INTRODUCTION give a high PSNR when put in a lossy compression chain.
The permanent use of the discrete cosine transform (DCT) in
multimedia application, especially in audio, image and video Table 1: Coding gain for several transforms. Results are given
compression techniques [1], has engendered the development of for a correlation coefficient ρ = 0.95.
many fast algorithms for its computation. Among these DCT DTT WHT BDCT[9] Haar SDCT[10]
algorithms, the Loeffler DCT [2] achieves the minimal Cg(dB) 8.8257 8.6756 7.9461 7.9461 7.9425 7.7905
theoretical limit of the algorithmic complexity of the exact 8-
point DCT but it is still insufficient. The aforementioned transforms have been integrated into a
In recent years, the Discrete Tchebichef Transform (DTT) JPEG like simulation; the PSNR has been used as a distortion
[3] is increasingly used as an alternative method to the DCT due metric. The results depicted in the figures bellow show a Rate-
to its low algorithmic complexity. It has already been used in Distortion (RD) comparison for a set of test images according to
several image processing applications such as image analysis, the number of retained coefficients (a transform block contains
artifacts measurement, lossy compression of grayscale images 64 coefficients). Having a PSNR very close to the DCT, it can
and videos coding [4-5] and recently in lossless compression of be concluded from these curves that the DTT is the best
images [6]. Unfortunately, the lossy compression of color candidate for low complexity image compression.
images using the DTT has not been investigated.
In this paper, we investigate the lossy compression of color
images using the DTT and we propose a new color compression
technique based on the pruned DTT [7]. The rest of the paper is
organized as follows: In section 2, a review of the DTT is
presented. In section 3, the proposed DTT-based color image
compression method is described. Results and discussion are
shown in section 4 and section 5 presents some concluding
remarks.
II. DISCRETE TCHEBICHEF TRANSFORM
The DTT is based on discrete orthogonal Tchebichef
polynomials and has almost the same properties as the DCT: (1) Figure 1. RD curve for Cameraman image

978-1-7281-3896-1
 The algorithm for computing the matrix T × requires 44
additions and 29 shifts operations.
To further reduce the arithmetic complexity one can use the
pruned DTT (PDTT) [7]. It consists in computing only the top
left sub-matrix of size L×L (where L < 8) which contains most
of the block energy. The PDTT is expressed by equation (6):

Y × =C × ∙X × ∙C × (6)

Where the 4-pruned matrix is given by equation (7):

C × =S × ∙ × (7)
Figure 2. RD curve for Lena image
A. Modified JPEG quantization table
After a DTT stage, the DTT-domain coefficients are
quantized according to (8):
( )
Y, = round(Y , Ø Ԛ) (8)

( )
Where Y , , is the quantized DTT-domain
coefficients,round ∙ is the rounding function, Ø denotes the
( )
element-by-element division, Q is the JPEG standard
quantization table.
The DTT kernel C is given by the product of a diagonal
Figure 3. RD curve for Boat image matrix (S) and an integer matrix T as follows:

The forward and inverse 2-D DTT transforms for 8×8 matrix C =S∙T (9)
can be expressed according to equation (1) and (2), respectively.
The multiplicative nature of S can be eliminated from the
Y × =C × ∙X × ∙C ×    transformation step by integrating its entries into the
quantization matrix as follows:

X × =C × ∙Y × ∙ C ×   
Y , = (S ∙ T) ∙ X , ∙ (S ∙ T) = (T ∙ X , ∙ T ) ʘ (s ∙ s ) (10)

Matrix × can be expressed by equation (3)

Where s, is a column vector corresponding to the main
diagonal S. Hence, the combined transform and quantization
C × =S × ∙ ×    pair is given (11):

Where × , is a diagonal scaling matrix and × is an integer ( )
matrix. Y, = round[( T ∙ X , ∙ T ) Ø Q ] (11)

From [3], × and × are given by (4) and (5) as follows: The modified quantization table is given by (12):

Q = Q Ø(s ∙ s ) (12)
S × = , , , , , , ,   
√ √ √ √ √ √ √ √
We note that the equations of the inverse process can be
 derived using the same methodology.
1 1 1 1 1 1 1 1 III. DTT-BASED COLOR IMAGE COMPRESSION
⎡−7 −5 −3 −1 1 3 5 7⎤
⎢7 1 −3 −5 −5 −3 1 7⎥ The proposed method starts by encoding each color plane
⎢ ⎥
−7 5 7 3 −3 −7 −5 7⎥ following the steps shown in Fig. 4. After an 8×8 block
T × =⎢   
⎢7 −13 −3 9 9 −3 −13 7⎥ partitioning, each block is transformed using a DTT-based
⎢−7 23 17 −15 15 17 −23 7⎥ transform. The transformed samples are then uniformly
⎢1 −5 9 −5 −5 9 −5 1⎥ quantized by a modified JPEG quantization table. The
⎣−1 7 −21 35 −35 21 −7 1⎦ quantized data is scanned following a zigzag path. The
resulting vector is then input to Huffman coder. Due to the high

978-1-7281-3896-1
energy compaction property, the DTT samples are supposed to Table 2. Computational complexity given in terms of
be small for high frequencies and most are set to zero by the number of operations for 8by8 blocs.
quantization process. In essence, the more zero-values
quantized samples there are in the zigzag scanned vector, the Methods Add Mul (float) Shifts
higher is the compression. JPEG-DCT[2] 1344 528 0
JPEG-DTT[3] 2112 0 1392
The first step in our method is the RGB-YCbCr color Proposed 1328 0 800
transformation. This step aims to reduce the correlations
between the RGB components. Doing so, we enhance the Table 3. Comparison between the energy consumption of
compression efficiency. Each pixel is first transformed the proposed method, JPEG-DCT [2], and JPEG-DTT[3].
separately to remove color redundancy, and then a transform is
applied to the whole image to reduce the spatial redundancy. Energy for 8×8 Energy for 512×512
The subsampling step is ignored and the 2-D DTT-based blocs ( J) images (mJ)
transform is applied directly on 8×8 blocs. It can be carried out JPEG-DCT[2] 21,85 89,50
with row-column decomposition by performing 1-D transform JPEG-DTT[3] 11,55 47,31
to each row of the pixels block and then 1-D transform to each Proposed 7,01 28,71
resulting column of 1-D coefficients. Since we are less
sensitive to high frequency in chroma components, one could To further prove the low complexity of the proposed method
apply the PDTT from start, on Chroma blue (Cb) and Chroma we have compared the aforementioned methods in terms of
red (Cr), without loss of fidelity. On this basis, in our method energy consumption. For the computation of the energy
we have applied the DTT to the luma component and the consumption, the characteristics of Mica2 sensor node, a real
PDTT [7] to the chroma components. example of a wireless sensor network node (WSN) frequently
used, have been used and the model presented in [7] has been
adopted. Table 3 represents a comparison between the energy
IV. RESULTS AND DISCUSSION consumption of the proposed method, JPEG-DCT [2] and
JPEG-DTT[3].
A. Complexity assessment
The computational complexity of the proposed color
compression method is expressed as a number of elementary
operations, and compared with the complexity of the color
compression techniques corresponding to JPEG-DCT [2] and
JPEG-DTT[3]. Since the transformation step takes the largest
part of the computational cost in image compression, it is
assumed that the complexity of each compression method is
taken by the transformation and hence the complexity of the
quantization and the entropy coders is neglected. Adopting the
row-column method to compute each 2-D transforms, number
of arithmetic operations required by the proposed method,
JPEG-DCT[2] and JPEG-DTT[3] to compute the luma and the
chroma components is shown in Table 2.
From this table, we observe that the proposed method is of Figure 5. Average PSNR against bitrate for the proposed
much less complexity than the two other compression schemes. method, JPEG-DCT [2] and JPEG-DTT[3].
It requires 37% and 42% less addition and shift operations
respectively, compared with JPEG-DTT [3]. Furthermore, it
does not involve any float multiplications, which are
computationally intensive, as it is needed for the JPEG-DCT
[2].

Figure 6. Average SSIM against bitrate for the proposed

Figure 4. DTT-based color image compression method, JPEG-DCT [2] and JPEG-DTT[3].

978-1-7281-3896-1
B. Image compression V. CONCLUSION
This section compares the compression efficiency of the In this paper, a lossy compression technique of color images
proposed method, JPEG-DCT [2] and JPEG-DTT[3] as rate- based on the discrete Tchebichef transform is proposed. The
distortion curves (R-D). PSNR and SSIM have been used as proposed method uses the DTT for compressing the luma
metrics for distortion. For the bitrate, a range of [0.3-1.5] bpp component and the PDTT for compressing the chroma
which correspond to high compression ratio (> 95 %) is components. Simulation results show that the image
selected. To do so, we used ten standard 512×512 and 256×256 degradations introduced by the proposed compression technique
color images (each 24-bits/pixel(bpp)), that were taken from the are negligible while the computational complexity has been
Waterloo Repertoire (http://links.uwaterloo.ca/Repository.html). reduced by more than 37% in comparison to the DCT and the
From the PSNR curves in Fig. 5, it can be seen that up to 1 DTT based algorithms. This technique can be a viable
bpp the proposed method has the same performance as JPEG- alternative to JPEG and it could be used in resource-limited
DCT [2] and JPEG-DTT [3]. At bitrates higher than 1 bpp, the WSN and embedded systems, requiring low energy.
performance of the proposed method tends to be slightly lower.
On average, a degradation less than 0,15 dB is noticed. From
SSIM results shown in Fig 6, we observe that the difference REFERENCES
between the three methods gets closer for the whole range of [1] S. Harize, A. Mefoued, N. Kouadria, and N. Doghmane, “HEVC
bitrates. Fig. 7 shows compressed color images at bitrate of 1 transforms with reduced elements bit depth,” Electronics Letters, IET,
54(22), pp. 1278 -1280, 2018
bpp with the aforementioned methods. Resulting images are
[2] C, Loeffler, A, Ligtenberg, and G.S. Moschytz, “Practical fast 1-D DCT
almost indistinguishable. algorithms with 11 multiplications,” Acoustics, Speech, and Signal
Process, pp. 988–991, May 1989.
[3] S, Prattipati, S, Ishwar, MNS, Swamy, and PK, Meher, “A fast 8by8
integer Tchebichef transform and comparison with integer cosine
transform for image compression, ” 56th IEEE International Midwest
Symposium on Circuits Sys (MWSCAS), pp. 1294–1297, 2013.
[4] V.A, Coutinho, R.J, Cintra, F.M, Bayer, P.A. Oliveira, R. S, Oliveira, and
A, Madanayake, “Pruned Discrete Tchebichef Transform Approximation
for Image Compression,” Circuits, Systems, and Signal Processing,
Springer, 37(10), pp. 1-21, 2018.
[5] B.Lu.G, Xiao, Y.Li.W Zhang, G, Wang, “Lossless image compression
based on integer Discrete Tchebichef Transform,” Neurocomputing,
Elsevier, 214, pp.587-593, 2016.
[6] G, Paim, L.M.G, Rocha, G.M, Santana, L.B, Soares, E.A.C, da Costa,
and S, Bampi, “Power-, Area-, and Compression-Efficient Eight-Point
Approximate 2-D Discrete Tchebichef Transform Hardware Design
Combining Truncation Pruning and Efficient Transposition Buffers,”.
IEEE Transactions on Circuits and Systems I: Regular Papers, pp. 1-14,
2018.
[7] N, Kouadria, K, Mechouek, D, Messadeg, and N, Doghmane, “Pruned
discrete Tchebichef transform for image coding in wireless multimedia
sensor networks,” AEU Int. J. Electron. Commun, Elsevier, 74, pp. 123-
127, 2017.
[8] N, Kouadria, K, Mechouek, K., S, Harize, and N, Doghmane, “ Region-
of-interest based image compression using the discrete Tchebichef
transform in wireless visual sensor networks”. Computers & Electrical
Engineering, Elsevier, 73, pp.194-208, 2019.
[9] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, “Binary discrete
Figure 7. Compressed images at bitrate of 1 bpp. From left to cosine and Hartley transforms,” IEEE Trans. Circuits Syst., 60,(4), pp.
989–1002, 2013.
right: JPEG-DCT [2], JPEG-DTT [3] and Proposed
[10] T. I. Haweel, “A new square wave transform based on the DCT,” Signal
processing, 81(11), 2309-2319. 2001

978-1-7281-3896-1