Pattern Recognition

Le~ers
ELSEVIER Pattern Recognition Letters 17 (1996) 37-46

An approach to detect lofar lines

J-C. D i M a r t i n o 1 S. T a b b o n e *
CRIN/CNRS-INRIA Lorraine, Campus Scientifique, BP 239, 54506 Vandoeuvre-les-Nancy,France
Received 1 October 1994; revised 26 April 1995

Abstract
The problem of extracting spectral lines from sonar images is treated using a complete scheme that combines an edge
detector with a line tracing process. Comparative performance results confirm the efficiency and the reliability of the
proposed approach.
Keywords: Line detection; Edge detector; Tracing; Filtering

1. Introduction • Multi-lines can be displayed on the lofargram de-

1.1. The passive sonar signal recognition problem
The fundamental objective of a passive sonar system
is to detect the presence of target-like signals in under-
water acoustic fields. A typical passive sonar system
is depicted in Fig. 1. Signals emitted from the acoustic
source are recorded by an array ofhydrophones, beam-
formed and spectrum analysed. Narrow-band compo-
nents of the signals' result provide an image of fre-
quency power versus time, more commonly referred
to as a "lofargram" (Fig. 2).
Lofar is an acronym for LOw Frequency Analysis
and Recording. Its main characteristics are:
• Signal to noise ratio is low due to the discreteness
of the sources.
• A constant frequency tonal produces a darkening
over time periods, which appears as a vertical line on
the lofargram display.
* Corresponding author. Email: tabbone@loria.fr
1This work was supported by the research centre of underwater
signal processing, DCN Toulon Cedex. Email: dimartin@loria.fr
pending on the spectrum band-width.
• Slanted lines or slightly curved lines can be
shown on the lofargram. They are due to the Doppler
effect.
By characterizing the spectral lines on a lofargram,
one can determine the acoustic source of the sound.
Actually, trained operators currently perform this char-
acterization, partially aided by machine measurements
of the signals.
1.2. Related works
In this section we review several methods proposed
in the literature to extract lines from images. Our re-
view is not intended to be exhaustive, but some of
the related works represent important advances in this
area, and we have experimentally tested some of them.
Abel et al. (1992) perform a detection step using
a statistical likelihood test, Then mathematical mor-
phology operators are used to extract regions which
encompass lines. This method, well tested on sonar
images, requires however a priori information regard-
ing the noise environment and the number and type of

0167-8655/96/$09.50 (~) 1996 Elsevier Science B.V. All rights reserved

SSDI 01 67-8655 (95) 00088-7
38 J-C. Di Martino, S. ~bbone/Pattern RecognitionLetters17(1996) 37-46
-* f;7 P
.f;Tq TIME
DELAY mmmm~
BEAM
FORMER
Fig. I.Passivesonarsystem block diagram.
(a)
Fig. 2. Lofargram image containing four spectral lines. (a) 2D grey-level image. (b) 3D surface.
sinusoids present.
Other contributions in the field of curved line ex-
traction are proposed by introducing perceptual con-
siderations. All these methods rely on a two-step pro-
cess. Initially, a line detector is applied to the im-
age in order to determine for each pixel the magni-
tude and direction of the image gradient. Then, several
grouping strategies are used to enhance points lying
on a curve while reducing those belonging to noise.
Sha'ashua and Ullman (1988) define a saliency mea-
sure based on perceptual constraints. A dynamic pro-
gramming algorithm is used to guide the grouping pro-
cess for extracting long lines of low total curvature.
Other researchers (Parent and Zucker, 1989; Yang,
1990; H6rault, 1991) have studied the tangent fields
consistency using relaxation constraints or simulated
annealing algorithms. Guy and M6dioni (1992) define
a vote system to execute a directional convolution of
the image using saliency-enhancing operators. Di Mar-
tino and Alexandre (1994) also use a vote system
including an interpretation step that allows the pro-
cessing of fluctuating and noisy curves such as those
encountered in sonar images. Globally, these methods
SPECTRUM mmmm~
LOFARGRAMS
ANALYSIS
(b)
give good results for constant curvature curves but they
are not well adapted to fluctuating curves (curvature
sign often changes) such as spectral lines.
Neural network approaches are also an important
field of methods that have been proposed to deal with
the line detection problem. In this case the line de-
tection is performed using a supervised learning neu-
ral network (Khotanzad et al., 1989; de Bollivier et
al., 1988; Leeming, 1993). These methods need a su-
pervised learning set that reduces their utility in real
cases. Di Martino et al. (1994) propose an alternative
approach using a non-supervised neural network. This
one only provides a contrast enhancement process and
so needs a detection step to extract the lines.
Other usual approaches to solve the curved line ex-
traction issue are based on tracing processes. Some (di
Martino et al., 1993; Fernandez et al., 1990) based on
dynamic programming (Bellman, 1957) consist of ex-
tracting optimal paths according to the minimization
of a cost function (Montanari, 1971 ). These methods
give interesting results on IR (infrared) and sonar im-
ages in the sense that they provide a connected chain
of points accurately localized. However, the initializa-
 J-C. Di Martino, S. Tabbone/ Pattern Recognition Letters 17 (1996) 37-46
tion is not easy, and in some cases overlapping search
trees can give wrong paths.
Line detectors, based on a differentiation operation
(Ziou, 1991 ) or on a local energy computation (Free-
man and Adelson, 1991 ), can also be provided to ex-
tract spectral lines. Since the images are very noisy,
these approaches require a smoothing stage. So there
is a tradeoff between the localization and the detec-
tion of edges. On one hand, with low smoothing the
detector is more noise-sensitive, but fine structures are
preserved and edges are well localised. On the other
hand, high smoothing removes noise and fine struc-
tures too. Moreover, in this case the detection is good
but the localization error is large. These approaches
are not well adapted to lofar images because the im-
ages are very noisy and we need to restore very fine
structures.
The work described in this paper is toward auto-
matically processing lofargrams in order to suppress
the noise and to detect the spectral lines. This pro-
cess is very important because the quality of further
automatic processing such as classification or tracing
acoustic sources relies on the quality of the extracted
spectral lines. Thus, we propose to combine an edge
detector with a line tracing process. The edge detec-
tion process is used to initialize the tracing process.
More precisely, the image is highly filtered in order to
remove noise, so lines with good detection and poor
localization are extracted. Regions encompassing the
lines (the zero-crossings) are detected to initialize and
to limit the extent of the line tracing process.
The proposed method is intended to be used in a
real-time situation, without using any information such
as signal duration, number of lines or noise level. Re-
suits on real lofargram images confirm, in practice, the
efficiency and the reliability of the proposed approach.
2. Our approach
The different steps of our approach are described
in Fig. 3. Spectral lines are extracted by the detection
process in order to find surrounding regions. Then,
a tracing step is executed in each region to find the
lines accurately. We describe these three stages more
precisely in the following sections.
39
2.1. Detection process
The detection process consists of detecting and
bridging spectral line contours in order to extract
regions. The line detection, further described, is
achieved in a two-step process: smoothing and differ-
entiation.
2.1.1. Line detection
Smoothing. Noise in lofargram images is essentially
characterized by high frequency components which
can be reduced by a smoothing step. It has been
pointed out that the Gaussian ensures an optimal com-
promise between the detection and localization errors
(Torre and Poggio, 1986). The Gaussian is given by:
1 ( lx2+y 2)
g,~(x,y) = 2 - ~ 2 e x p 2 0-2 (1)
where o- is a standard deviation expressed in terms
of pixel dimensions. The application of this filter to a
real lofargram image (Fig. 2) is shown in Fig. 4.
Edge detection. The spectral lines in a lofargram can
be modelled by peak profiles (see Fig. 2). To perform
the detection, we run a detector based on steerable fil-
ters designed in quadrature pairs as described by Free-
man and Adelson ( 1991 ). The method consists of cal-
culating the response of an arbitrary linear filter by a
linear combination of known filter responses, which
will in general be significantly faster than computing
the convolution of a rotated filter with the image (see
(Freeman and Adelson, 1991 ) for more details). Ta-
ble 1 shows the steering equations of the pair of filters
used by the detector where G is the second derivative
of the Gaussian and H its Hilbert transform. In prac-
tice, all these filters are separable, which improves the
computation time.
The detection first computes in each image point a
local dominant direction 0a (the angle of maximum
response of E ( O ) ) . This is performed by defining a
local energy function E(O) = [G°] 2 + [H°] 2 whose
maxima correspond to visual features (Morrone and
Owens, 1987).
Once this is done, we can easily compute the second
derivative of the image in order to enhance the lines
and also to locate regions encompassing these lines.
This is simply achieved by taking the appropriate com-
40 J-C. Di Martino, S. Tabbone/ Pattern Recognition Letters 17 (1996) 37-46
IMAGE
Fig. 3. Approach scheme.
(a)
Fig. 4. Lofargram image after smoothing. (a) 2D grey-level image. (b) 3D surface,
Table 1
Steering equations (Freeman and Adelson, 1991)
Go = gl ( 0 ) ~ 0 q- g2 ( O ) ~ 1r/4 + g3 ( 0 ) ~ r/2
gl(O) = cos2(8)
g2(0) = --2cos(0)sin(0)
g3(0) = sin2(0)
Go = 0.9213(2x 2 - l ) e x p - ( x 2 + y 2)
G~r/4 = 1.843( x y ) e x p - - ( x 2 + ) ,2)
G ~r/2 - 0.9213(2y 2 - 1) e x p - - ( x 2 + 3,2 )
binations of the G basis filter outputs, and adaptively
steering them along the local direction of dominant
orientation (see Fig. 5 ( a ) ) . After finding, in the en-
hanced image, the 2D local maxima in the direction
perpendicular to Od, the false contours are suppressed
using an hysteresis threshold (Canny, 1986). No ad-
ditional filtering is required for this step and the whole
process (i.e., finding the local dominant direction and
the derivation of the image) involves only one pass of
the image.
2.1.2. Gap bridging
Because of its local nature, the detection process is
not able to provide continuity and coherence of im-
age curves. In our case, the major drawbacks of this
process result in:
• gaps, where weak but significant information is
missed due to thresholding errors and intensity discon-
tinuity stemming from various and complex physical
phenomena;
TRACING IMAGE
(b)
H° = hi (0)7"( 0 + h2(0)7-[ 7r/6 + h3 (0)']-/"/r/3 q- h4(0)7"~ rr/2
hi(O) = cos3(0)
h2(O) = --3cos2(O)sin(O)
h3(0) = 3cos(0)sin2(0)
h4(0) = -sin3(0)
7-(0 0.9780(--2.254x + x 3) e x p - - ( x 2 + y2)
7grr/6 = 0.9780( --0.7515 + x 2 ) ( y ) exp - (x 2 + y2)
7-U/3 - 0.9780(--0.7515+ y 2 ) ( x ) e x p - - ( x 2 + y2)
7 ~ / 2 _ 0.9780( --2.254y + y3 ) exp -- (x 2 + y2 )
• deviations in position and uncertainties o f ori-
entation where local estimates of these quantities are
adversely affected by noise.
Our aim is to restore continuous curves as long
as possible. A linking process is first achieved on
the thresholded image in order to obtain elementary
curves, called tokens. Each token, defined by a linked
edge points lying on the same curve, is characterized
by a local dominant orientation at each end of the to-
kens and its main direction. The grouping step relies
on two fundamentals concepts:
( 1 ) A strategy of association based on an iterative
techniques of grouping by successive refinement (Mo-
han and Nevatia, 1989; Tabbone, 1994a). The process
is iterated on a set of tokens and ends when no asso-
ciation is achieved at one stage.
(2) A set of criteria that ensures a good robustness
of the grouping process using adaptative threshold val-
ues. Let ti and tj be two tokens, longi and longj their
length:
 J-C. Di Martino, S. Tabbone/ Pattern Recognition Letters 17 (1996) 37-46
(a)
Fig. 5. Contours found by the detection step. (a) Line contours from Fig. 2. (b) Line contours after gap bridging.
• Proximity: tokens must satisfy the following rule:
d ( ti, t j ) < min( dmax ( n ) , longi + longj)
where dmax(n) is growing at each level n allowing the
bridging of greater gaps.
• Curvature consistency: tokens must satisfy a co-
circularity measure. Let D be the straight line that
reaches two tokens, and Yi and yj the deviation of the
orientation of D from each of the tokens. Then ti and
tj must satisfy:
[~/i--'Zjl "( ~/max -- (~max -- ~/min)~/ "-if----
Y t/max
so, when d ( c i , c j ) = 0, it means that the end points
of the tokens are contiguous, so we let the curvature
be large and bounded by "Ymax.Inversely, when the
distance between the two tokens increases, then the
maximum curvature allowed is weak and equal to ~min-
• Overlapping: sometimes, tokens overlaps due to
quantification or linking errors. A maximum overlap
value Pij is fixed to take these cases into account.
2.1.3. Discussion
We have shown a global scheme to efficiently detect
spectral lines embedded in a very noisy background.
Nevertheless, even if they are well detected, fine struc-
tures are blurred by the smoothing step and so are not
well localised (see Figs. 7(a) and 7 ( b ) ) . We pro-
pose to tackle this problem by first locating regions
that encompass the contours and then using a tracing
algorithm to accurately find the lines in each region.
2.2. Region locating
This process aims to provide the tracing process
with a limited search space where we have a high
41
(b)
chance of detecting lines. Such an approach enables
us to avoid initialization and solution tree overlap-
(2) ping problems inherent in multistage decision pro-
cesses. The main idea is to determine for each con-
tour the boundaries of the smallest region it encom-
passes (Tabbone, 1994b). We first compute the zero-
crossings of the second derivative of the image on each
side of the contours (zero-crossings are detected in
four direction scans). This information cannot be di-
rectly used as boundaries of regions because of effect
of noise that can produce wrong zero crossings. Our
solution consists of estimating each region width by
(3) calculating the mean distance from the curve it con-
tains and its zero crossings. A mathematical morphol-
ogy operator is thus ran to dilate each curve until we
obtain the desired width (see Fig. 7(c) ).
This method does not insure that a region contains
only one curve. It is always possible that some closed
spectral lines will be merged by the smoothing step
into one contour. In our experiments, we have chosen
the smoothing parameter o- equal to 3 in order to have
good noise reduction and to be able to distinguish two
curved lines 5 pixels from each other.
2.3. Tracing guided by contours
A multistage decision process (Bellman, 1957;
Montanari, 1971 ), based on the optimisation of an
objective function qh is used to extract from each re-
gion R the spectral line going through it. The problem
consists of finding an optimal path from a set of graph
nodes defined by the pixeis belonging to a region.
This operation is achieved on the original image. The
fundamental property underlying this method is that
any optimal path between two nodes of a graph has
optimal subpaths for any node lying on it. Thus the
 42 J-C. Di Martino, S. Tabbone/Pattern Recognition Letters 17 (1996) 37-46
] f

.......i i 3 t
-~(!.4 ; -
t

L_i'
! r-~t'-

i....4.,4-,, ..
I i i/~--5
{a}
r.............. .................. !
(b)
L
./' i Pij

Fig. 6. Gap bridging. (a) Successive refinements. (b) Bridging cnteria.

(a) (b) (c)

Fig. 7. Image for which the detector does not give a good result. (a) Fluctuating lines in noisy image. (b) Results of the detection, lines
are blurred and poorly located. (c) Regions encompassing the detected lines.

optimal path between two nodes Pa, P8 can be split and

into two optimal subpaths PAPi and PiP8 for any Pi
lying on the optimal path PAPs. More precisely: 4,(c. pj,,*, U {Pi}) = p,, max
,ev(p,~ qs(C~/- , U {Pi})
• Let R be a region defined by its set of left and
right boundaries {(li, ri)} where i is a row index. with
• Let Pi be a point in the current row i of the region
R. m , U {Pi}).
j0 = argmax @(C*
• Let V(Pi) = {Pil'_-l ' , .... ~'~i ~} be the set of the
k neighbours of Pi on the row i - 1. The function @ is chosen so that the values for a
The optimal path C*Pi of length i ending to Pi is built path increase as the path's total amplitude increases
(see Fig. 8 ( a ) ) from a previously computed optimal and as its global curvature decreases:
path C*ej,,
" of length i - 1 ending at a Pi's neighbour N N--I

point P/" j :
qO(e) = Z a(Pi) - oe ~ ]s(P,_,,P,) -s(P,,Pi+,)]
i=1 i=2

c~, = % u {p,} (4)

i I

where where a(Pi) is the amplitude of Pi, s(Pi, Pj) is the

slope of the segment [ Pi, Pj ]. The gain coefficient a
Pi~J C V ( Pi) and 1 ~<j0~<k allows more or less local distortions to be obtained.
 J-C. Di Martino, S. Tabbone / Pattern Recognition Letters 17 (1996) 37-46 43

,)
pi©
/
/I \ \
~
i I \

'__"
(,

) ¢
(a) (b) (c)
Fig. 8. Tracing guided by contours. (a) Principle of the multistage decision process. ( b ) - ( c ) Extracted spectral lines (compare these
results with those obtained only with the detection step on Figs. 5(b) and 7 ( b ) ) .

In our experiments, its value was fixed to 3 in order
to allow tracing very fluctuating lines (see Figs. 8 (b)
and 8(c) ).
3. Experimental results
Our method has been tested on a set of real lofar-
grams with different signal to noise ratios. The exper-
imental results show that this method is well suited to
detect spectral lines accurately, even with a low signal
to noise ratio. Fig. 9(a) shows a real iofar containing
noisy fluctuating spectral lines. A histogram equalisa-
tion was performed on the printed picture in order to
improve the contrast. The result of the tracing process
is presented in Fig. 9(b). Fig. 10 shows the principal
steps of our method on a noisy image containing five
spectral lines.
We have compared the location performance of the
detection process (edge detector) and the detection
combined with the tracing process (tracing guided by
contours). Line location accuracy is assessed by the
figure of merit proposed by Pratt ( 1991 ):
/A
1 ~ 1 each point. However, we have experimentally set the
F (5)
max(It,la) ~.= 1 + ad2(i)
where I~ is the ideal number of edge points, I A is
the actual number of edge points, d(i) is the shortest
distance of the ith actual edge point to an ideal edge
point, and ce is a positive constant. F is maximum
(that is, F = 1 ) when It = la and d(i) = 0 for all i.
Table 2 shows results obtained from a set of im-
ages (Fig. 11 ) containing spectral lines with increas-
ing spatial frequency (increasing fluctuation of the
curves). We can see that, globally, our approach per-
forms well, especially when the spatial frequency in-
creases.
We have experimentally tested the performances of
different approaches. The tested detectors are based
on four methods: (A) steerable quadrature filters
(Freeman and Adelson, 1991) briefly described in
this paper, (B) adaptative smoothing (Saint-Marc
et al., 1991), (C) a set of oriented masks (set of
template pseudo-laplacian impulse response arrays)
(Nevatia and Babu, 1980) and (D) a tracing pro-
cess guided by the contours (see Section 2). These
detectors were chosen because (1) the contours are
extracted in the same manner with a non-maxima
suppression scheme (Canny, 1986) but differ in the
way they compute the orientation of contours; and
(2) the smoothing is different. The methods A, C and
D use a gaussian filter where o- = 3 (see Eq. (1)).
The adaptative smoothing (method B) is achieved by
iteratively convolving the image with a small averag-
ing mask weighted by the modulus of the gradient at
number of iterations so that the quantity of smoothing
is the same as in the methods A, C and D.
Tables 3 and 4 show a comparative array of per-
formance for images with different signal to noise ra-
tios. In the first case, images contain curves with low
spatial frequency, while in the second case curves are
very fluctuating. Table 5 indicates, for each method,
44 J-C. Di Martino, S. Tabbone / Pattern Recognition Letters 17 (1996) 37-46

(a) (b)

Fig. 9. Lofargram processing. (a) Lofargram. (b) Extracted spectral lines.

(a) (b)

l
i;

(c) (d)
Fig. I 0. Main steps of our line extraction method. (a) Real lofargram. (b) Result of line detection. (c) Regions encompassing the detected
lines. (d) Extracted spectral lines.

Table 2
Line-location figure of merit as a function of spatial frequency
Spatial frequency 0.10 0.20 0.30 0.40 0.50 0.60
Edge detector based on steerable quadrature filters 0.90 0.90 0.89 0.85 0.82 0.81
Tracing process guided by contours 0.94 0.95 0.95 0.94 0.92 0.88
 J-C. Di Martino, S. Tabbone / Pattern Recognition Letters 17 (1996) 37-46 45

(a) (b) (c) (d) (e) (f)

Fig. 1 I. ( a ) - ( f ) Images ( 128 × 128-pixel array) containing spectral lines with the following spatial frequency values: 0.10, 0.20, 0.30,
0.40, 0.50, 0.60.

Table 3
4. Conclusion and perspectives
Line-location figure of merit as a function of signal to noise ratio
(spatial frequency - 0.05)
SNR 1.50 1.67 2.48 2.95 3.41 4.05 5.96 7.45 We have proposed a complete scheme of processes
Method A 0.91 0.92 0.93 0.92 0.93 0.93 0.93 0.93 to perform the spectral line extraction from lofargram
Method B 0.80 0.87 0.88 0.83 0.88 0.87 0.91 0.92 images (sonar images). Our method is built on the
Method C 0.91 0.85 0.94 0.95 0.95 0.95 0.96 0.96
cooperation of two techniques which are for our pur-
Method D 0.95 0.95 0.98 0.95 0.94 0.96 0.95 0.99
poses complementary: the detection of contours and
the tracing process. Comparative results are presented
Table 4 that show the performance of the tracing guided by
Line-location figure of merit as a function of signal to noise ratio contours versus the contour detection alone, especially
( spatial frequency = 0,5 )
SNR 1.50 1.67 2.48 2.95 3.41 4.05 5.96 7.45
when fine structures are to be detected (high spatial
Method A 0.87 0.87 0.87 0.87 0.87 0.86 0.91 0.88 frequency). A set of experiments performed on im-
Method B 0.71 0.71 0.75 0.74 0.73 0.74 0.72 0.74 ages with different signal to noise ratios shows the ro-
Method C 0.87 0.85 0.92 0.89 0.94 0.92 0.93 0.94 bustness of our approach against noise. This method
Method D 0.90 0.87 0.89 0.92 0.91 0.96 0.97 0.98 is currently applied to real lofargram images in a real-
time situation.
Table 5 At the present, the amount of smoothing, deter-
Comparative performance measured in cpu time mined experimentally, is chosen to be high. The re-
Methods A B C D sults is that very near spectral lines can merge after
Cpu time (sec) 27.85 24.13 51.60 36.74
the smoothing stage and thus cannot be accurately lo-
calized by our tracing process, which supposes that a
the execution time resulting of the process, on a sparc unique line belongs to each region. In other respects,
20 sun station, of an image of size 300 × 700. our global schema of line extraction is not entirely au-
The results show the robustness of our approach tomatic, due to the manual choice of the thresholding
against noise, especially when the images are noisy parameters. Further work will deal with these issues.
and the spectral lines have a high spatial frequency.
The gaussian smoothing performs better than the
adaptative smoothing because the latter relies on the
value of the gradient modulus. The oriented mask ap- References
proach (method C) performs as well as our approach
Abel, J.S., H.J. Lee, and A.E Lowell (1992). An image processing
(method D)" applying different oriented masks gives approach to frequency tracking. In: Proc. IEEE Internat. ConJ~
a good estimation of the orientation of the contour. Acoust. Speech Signal Process., San Francisco, CA, Vol. 2,
However, this method is very time-consuming. 561-564.
46 J-C. Di Martino, S. Tabbone/ Pattern Recognition Letters 17 (1996) 37-46
Bellman, R. (1957). Dynamic Programming. Princeton Univ.
Press, Princeton, NJ.
Canny, J. (1986). A computational approach to edge detection.
IEEE Trans. Pattern Anal. Mach. lntell. 8 (6), 679-698.
de Bollivier, M., A. Lemer and J. Tanguy (1988). Reconnaissance
de bruits acoustiques sous-marins par r6seaux multi-couches.
In: Actes du congr~s Neuro-Nfmes, 423-430.
di Martino, J.C. and E Alexandre (1994). Curve line extraction
using statistical perceptual grouping. In: Proc. EUSIPCO,
Edinburgh, Scotland, Sept. 1994.
di Martino, J.C., J.P. Haton and A. Laportc (1993). Lofargram
line tracking by multistage decision process. In: Proc. IEEE
Internat. Conj'. Acoust. Speech Signal Process., Minneapolis,
MN, Vol. I, 317-320.
di Martino, J.C., B. Colnet and M. di Martino (1994). The use of
non supervised neural networks to detect lines in Iofargram. In:
Proc. Internat. Con[: Acoust. Speech Signal Process., Adelaide,
South Australia, Apr. 1994. IEEE Press, New York, Vol. If,
293-296.
Fernandez, M.E, A. Aridgides and D. Bray (1990). Detecting
and tracking low-observable targets using IR. In: Proc. SPIE-
lnternat. Society for Optical Engineering, Vol. 1305.
Freeman, W.T. and E.H. Adelson (19910. The design and use of
steerable filters. IEEE Trans. Pattern Anal. Mach. lntell. 13 (9).
Guy, G. and G. Mddioni (1992). Perceptual grouping using global
saliency-enhancing operators. In: Proc. lnternat. Conf. Pattern
Recognition, The Hague, Netherlands, 99-103.
Hdrault, L. (1991). Rdseaux de neurones rdcursifs pour
I'optimisation combinatoire; application b. la thdorie des graphes
et h la vision par ordinateur. Th6se de doctorat, Institut National
Polytechnique de Grenoble, Febr. 1991.
Khotanzad, A., J.H. Lu and M.D. Srinath (1989). Target detection
using a neural network based passive sonar system. Proc.
Internat. Joint Con]: Neural Networks I, 335-340.
Leeming, N. (1993). The use of artificial neural nets to detect
lines in lofargrams. In: Proc. Underwater Defense Technology,
48-53.
Mohan, R. and R. Nevatia (1989). Segmentation and description
based on perceptual organisation. In: Proc. 1EEE Con]:
Computer Vision and Pattern Recognition, San Diego, CA, June
1989.
Montanari, U. (1971). On the optimal detection of curves in
noisy pictures. Comm. Assoc. Comput. Mach. 14, 335-345.
Morrone, M.C. and R.A. Owens (1987). Feature detection from
local energy. Pattern Recognition Lett. 6 (5), 303-313.
Nevatia, R. and K.R. Babu (1980). Linear feature extraction.
Computer Vision, Graphics, and Image Processing 13,257-269.
Parent, P. and S.W. Zucker (1989). Trace inference, curvature
consistency and curve detection. IEEE Trans. Pattern Anal.
Mach. lntell. 1 1 (8), 823-839.
Pratt, W.K. (1991). Digital Image Processing, second edition.
Wiley Interscience, New York.
Saint-Marc, E, J. Chen and G. Medioni (1991). Adaptative
smoothing: a general tool for early vision. IEEE Trans. Pattern
Anal. Mach. Intell. 13 (6), 514-529.
Sha'ashua, A. and S. Ullman (1988). Structural saliency: the
detection of globally salient structures using a locally connected
network. In: Proc. 2nd Internat. Conf. Computer Vision, Tampa,
FL, 321-327.
Tabbone, S. (1994a). Cooperation between edges and junctions for
edge grouping. In: First IEEE lnternat. Con]: Image Processing,
Austin, TX, Nov. 1994.
Tabbone, S. (1994b). Multi-Scale Junctions and Edges Detection
with Subpixel Accuracy. PhD thesis (in French), lnstitut
National Polytechnique de Lorraine.
Torre, V. and T.A. Poggio (1986). On edge detection. IEEE
Trans. Pattern Anal. Mach. Intell. 8 (2), 147-163.
Yang, Y.H. (1990). Relaxation Method Applied to Lofargram.
M.S. thesis, Naval Postgraduate School, Monterey, CA.
Ziou, D. (1991). Line detection using an optimal IIR filter.
Pattern Recognition 24 (6), 465-478.