DTIC ADA265862: Detection of Point Objects in Spatially Correlated Clutter Using Two Dimensional Adaptive Prediction Filtering

9 Pages·0.65 MB·English

by  Defense Technical Information Center

AD-A265 862
MENTATI -NPA GE .
I1tO average I hour per response, inctuling the time for reviewing i(cid:127)stiorrui ose, eOavsrgs ivir s(cid:127)cl'.ces gttr ar
mlono t nformation Sena comments ;egaro.rg ths burden estimate an y ithera spect u(cid:127)tIt hr coleC!&I, mcro Irnatx cir
rs Services. Diretorate for Intormnalion Operations and Reports. 1215 Jefferson Davrs H.ghdVay Sine 1204. Arirnglu VA
2262,4302, and to the Office of Management anid Budget, Paperwork Reduccbon Pro/ect (0704 0188), Washington, DC 20503
1. AGENCY USE ONLY (Leave bankj 2 RAEPpOrRilT D1A9T9E3 3 REPORT TYPE AND DATES CO'V RED'
Professional paper
4. TITLE AND SUBTITLE 5 FUNDING NUMBERS
DETECTION OF POINT OBJECTS IN SPATIALLY CORRELATED CLUTTER PR: ZW51
USING TWO DIMENSIONAL ADAPTIVE PREDICTION FILTERING PE: 0601152N
6. AUTHOR(S) WU: DN300155
T. Soni, J. R. Zeidler, W H. Ku
7. PERFORMING ORGANIZATION NAME(S) AND ADDREbZ,,rS) B PERFORMING ORGANIZATION
REPORT NUMBER
Naval Command, Control and Ocean Surveillance Center (NCCOSC)
RDT&E Division
San Diego, CA 92152-5001
9, SPONSORING/MONITORING AGENCY NAME(S) AND ADORESS(ES) - 10 SPONSORING)MONITORING
Office of Chief of Naval Research "N f, AGENCY REPORT NUMBER
Independent Research Programs (IR)
OCNR- 10P
Arlington, VA 22217-5000 _______, Wk....
11.S UPPLEMENTARY NOTES R
12a. DISTRIBUTION/AVAILABILflY STATEMENT 12b. DISTRIBUTION CODE
Approved for public release; distribution is unlimited.
13. ABSTRACT (Maximum 200 words)
This paper studies the performance of a two dimensional least mean square (TDLMS) adaptive filter as a prewhitening
filter for the detection of small signals in infrared image data. The spatially broad clutter with long correlation length is
seen to be narrowband in the two dimensional frequency domain. This narrowband clutter is predicted and subtracted
from the input, leaving the spatially small signal in the residual output. The output energy in the residual and prediction
channels of such a filter is seen to depend on the correlation length of the various components in the input signal, thus per-
mitting the separation of short correlation targets from the longer correlation clutter. False alarm improvements and detec-
tion gains obtained by using this detection scheme on thermal infrared sensor data with known target points is presented.
Published in Record of the 26th Asilomar Conference on Signals, Systems, & Computers, Vol 1 pp 846-851.
14 SUBJECT TERMS 15 NUMBER OF PAGES
image processing
signal detection i,' PRICE CODE
'17 SECURITYCLASSIFICATION 18 SECURITY CtASSIFICATION 19 SECURITY CLASSIFICATION 20 IIMITATION OF ABSTRACT
OF REPORT OF THIS PAGE OF ABSTRACT
UNCLASSI FI E UI 'NCIASSIF I E) UNCLASSIFI ED SAME AS REPORT
NSN 754001 280 0 STa5r0ar0lo~nCQ (FRONT)
"UiQCLASSIFIED
21 a. NAME OF RESPONSIBLE iNDIMIOUAL 21b TELEPHONE f(includ& Area Cowe, 2 1c OF r,C;E SYMBO0L
J. R. Zeidler [phone] Code 804
Acoe~iOUFor
, By .....-- - - --
Dist ivz(cid:127)ttOL' _ _
Diat
FI~
S tan dard In -2 %(. RAKC)
N S N 7540 0 1-280 5500
UTNCLASSIFIED
Conference Record of
The Twenty-Sixth
Asilomar Conference on
Signals, Systems & Computers
Volume 2 of 2
October2 6-28, 1992 Pacific Grove, California
Edited by
Avtar Singh
San Jose State University
Sponsored by
Naval Postgraduate School
Monterey, California
and
San Jose State University
San Jose, California
In cooperation with
IEEE Signal Processing Society
IEEE Computer Society
IEEE Computer Society Press
Los Alamitos, California
Washington * Brussels • Tokyo
9 3
(cid:127) 93-13491
Detection of Point Objects in Spatially Correlated
Clutter Using Two Dimensional Adaptive Prediction
Filtering*
Tarun Sonit, James R. Zeidlertt, and WaltertI . Kut
t Department of Electrical and Computer Engineering,
University of California at San Diego,
La Jolla, CA 92093-0407.
TCode 804, Naval Ocean Systems Center,
San Diego, CA 92152-5000.
large and correlated clutter in image data. Methods
based on image models and prediction of clutter are
Abstract often applied to such applications [1,21. In the ab-
sence of a well defined model for the image data, the
This paper studies the performance of a two di- relevant parameters of the image clutter have to be
mensional least mean square (TDLMS) adaptive flj- estimated. Further in the case of nonstationary clut-
ter as a prewhitening filter for the detection of small ter these parameters may have to be continuously
signals in infrared image data. The spatially broad updated, before cancellation can be done. One di-
clutter with long correlation length is seen to be nar- mensional adaptive linear prediction filters have been
rowband in the two dimensional frequency domain, applied to the detection of narrow band signals em-
This narrowband clutter is predicted and subtracted bedded in non-stationary noise as well as to the re-
from the input, leaving the spatially small signal in moval of narrow band interference from broad band
the residual output. The output energy in the resid- data [3-5]. In the first case a narrow band signal
ual and prediction channels of such a filter is seen to of interest is extracted from the prediction channel
depend on the correlation length of the various corn- of the adaptive filter. In the second case the broad
ponents in the Input signal, thus permitting the sepa- band signal of interest is extracted from the residual
ration of short correlation targets from the longer cor- error of the adaptive filter 13,41.
relation clutter. False alarm improvements and de- Adaptive filters analogous to the LMS and lat-
tection gains obtained by using this detection scheme tice implementations in one dimension, have been
on thermal infrared sensor data with known target recently extended to two dimensions with applica-
points is presented. tions in image processing t6, 7]. These algorithms
update the filter weights based on the spatial coher-
ence between the signal and noise components of the
1 Introduction data and minimize the variance of the prediction er-
ror (residual) without explicit assumptions about the
This paper addresses the problem of detecting noise statistics. In this paper the performance of two
small objects of very small spatial extent masked by dimensional least mean square(TDLMS) adaptive fil-
"*This work was supported by the NSF I/UC Reach ters in the line enhancer configuration [8,91 for the de-
Center on Ultra High Speed Integrated Circuits and Systems tection of small targets is studied. The components of
(ICAS) at the University of California, San Diego. The Office the input image data are seen to be separated based
of Naval Research and the National Sci.nce Foundation under
grant #ECD89-16669 and the U.S. Naval Command Control on their correlation lengths. Thus long correlation
Ocean Surveillance Center, R.D.T.&E. Division's Independent clutter is predicted and can be cancelled from the in-
Research Program. put with the residual containing the signal of interest.
846
1058-6393/92 $03.00 CD1 992 1EEIE
iG,"
Ad pti've Alorithln - . "
. . ..... Figure 2: A two stage filtering an(d d<ltection
Figure 1: The Two Dimensional LMS Adap- process
tive Filter.
A detector based on TDLMS prewhitening filters is
applied to thermal niultispectral infrared sensor data
with an appreciable decrease in the number of false
alarms.
Section 2 describes the TDLMS filter used briefly.
Section 3 describes the separation of the input image
energy into two channels, the prediction and the er-
ror channels based on the correlation lengths of the
various components of the input image. Section 4 de-
scribes the gain obtained by such a filter and its de-
pendence on the adaptive time constant of the filter.
Section 5'describes the application of such prepro-
cessing on a single band image from a multi-spectral
data set and the improvement in the false alarm rate
obtained.
2 The TDLMS Filter Figure 3: Signal with two components: object
of interest at (100,100) with corre-
The TDLMS adaptive filter f61 shown in Fig. 1 pre- lation spread of a2 = 2 and spurious
dicts an image pixel as a weighted average of a small component at (200,200) with corre-
window of pixels as lation spread of or2 = 98.
N-I N-1
Y(m, n) = E 1 Wj (1, k)X(m - I, n - k) (1) In our implementation a causal window with quarter
t=0 k=O plane support and a left to right lexicographic scan
where, X is the input image of size M x M, Y is were used.
the predicted image and Wj is the weight matrix at
the P iteration. The window size (and hence the
weight matrix) is N x N. If the image is scanned 3 Separation by Differences in Corre-
lexicographically, j = mM + n. The predicted pixel lation Lengths
value is ;ompared with a reference image ,D(m, n),
which is a shifted version of the primary image in the In this section we present the results obtained by
line enhancer configuration. The error is then found applying the two stage augmented detector shown in
Sn = )Fig.2 to infrared image data. The adaptive filter used
"E (m, n) = =D(rn, n) - Y(m, n). (2) for prewhitening was the two dimensional adaptive
In the absence of any knowledge of the statistics of LMS filter described in section 2. The image
the input, the LMS algorithm uses instantaneous es- used was channel 4 of a 6 channel data set collected
timates. Then, the stee'pest descent algorithm leads by the NASA Thermal Infrared Multispectral Scan-
to the weight update equation ner(TIMS) sensor and includes a rural background
over the hills of Adelaide, Atistralia [10. This sec-
W j+i(I, k) = W -+i pjX(m - 1, n - k) (3) tion includes performance data from injected point
i47
~
------------
S*.\-. ...
Figure 5: Energy at the pixel of interest ini the
output of the TDLMS as a function
of the correlation length.
Figure 4: The residual channel output for i
-I
Fig.15. TDLMS was used with a3x3 I F i .-rl
*-.~
window and 1A= le - 7. LSBR at ~
(100,100) = 16.024dB and LSB9. at
(200,200) = -16.75dB. IL 1. 1 t I
objects in TIMS sensor output data. The injected
.
objects were utilized to illustrate the performance of3
the TDLMS filter for known object parameters.1 ,fI
To study the ability of the TDLMS prewhitener IIL
to separate signals based on spatial correlation, sim-
ulation studies were conducted with varying signal
spread. The infrared signal model developed by Chan Figure 6: Energy at the pixel of interest in the
et. al. [Il] was used for these simulations. This model prediction channel of the TDLMS as
leads to a Gaussian intensity function for the object, a function of the correlation length.
of the form
01' =2 is seen to be present, while the output for
(Y -
I(Z' t') = re (4) the a12 = 98 case is absent.
Figure 5 shows the energy at the pixel of interest
where 1r is the maximum value of the object intensity in the residual channel output of the TDLMS filter
function, (zo, y/o) is the position of the center of the as a function of the signal correlation length (o, ).
object, and (a,,, a,,) define the spatial spread of the The pixel intensity at the output was observed for an
object. object defined by Eq.4. It is seen that as the correla-
Two objects with symmetric Gaussian intensity tion length increases, the output energy at the pixel
functions (a., = a.,) defined by Eq.4 were inserted in of interest decreases. Further, as shown in Fig.6, the-
the infrared background. Fig.3 shows an image with energy in the prediction channel of the TDLMS filter
the two components inserted in the background. One correspondingly increases as the correlation length in-
is the object of interest (at pixel location 100,100) creases. It should be noted that the prediction chan-
which is a Gaussian shape with a12 = 2 and the other nel output always contains the energy due to the non
is a component of the same shape but with a'2 = 98. zero mean of the image which is seen in Fig.6 as a
Fig.4 shows the output of the TDLMS whitener for base level of 1.3 x 104.
this case, where the component of the input with In both these plots, depending on p, three regions
848
.ial
W.3
10 IJ t1.1 41. f-I LL I- I I-111
2~ A I I*2 4 I ll
C...W- L.A* a' Ad.9P -
Figure 7: Energy at the pixel of interest in the Figure 8: Gain vs. p for the Adelaide back-
output of the Local Demeaning flu- ground with one pixel signal at
ter as a function of the correlation (100,100).
length.
sensitive to changes in the input, and some of the
can be defined for this TDLMS adaptive filter: energy from the signal of interest is also predicted and
cancelled. The misadjustment noise in the adaptive
SFor objects with very small correlation length, filter weights increases with p and is another factor
the energy is contained almost totally in the out- contributing to the drop in gain as p increases.
put channel.
* For objects of long correlation length, the energy
is absent from the output. 5 Reduction in False Alarm
* For objects of intermediate correlation length, In [101 Hoff et. at. identified 14 pixels which
there is a partial deccrrelation between the input contained small objects in the TIMS sensor data de-
and reference channels. In such cases the energy scribed above. The 14 pixels were identified on the
of the object is split between both channels. basis of multispectral data in the 6 separate bands.
A s a The locations designated in [101 are illustrated in by
shwA similar plot for the local demeaning filter is boxes in Fig.9. The pixel intensities at the output
Si shown in Fig.7, and it is seen that the local demean- of the TDLMS filter (residual channel) is shown in
ing filter is not able to provide a clear boundary be- Fig. 1r.
tween the two regions of differing correlation spread. Fig.1l shows a high resolution intensity plot
Thus we see that the TDLMS is able to separate the around the pixel (199,92). As seen in the figure, this
orebajetcot ofs ipnetedre st from the clutter based on their cor- pixel hhpaaixss eaal vveerryy high Local Signailn to tBBo acckggrroounn d RRaa--
Sre lation spread, tio(LSBR) and this causes the object to be predicted.
Fig.12 shows the corresponding output for this pixel.
It is seen that there are a number of pixels in the
4 Dependence on the Adaptive Time local region with comparable intensity.
Constant Fig.13 is a similar high resolution plot around the
pixel (188,25). In this case, due to the low LSBR, the
Fig.8 shows the behavior of the TDLMS filter as output (shown in Fig.14) is seen to contain energy
a function of p for a one pixel signal at two different due to the object pixel and there is a vast improve-
signal intensities. There is an optimum value of p at ment in the detectability.
which the maximum gain is obtained. Ifp is less than Though there is a loss in LSBR at some pixels it
this optimum value, the filter is not able to converge is seen that the detection performance of the filter
to the statistics of the clutter. Hence the residual improves considerably when the input image is pro-
channel contains some component of the correlated cessed by the TDLMS adaptive filter. For a threshold
clutter leading to a reduced gain. If the value ofp is set to detect all 14 of the signals defined in 110], the
higher than the optimum, the adaptive filter is very number of false alarms in the image is significantly
849
_ .w!
Figure 9: The infrared imiage, data with the 14
pixelIs of interest. Note that some.~ Fignire 12. Hiighi r1esolutioii initenii ty plot oif
objc ts are clustered close to each a 33x33 window around1 tho pixel
other their eniclosinig boxes overlap. (199.92) after proc-essijig.
Figure 13: Highi resolution intensity plot of
a 33x33 window around the p)ixel
(188,235) before processing.
F-igre 10- The infrared image data otitpu t Figu.i'in 14H igh rosoltit ion ilitelisi ty plot of
from TDLMS ndaaptive fiter for a 33-,33 window around tjhoe IXel
Fig .20. (188.235) after pirocess1ing.
though the background clutter is not stationary, the
_- 9 - -. --..-- T- /n TanDdL cManSc efli ltetlre aisd ainblge ttoo apnr eadpipcrte cpiaabrtlse oifm tihpero vcelruntetnetr
a-st M the detection performance.
References
S. - .- [1] A. K. Jain, Fundamentals of Digital Image Process-
""j ing. Englewood Cliffs, NJ.: Prentice Hall, 1989,
t,(cid:127)y.!.o'
.. [21 J. Y. Chen and I. S. Reed, "A Detection Algorithm
for Optical Targets in Clutter," IEEE Transactions
0.1 - .-- on Aerospace and Electronic Systems, vol 23. no. 1,
01 ____0_ _____-___.r. 1 .- __________________ ""Ip:p. lot~alfa 46-59, 1987.
lO 1i0, D i[31 J. R. Zeidler, "Performance analysis of LMS Adap-
tive Prediction filters," Proceedings of the IEEE,
vol. 78, pp. 1781-1806, Dec. 1990.
! Figur1e5 : False alarnw in the Adelaide image [4] L. Milstein, "Interference Rejection Techniques in
for known objects as the percentage Spread Spectrum Communications," Proceedings of
of objects detected is lowered, the IEEE, vol. 76, pp. 657-671, June 1988.
TDLMS was used with p = le - 7. [5] N. Ahmed, R. J. Fogler, D. L. Soldan, G. R. Elliott,
and N. A. Bourgeois, "On An Intrusion Detection
reduced. If the detection criterion is changed to re- Approach Via Adaptive Prediction," IEEE Transac-
reduced.thI enumbe detectedion i tens theanubteofr rtions on Aerospace and Electronic Systems, vol. 15,
duce the number of detected objects the number of pp. 430-437, May 1979.
false alarms in the image are further reduced. A plot [6] M. M. Hadhoud and D. W. Thomas, "The Two-
of the percentage of objects detected as a function Dimensional Adaptive LMS(TDLMS) Algorithm,"
l -thoef number of false alarms per pixel is shown in IEEE Transactionso n Circuits and Systems, vol. 35.
Fig.15. A similar plot for the local demeaning filter no. 5, pp. 485-494, 1988.
is also shown in this figure and it is seen that for a [7] M. Ohki and S. Hashiguchi, "Two-Dimensional LMS
fixed number of false alarms per pixel the percentage Adaptive Filters," IEEE Transactions on Consumer
of of objects detected is much higher when the data Electronics, vol. 37, pp. 66-73, Feb. ;991.
is processed with the TDLMS based adaptive filter [8] T.Soni, B. D. Rao, J. R. Zeidler, and W. H. Ku, 'Sig-
than with the local demeaning filter. nal Enhancement in Noise and Clutter Corrupted
Images Using Adaptive Predictive Filtering Tech-
niques," in Proceedings of SPIE, vol. Conference
6 Conclusions 1565, Adaptive Signal Processing, (San Diego, Cali-
fornia, USA), pp. 338-344, July 1991.
The line enhancer using a TDLMS adaptive pre- [9] T.Soni, J. R. Zeidler, and W. H. Ku, "Recursive Es-
np etg timation Techniques for Detection of Small Objects
"diction filter wns applied to the problem of detecting in Infrared Image Data," in Proceedings IEEE Inter-
a small object in infrared backgrounds. It is seen that national Gonf. on Acoustics Speech .,,d Ji(cid:127)'c! P':-
for correlated background clutter, the adaptive fil- cessing, vol. 3, (San Francisco, USA), pp. 581-584,
ter predicts the background. The filter residual then March 1992.
contains the signal of interest, without the cluttering [10] L. E. Hoff, J. R. Evans, and L. E. Bunney, "Detec-
background. tion of Targets in Terrain Clutter by Using Multi-
The prediction characteristics of the TDLMS filter spectral Infrared Image Processing," Technical Re-
were seen to depend on p and permit the separation port for Naval Ocean Systems Center, Dec. 1990.
of components in the input of differing correlation [11] D. S. K. Chan, D. A. Langan, and D. A. Stayer,
lengths. This allows the detection system to distin- "Spatial Processing Techniques for the Detection of
guish between a signal of interest and a background Small Targets in IR Clutter," in Proc. of SPIE,
cluttering phenomenon. It is seen that this cancella- Technical Symposium on Optical Engineering and
tion of background clutter leads to a reduction in the Photonics in Aerospace Sensing, (Orlando, Florida),
number of false alarms obtained for any given thresh- Apr.,1990.
old of the detector. For real data, it is seen that,
851