1
Gaussian Golomb Codes Seishi TAKAMURA and Yoshiyuki YASHIMA NTT Cyber Space Laboratories, NTT Corporation Y-517A 1-1 Hikarino-oka, Yokosuka, Kanagawa, 239-0847 Japan Email: {takamura.seishi, yashima.yoshiyuki}@lab.ntt.co.jp Gaussian sources are observed in vast wide areas. Contrarily to what happens for the geometric and double-sided geometric distributions, there is no simple, instantaneous code for the normal distribution. This paper tackles this problem by mapping the normal distribution into the geometric distribution before applying Golomb codes, which is optimal for geometric distributions. In our mapping, a pair of normally-distributed i.i.d. integers (say ( x, y)) is concatenated and then mapped to one natural number z( x, y). The conditions that z shall satisfy are: min x,y z( x, y) = 0, l( x, y) < l(a, b) z( x, y) < z(a, b) and z( x, y) = z(a, b) ( x, y) = (a, b), where l( x, y) is an arbitrary distance measure between the origin and the grid point ( x, y), such as the Euclidean norm. Fig. 1 is an example of such a mapping. When the signal is correlated, l( x, y) = x 2 - 2ρ xy + y 2 , where ρ is the autocorrelation of the source, is suitable. The mapping can be easily obtained using a computer program. In addition, if the upper- and lower- bounds of the source is known, pre-calculated mapping table can be stored in the memory because it is independent of source statistics. Of course, this table is not needed to be downloaded / transmitted. After this mapping, z is made geometrically-distributed and conventional Golomb codes can be eciently applied. Coding eciency comparison for our codes (Gaussian Golomb) and ordinary Golomb codes for quantized Gaussian sources is shown in Fig. 2. The unit-variance Gaussian dis- tribution is quantized with step size of d. The quantizer maps input values within each bin into integers. Our codes constantly yield better coding eciency, which is higher than 98%. In particular, the coding eciency for d < 0.1 area is stably higher than 99.5%. We also conducted the experiment on actual pixel value residual data of decoded video, which was observed to be quasi-Gaussian in Fig. 3. Our coding eciency is no worse than 90%, in average 94.1%. Conventional Golomb codes yield 87.4% in average, which is about 7 points worse. In addition, we propose an estimation method of optimal Golomb parameter using source variance to prevent exhaustive parameter search, which is verified to be practically ecient enough. 0 1 2 7 3 8 14 9 15 16 10 6 11 19 20 12 13 21 25 24 23 17 22 4 5 18 -3 -2 -1 0 1 2 3 3 2 1 0 -1 -2 -3 x y Figure 1: Mapping example of ( x, y) 7z 92 93 94 95 96 97 98 99 100 0.01 0.1 1 Coding efficiency (%) Quantization step size d Gaussian Golomb Golomb Figure 2: Coding eciency for quantized Gaussian source Figure 3: Coding e- ciency for actual data sets 2007 Data Compression Conference (DCC'07) 0-7695-2791-4/07 $20.00 © 2007

[IEEE 2007 Data Compression Conference (DCC'07) - Snowbird, UT, USA (2007.03.27-2007.03.29)] 2007 Data Compression Conference (DCC'07) - Gaussian Golomb Codes

Embed Size (px)

Citation preview

Page 1: [IEEE 2007 Data Compression Conference (DCC'07) - Snowbird, UT, USA (2007.03.27-2007.03.29)] 2007 Data Compression Conference (DCC'07) - Gaussian Golomb Codes

Gaussian Golomb CodesSeishi TAKAMURA and Yoshiyuki YASHIMA

NTT Cyber Space Laboratories, NTT CorporationY-517A 1-1 Hikarino-oka, Yokosuka, Kanagawa, 239-0847 Japan

Email: {takamura.seishi, yashima.yoshiyuki}@lab.ntt.co.jp

Gaussian sources are observed in vast wide areas. Contrarily to what happens for thegeometric and double-sided geometric distributions, there is no simple, instantaneous codefor the normal distribution.

This paper tackles this problem by mapping the normal distribution into the geometricdistribution before applying Golomb codes, which is optimal for geometric distributions.In our mapping, a pair of normally-distributed i.i.d. integers (say (x, y)) is concatenated andthen mapped to one natural number z(x, y). The conditions that z shall satisfy are:minx,y z(x, y) = 0, l(x, y) < l(a, b)⇒ z(x, y) < z(a, b) and z(x, y) = z(a, b)⇔ (x, y) = (a, b),where l(x, y) is an arbitrary distance measure between the origin and the grid point (x, y),such as the Euclidean norm. Fig. 1 is an example of such a mapping. When the signal iscorrelated, l(x, y) = x2 − 2ρxy+ y2, where ρ is the autocorrelation of the source, is suitable.The mapping can be easily obtained using a computer program. In addition, if the upper-and lower- bounds of the source is known, pre-calculated mapping table can be stored in thememory because it is independent of source statistics. Of course, this table is not needed tobe downloaded / transmitted. After this mapping, z is made geometrically-distributed andconventional Golomb codes can be efficiently applied.

Coding efficiency comparison for our codes (Gaussian Golomb) and ordinary Golombcodes for quantized Gaussian sources is shown in Fig. 2. The unit-variance Gaussian dis-tribution is quantized with step size of d. The quantizer maps input values within eachbin into integers. Our codes constantly yield better coding efficiency, which is higher than98%. In particular, the coding efficiency for d < 0.1 area is stably higher than 99.5%.

We also conducted the experiment on actual pixel value residual data of decoded video,which was observed to be quasi-Gaussian in Fig. 3. Our coding efficiency is no worse than90%, in average 94.1%. Conventional Golomb codes yield 87.4% in average, which isabout 7 points worse.

In addition, we propose an estimation method of optimal Golomb parameter usingsource variance to prevent exhaustive parameter search, which is verified to be practicallyefficient enough.

0 1

27

3

8

14 9

15

16

10

6

11

19

20

12

13

21

25

24 23

17

22

4 5

18

-3 -2 -1 0 1 2 3

3

2

1

0

-1

-2

-3

x

y

Figure 1: Mappingexample of (x, y) 7→ z

92

93

94

95

96

97

98

99

100

0.01 0.1 1

Cod

ing

effic

ienc

y (%

)

Quantization step size d

Gaussian Golomb

Golomb

Figure 2: Coding efficiencyfor quantized Gaussian source

707580859095100

data1 data2 data3 data4 data5 average

Coding Efficiency (%)

Gaussian Golomb Golomb

Figure 3: Coding effi-ciency for actual data sets

2007 Data Compression Conference (DCC'07)0-7695-2791-4/07 $20.00 © 2007