International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056

Volume: 04 Issue: 01 | Jan -2017 www.irjet.net p-ISSN: 2395-0072

Image Compression using DPCM with LMS Algorithm

Reenu Sharma, Abhay Khedkar
SRCEM, Banmore
-----------------------------------------------------------------****---------------------------------------------------------------

Abstract: The Differential pulse code modulation (DPCM) [1] 1.1

may be used to remove the unused bit in the image for image In figure 1: Where Q=Quantizer, is the nth input
compression. In this paper we compare the compressed image sample, is the predicted value, and is the
for 1, 2, 3, bit and also compare the estimation error. The LMS
associated prediction error. This is performed in the
[2] Algorithm may be used to adapt the coefficients of an
predict-and-compare loop, the loop shown in figure 1.
adaptive prediction filter for image source coding. In the
Its prediction by forming the sum of its prediction and
method used in this paper we decrease the compressed image
the prediction error
distortion and also the estimation error. The estimation error is
reduced as much as 7-8 dB using DPCM with LMS Algorithm.
1.2
Key Words: - Adaptive filter, LMS algorithm, DPCM,
Quantization. Where quant (.) represents the quantization operation,

1. INTRODUCTION is the quantization [4] version of the prediction

error, and is the corrected and quantized version
In a communication environment, the difference
of the input sample. This is performed in the predict-
between adjacent time samples for image is small,
and-correct loop.
coding techniques have envolved based on transmitting
sample-to-sample differences rather than actual sample
+
value. Successive differences are in fact a special case of
Q
a class of non-instantaneous converters called N-tap
linear predictive coders. These coders, sometimes Predict and compare loop
called predictor-corrector coders, predict the next input
sample value based on the previous input sample +
LMS
values. This structure is shown in figure 1. In this type Predictor
of converter, the encoder forms the prediction error (or
the residue) as the difference between the next Predict and correct loop +
measured sample value and the predicted sample value.
The equation for the prediction error [3] is Figure 1 Basic Block diagram of DPCM with LMS
Algorithm image compression system

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 824
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 01 | Jan -2017 www.irjet.net p-ISSN: 2395-0072

quantized samples difference

e(n)=x(n)-y(n) is now transmitted via PCM. At the
50
receiver, we can generate and from the received

100
we can reconstruct Figure 1 shown a DPCM
predictor. We shall soon show that the predictor input
150
is Naturally, its output is the predicted value

200
of The difference of original image data, and
prediction image data, is called estimation
250
50 100 150 200 250
residual, . So
2.1
Figure 2 Original image is quantized to yield

The communication task is that of transmitting the

Where is the quantization error,
difference (the error signal) between the prediction and
quantized signal. And
the actual data sample. For this reason, this class of
2.2
coder is often called a differential pulse code modulator
(DPCM) [3]. If the prediction model forms predictions The prediction output is fed back to its input so

that are close to the actual sample values, the residues that the predictor input is

variance (relative to the original signal).

2. ANALYSIS OF DPCM 2.3

This shows is quantized version of The
In DPCM [3] we transmit not the present sample x(n),
prediction input is indeed , as assumed. The
but e(n) (the difference between x(n) and its predicted
quantized signal is now transmitted over the
value y(n)). At the receiver, we generate y(n) from the
channel.
past sample value to which the received x(n) is added to
generate x(n). There is, however, one difficulty
3. IMAGE COMPRESSION USING DPCM AND
associated with this scheme. At the receiver, instead of
LMS ALGORITHM
the past samples as well as we
have their quantized version this A block diagram of the LMS adaptive image
will increase the error in reconstruction. In such a case, compression system is shown in figure 1. It is seen that
a better strategy is to determine the estimate of the image prediction is formed in a linear manner
(instead of ), at the transmitter also from the at the output of the LMS filter:

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 825
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 01 | Jan -2017 www.irjet.net p-ISSN: 2395-0072

3.1 between the original discrete image x(n) and the

reconstructed value y(n) at the receiver is given by
3.7
3.2
(Assuming the no channel-induced errors)
In equation 3.2, the are N adaptive predictor Therefore, if the goal of the system is an accurate
coefficients, the are the reconstructed image data, reconstruction of the image, then an algorithm is
and k is 1, 2.N integer values which select the desired which will form an accurate so that e(n)
previous image pixel on which base the current will have smaller variance and the quantizer levels may
prediction. At each scanned pixel a prediction residual be adjusted to give a smaller quantization error.
(error), is computed Hence, a lower reconstruction error, or distortion, will
be present at the receiver. The quantizer levels
3.4 themselves may be fixed or may vary as some function
of the residual sequence . Although, in general, the
This quantized residual is send to the receiver. The
position of the quantizer levels could be adaptive, for
quantization residual is determine
simplicity, in this correspondence we only examine the
3.5 case of a quantizer with fixed levels.
Alternatively, if the goal of the system is to reduce the
This residual is then quantized to form and the
bit rate over the channel subject to some distortion
quantized residual is also used to update the predictor
criteria, then we may reduce the number of quantizer
coefficient for the next iteration by the well known least
levels which span the residual signal range and, hence,
mean squares (LMS) [5] algorithm.
produce shorter code words per level. In this situation

3.6 the LMS adaptive predictor reduces the average

The parameter is known as the step size parameter number of bits per image while maintaining an

and is a small positive constant, which control steady- acceptable visual appearance at the receiver.

state and convergent mean-square residual

4. SIMULATION RESULT
characteristics of the predictor. The LMS algorithm is an
approximation to the gradient search method for In this paper we use 256256 image in figure 2 were
iteratively computing the N optimal coefficients used in experimental work to illustrate the
which minimize the mean square prediction residual. It performance advantages of using LMS [7] as an
is known by [6] that the error between the original adaptive predictor. The image of figure 2 was processed
image and the reconstructed image at the receiver is with the residual quantizer consistiting of b=1, 2 and 3
simply the quantization error Thus, the distortion bits (2, 4 and 8 quantization levels respectively) the

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 826
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 01 | Jan -2017 www.irjet.net p-ISSN: 2395-0072

DPCM image quantization [8], [9]. The dynamic range of

data was eight bits from grey level 0 to 255. The figure 50

3 plots the average square distortion versus

transmitted bit rate for the woman image. All values of 100

average square error are in dB referenced to the

150

performance of the 1bits/pixel fixed coefficient

predictor. The bit rate is in bits/pixel and is controlled 200

by the number of levels in the quantizer. If number of

250
bit increasing and distortion will be decrease. Figure 7, 50 100 150 200 250

8, and 9 is shown the prediction mean square versus Figure 4 1bits/pixel LMS images
gray level respectively for 1, 2, and 3 bits reconstructed
image. And figure 10 is shown the comparison of PMSE.
If the number of bits is increasing then PMSE will be 50

decreasing.
100

Table 1 Condition in Simulation Experiment

150
Image Matrix size 256256
No of Filter Taps 110
No of Bits 1, 2, 3 bits 200

Quantization level 2, 4, 8 quantizer

LMS Parameter 250
Average square distortion versus transmitted bit rate 50 100 150 200 250
-22.9
LMS

-22.95
Figure 5 2bits/pixel LMS images
Average Square Distortion[dB]

-23

-23.05

-23.1
50
-23.15

-23.2
100

-23.25
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
bit/pixel
150

Figure 3 average square distortions versus transmitted

bit rate. 200

250
Figure 650 3bits/pixel
100 LMS
150 images
200 250

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 827
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 01 | Jan -2017 www.irjet.net p-ISSN: 2395-0072

PMSE of LMS Algorithm Comparision of PMSE For 1,2,3 Bits

-25
-26
1bit -30
-28
-35
-30
-40

-32 -45

PMSE [dB]
-34 -50
PMSE [dB]

-55
-36
-60
-38
-65 1bit
-40 -70
2bit
3bit
-42 -75
0 50 100 150 200 250 300
sample number
-44

-46
0 50 100 150
sample number
200 250 300
Figure 10 PMSE [dB] versus Sample number for 1, 2, 3
bits/sample comparison.
Figure 7 PMSE [dB] versus Sample number for
1bits/sample 5. CONCLUSSION
The LMS is a simple and robust adaptive algorithm and
-30
PMSE of LMS Algorithm
DPCM use the LMS for prediction. At last the distortion is
reduce for 1, 2, 3 bits and also reduce the estimation
2bit

-35

-40
mean square error. The distortion and the estimation
mean square error is very less. We compare the
PMSE [dB]

-45

estimation mean square error in dB. This difference is 7-

-50

9 dB respectively for 1, 2, 3 bits as shown in figure 10

-55
and the reduce image shown in figure 4, 5, and 6
respectively this work carried out in future also.
-60

-65
0 50 100 150 200 250 300

REFERENCES
sample number

Figure 8 PMSE [dB] versus Sample number for

2bits/sample. [1]. A. Habbi, Comparison of Nth-order DPCM encoder
with linear transformation and block quantization
-30
PMSE of LMS Algorithm
techniques, IEEE Trans. Commun., vol. COM-19, pp. 948-
3bit
-35 956, Dec. 1971.
-40

-45 [2]. S.Haykin and T.Kailath Adaptive Filter Theory

Fourth Edition. Prentice Hall, Pearson Eduaction 2002.
PMSE [dB]

-50

-55

-60
[3]. B. P. Lathi and Zhi ding Modern Digital and Analog
Communication Systems International Fourth Edition.
-65

-70

-75
New York Oxford University Press-2010, pp.292.
0 50 100 150 200 250 300
sample number

[4]. J. E. Modestino, and D. G. Daut, Source-channel

Figure 9 PMSE [dB] versus Sample number for coding of images, IEEE Trans. Commun., vol. COM-27,
3bits/sample. pp. 1644-1659, Nov. 1979.

[5]. J. R. Zeidler et al., Adaptive enhancement of

mulyiple sinusoids in uncorrelated noise, IEEE Trans.

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 828
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 04 Issue: 01 | Jan -2017 www.irjet.net p-ISSN: 2395-0072

Acoust., Speech, Signal Processing, vol. ASSP-26, pp. 240- [8]. W. K. Pratt, Digital Image Processing. New York:
254, June 1978. Wiley, 1978.

[6]. J. G. Prokis, Digital Communications. New York: [9]. J. E. Modestino, and D. G. Daut, Source-channel
McGraw-Hill, 1983. coding of image, IEEE Trans. Commun., vol. COM-27, pp.
1644-1659, Nov1979.
[7]. S. T. Alexander and S. A. Rajala, Analysis and
simulation of an adaptive image coding system using the
LMS algorithm, in Proc. 1982 IEEE Int. Conf. Acoust.,
Speech Signal Processing, Paris, France, May 1982.

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 829