Professional Documents
Culture Documents
Delp 1979
Delp 1979
COMMUNICATIONS,
ON VOL. COM-27, NO. 9 , SEPTEMBER 1979 1335
fll(4) = 3.9
iTz(4) = 2.5 Coding
Truncation
Block
Image Using
Compression
so that the decision would be to send the packet to node 2 . EDWARD J. DELP, STUDENT MEMBER, IEEE, AND
(The exact values of 3, and 3 2 would yield the same deci- 0 . ROBERT MITCHELL, MEMBER, IEEE
sion.)
REFERENCES
Paper approved by the Editor for Data Communication Systems of
[ 1] A. Livne, “Dynamic Routing in Computer Communication Net- the IEEE Communications Society for publication after presentation at
works,”DoctoralDissertation,Dept. of ElectricalEngineering ICC’78, Toronto, Ont., Canada, June 1978. Manuscriptreceived Oc-
and Electrophysics, PINY, 1977 tober 30, 1978; revised March 22, 1979. This work was supported by
[2] C. W. Brown and M. Schwartz,“AdaptiveRouting in Central- the Rome Air Development Center, Griffiss AFB-RADC, under Grant
izedComputerCommunicationNetworks,” h o c . ZCC, San F75C0082 covering the period fromApril 1, 1977 to October 31, 1978.
Francisco, June 1975 The authors are with the School of Electrical Engineering, Purdue
[ 3 ] J. M. McQuillanand D. C.Walden,“TheARPANetwork De- University, West Lafayette, IN 47907.
sign Decisions,” Computer Networks, l 243-289 (1977) At the time of presentation of our ICC paper [ 51 ,a method similar
[4] T. S. Yumand M. Schwartz,“AnalysisofaClassofLocally- to BTC was presented in Japan by Kishimotoet al. [6] -[8].
x = -mI r2n xi
(1)
i=l i=m-q
where
Figure 2. Results using various fidelity criterion. All data rates are 2.0
l m bits/pixel.Upper left:minimummeansquareerror; Upperright:
b=- C Yi minimummeanabsoluteerror;
Lower right: minimum mean
Lowerleft: moment preserving;
square error and also assuming image
4 i=m-q
data uniformly distributed each block.
where
b = median of ( Y m W q ,--, Y m ) .
TABLE I
COMPUTED MEAN SQUARE ERROR AND MEAN ABSOLUTE
ERROR MEASURES FOR VARIOUS QUANTIZATION
SCHEMES
Mean Square E r r o r Mean Absolute
Error
b a t a from Figure 2:
Using MSE q u a n t i z e r 3.54 32.94
Using MAE q u a n t i z e r 37.13 3.28
Using BTC 40.89 3.91
Using
parametric MSE q u a n t i z e r 44.64 4.23
and assumed uniformdensity
Datafrom Figure 3:
Using MSE q u a n t i z e r 4.39 47.14
4.10 MAE q u a n t i z e r
Using 53.22
Using BTC 58.34 4.85
Using
parametric' MSE q u a n t i z e r 64.02 5.42
and assumed uniformdensity
a
tizing 2 and jointly, using 10 bits. f is assigned most bits in
blocks where 0 is small and fewest bits where 0 is large.
Additional savingscouldberealizedbyefficientlycoding
the bitplane.Entropycalculationsindicatethat 0.85 bits/
pixelwouldbesufficientfor the bitplane.Suchsmallim-
provementdoes not justify the increased complexity of the
system and the increased variability to error that efficient cod-
ing entails.
Setting the threshold of the quantizer 2, at removes a possi-
ble degree of freedom for optimizing the encoding. Allowing Figure 4. Results of BTC using third moment preserving threshold
the threshold tobea variable permitstheencodingtopre- selection. Original images are atthetop.Codedimages are atthe
serve not on1.y the first two sample moments but also the thud bottom. Data rate is 2.0 bits/pixel.
sample moment. To analyze this, we make use of the sorted
pixel-values Yi. The third moment can beexpressed as In general the value of q obtained will not be an integer,so
inpracticeitmust be roundedtothenearestinteger. An
interesting interpretation of Equation 9 is that the threshold
is nominallyset to the samplemedianandbiasedhigher or
lower depending upon the value of the third sample moment
The problem then is finding a, b, and q to preserve 2,f 2 , and which is ameasureofskewnessof the Yi's. Thisthreshold
' T3 ( q definesthethresholdsince it specifies thenumberof selectionrequires no extra computation at the receiver(de-
Xi's greater than X , , ) , Equation 3 now becomes coder), but the transmitter (encoder) does have extra process-
ing. This method of threshold selection is far easier to imple-
m X = ( m - 4)" + qb ment than are the nonparametric quantizers discussed in Sec-
tion I11 since they require an exhaustive search.
Figure 4 shows results using this threshold selection. It im-
proves subtle features (such as near edges) of the image that
are usually important in analysis of aerial photography imagery.
This improvement will be discussed in the following section.
Having the solutions
V. PERFORMANCE EVALUATION OF BTC
Recent studies at Purdue University have evaluated various
image coding techniques for transmitting aerial reconnaissance
imagesovernoisychannels [ 161. BTC, transform,Hybrid
[ 17] and one other technique were evaluated by professional
photo analysts.
Although BTC did not perform as well as transform coding
in the photo interpreters evaluation, it proved superior to the
other techniques. In the presence of many channel errors, BTC
was superior to all of the other techniques. Some images used
in the study are shown in Figures 5-8. In these figures BTC is
compared with the Chen and Smith [ 2 ] method of transform
where
coding and Hybrid coding. The computed mean square errors
and mean absolute errors are shown in Table 11. Our study has
shown that the meansquareerrorandmeanabsoluteerror
measurecannot beeasilycorrelatedwith photo analysts'
evaluations.Insome cases, imageswithlargemean square
-
u # 0. errors wereevaluatedhigherthan imageswithsmallermean
square errors. It should be noted that BTC requires a signifi-
Ifa=O,thena=b=X. cantly smaller computational load and much less memory than
IEEE TRANSACTIONS
ON COMMUNICATIONS, VOL. CUM-27,NO. 9 , SEPTEMBER 1979 1339
TABLE I1
COMPUTED MEAN SQUARE ERROR AND MEAN ABSOLUTE
Figure 6 . Results of codingoriginal(Figure 5 ) using BTC with third ERROR MEASURES FOR COMPARISON IMAGES SHOWN IN
moment preserving threshold selection. Data rate is 1.63 bits/pixel. FIGURES 6-1 1 AND FIGURE 13
Mean Square E r r o r Mean Absolute E r r o r '
Figure 6 84.22 5.94
Figure 7 67.15 6.32
Figure 8 125.84 6.12
Figure 9 115.09 6.29
Figure10 115.31 7.06
Figure 11 140.33 6.67
Fisure13 74.67 5.72
1340 IEEE TRANSACTIONS
ON COMMUNICATIONS, VOL. COM-27, NO. 9 , SEPTEMBER 1979
Figure 9. BTC coding withchannelerrorprobability of 10-3. Data Figure 11. Hybridcodingwith channelerrorprobability of 10-3.
rate is 1.63 bits/pixel. Data rate is 1.63 bits/pixel.
ACKNOWLEDGMENT
The authors would like to thank Dr. John Gamble of the
Rome Air Development Center for his helpful criticism of this
work. We would also like to thank Richard P. Petroski and his
colleagues of the Rome Research Corporation whoprovided
the expert photoevaluations mentioned in Section V.
The authors would also like to thank Professor Steven Bass
of Purdue for suggesting the hybrid formulation of BTC and
Ali Tabatabai and Paul Stiling also of Purdue for providing the
transform and Hybrid coded images.
APPENDIX
Toillustrate coding a4 X 4picture block using Block
Truncation Coding, let usquickly review the basic BTC
algorithm:
a) the image is divided ih small nonoverlapping blocks such
Figure 13. Results of HybridFormulation of BTC: Data rateis 1.88 as 4 X 4.
bits/pixel. b) The first andsecond sample moments are computed.
c) A bit plane is constructed such that each pixel location is
coded as a “one” or a “zero” depending on whether that pixel
processing of the image can reduce the effects of both these
artifacts while simultaneously reducing the mean square error
is greater than x.
d) The bitplane, $, and 0 are sent to thereceiver.
and mean absolute error. It should be emphasized that these
coding artifacts are problems in high resolution aefial recon-
x
e) The picture block is reconstructed such that and (al-a
ternatively X 2 ) are preserved. That is, pixels in the bit plane
naissance images where minute objects are important. These that afe “0” are set to “a” and the “1”’s are set to “b” in
coding artifacts usually are not visible intypical “head and Equation 4. For example, suppose a 4 X 4 picture blockis
shoulders” imagery. given by the following:
[ :z
VI. HYBRID FORMULATION OF BTC
121 114 56 47
‘BTC uses only first-order informationanddoesnot ex-
ploit the two-dimensional sturcture of the image within each 37 200 247 255
block as do most other forms of image coding. For example xij= 12 16g]
in two-dimensional transform coding thetransform coeffi-
cients contain information about variations in the picture in 251
7
both directions. Also BTC generally has a poor response near
the spatial frequency of 1/2 cycle per block. so
One methodto overcomethisdisadvantage is ahybrid
formulation. First a highly compressed Cosine transform
x = 98.75
coded image is subtracted from the original image. For the re- -
u = 92.95
sults presented herethetransformpicture was obtained by
taking the two-dimensional Cosine transform over 16 X 16 q=7
pixel blocks. Only the eightnon-dc coefficients in thelow
frequency section of eachblock were retained. Thiscor- and
responds to a zonalfiltering method with abitrateof
0.25 bits/pixelforthe highlycompressed image. BTC was a= 1 6 . 7 s 17
then used on this difference pictureandtherecombination
formed at the receiver. While this did increase the computa- b = 204.2 S 204
tional load, the improvement seemed to be significant enough
the bit planeis
to give this method further attention. Figure 13 presents re-
sults of this hybrid method. TableI1 has the mean square error
and mean absolute error measures for Figure 13.
VII. CONCLUSIONS
A new,simple technique has beendescribedwithmodifi-
cationsto improve itsperformance.The resulting
recon- Lo o o IJ
The reconstructed block becomes Effects of Digital Demodulation on Component Coding o f
r 2 0 417 17
204 NTSC Color Signals
204
204
17
204 ERIC DUBOIS