PROCEEDINGS OF THE IEEE, VOL. 67, NO.

5 , MAY 1979 753

Quantitative Design and Evaluation of

Enhancement/Thresholding Edge
Detectors

Absmct-Quantitative design and performance evaluation techniques

are developed fortheenhancement/threshdding class of image edge
detectors. The design techniques are based on statistical detection theory
and detenninistic pattern-mmgnition dnssificab;on procedures. The
performance evaluation methods developed include: a) deterministic
measurement of the edge @ i t amplitude; b) comparison of the
probabilities of conect and false edge detection; and c) figure of merit
computation. The design techniques developed are used to optimally
design a variety of small and large ma& edge detectors. Theoretical and
experimental comparisons of edge detectors are presented.

I. INTRODUCTION Fig. 1 . Edge model.

E DGES are primitive features of an image that are widely

used in image classification and analysis systems to out-
line the boundaries of objects. An image edge, as defined
here, is a local change or discontinuity in image luminance.
A I ,
Spatial G,(j,k)
Operator

There are two basic approaches to image edge detection: the

enhancement/thresholding method and the edge fitting method.
In theformer, discontinuities in animage attribute are en-
hanced oraccentuated bysome spatial operator. If the en-
hanced discontinuity is sufficiently large-greater than some
threshold level-anedge is deemed present. The edge fitting Array
approach involves fitting of an ideal edge replica, a twodimen- Fig. 2. Edge enhancement /thresholding edge detection system.
sional ramp or step function, to the image over some region.
If the fit is close, an edgeis judged present.
Although there is abundant literature [ 1, p. 478 J on the
angle (e), base amplitude ( b ) , contrast ( h ) , and slope width
subject of edge detection; the topic has not yet been covered
in a unified, quantitative manner. This paper presentsquantita- (w).
The edge enhancement/thresholdingedge detection method
tive techniques for the design and evaluationof edge detectors
is described in Fig. 2. In this method, the discrete image array
based on the enhancement/thresholding method. A companion
FG, k) is spatially processed by a set of N linear operators or
paper, in preparation, will consider the design and evaluation masks HiG, k) to produce a set of gradient functions
of edge fitting edge detectors.

II. ENHANCEMENT/THRESHOLDING
LUMINANCE
EDGEDETECTORS where @ denotes two-dimensional spatial convolution. Next,
at each pixel, the gradient functions are combined by a linear
For the purpose of edge detector design and analysis, it is or nonlinear point operator 0 (.) to create an edge enhanced
convenient to defiie an idealized luminance edge, illustrated
may
in Fig. 1, as a planar ramp discontinuity. The ideal edge can
be described by its Cartesian pixel coordinate ( j , k), orientation
Typical forms of the point operator include the root mean
ManuscriptreceivedAugust 25,1978; November 28,1978. This square (rms), magnitude, and maximum. The enhanced array
work w88 supported by the Advanced Research Projects Agency of the
Department of Defense and was monitored by the Wright-Patterson Air A ( j , k) provides a measure of the edge discontinuity at the
Force Base under Contract F-33615-76-C-1203. center of the gradient mask. An edge decision is formed on
I. E. Abdou was withthe ImageProcessingInstitute,University of
SouthernCalifornia, Los Angela, C k He is now with IBM Research the basis of theamplitude of A ( j , k) with respect to a threshold
Laboratory, San J o s e . C k ( r ) . If
W. K. Ratt is withthe ImageProcessingInstitute,University of
Southern California, Los Angela, CA 90007.

0018-9219/79/0500-0753$00.75 O 1979 IEEE

 754 PROCEEDINGS OF THE IEEE, VOL. 67, NO. 5 , MAY 1979

I I -I 3 3 -5
I -2 -I 3 0 -5
I I -I 3 3 -5
i ) Canpass ii) Kirsch
gradient

I 0 -I I 0 -I
I 0 -I 2 0 -2
I 0 -1 I 0 -I
iii)3-level iv) 5-level

(a)
I -I -I 3 -5-5
I -2 -I 3 0 -5
Bs I l l 3 3 3
i) Compass ii) Kirsch
gradient

0 -I -I 0 -1 -2
I 0 -I I 0 -I
1 1 0 2 1 0
iii) 3- level iv)5-level

(c)
Fig. 3. Templatematchingoperators. (a) Compassdirections. @)
Mask H,. (c) Mask Hz.

an edge is assumed present, and if A magnitude pointnonlinearity yielding

4 , k) = IG'U, k)l+ IG2(i, k)l (6b)
no edge is indicated. The edge decision is usually recorded as a is often used for computational simplicity.
binary edge map E ( j , k) where a one value indicates an edge Edge orientation can be obtainedfromthe relationship
and a zero value, no edge. between the horizontal andvertical gradient functions. For
There are two types of spatial edge enhancement operators: the 2 X 2 operators, the edge orientation angle OU, k) with
the differential and the templatematching operators. The respect to the horizontal axis is defined to be
following subsectionspresentexamples of each type of operator.

A . Differential Operators
The differential operators perform discrete differentiation of and for 3 X 3 Operators
an image array to produce a gradient field. This group includes
the Roberts (21, Prewitt [3], and Sobel (4, p. 271 I operators.
The Roberts operatoris a 2 X 2 pixel mask in which'
- -
d ( j , k) = tan-'
E::
-
:3.
0 -1
B. Template Matching Operators
=[1 01
The template matching operators are a set of masks represent-

= [-1 0
0 11.
The Prewittand Sobel operators are 3 X 3 pixel operators where
ing discrete approximations to ideal edges of various orienta-
tion. Fig. 3 givesseveralexamples for two of eight possible
compass orientations. These operatorsinclude the compass
gradient introduced by Prewitt [ 31, the Kirsch [ 5 ] , and the

Hi = [.
1 0 -1

1 0 -1
0 -c]
With the Prewitt operator, c = 1 and with the Sobel operator,
c = 2. These operators usually utilize an rms point nonlinearity
to produce an edge enhanced array
A ( j , k ) = ( [ G l ( i , k ) I ' + [Gz(j,k)l')*- (64
'Note that the masks are rotated by 1 SOo to compensate for the 180' types ofsmallmask
rotation inherent to the convolution operation.
(54
3-and5-level template masks. The latter two operators are
related tothe Prewitt and Sobel differential operators,
respectively. With these operators, the enhancement is formed
as the maximum of the gradient arrays. Thus
A ( j , k) = m v E k)JI* (8)
(5b)
The edge orientation O ( j , k) corresponds to thecompass direc-
tion of the largest gradient. For the3-level and 5-level operators,
the outputs GiG, k) of the fust four masks suffice t o specify
the eight possible edge orientations.
C. Discussion
The specific edge detectors introduced above are widely used
detectors. Many other enhancement/
thresholding edge detectors employing different, often larger,
ABDOU AND PRATT: ENHANCEMENT/THRESHOLDING EDGE DETECTORS 155

@) (c)
Fig. 4 . Edge models for edge orientation sensitivity analysis:
(1 -tan @I1 (12 = (3 tan @ - 1)'
a1 =
8 tan @ 8tan@
(a) 2 X 2 yodel. (b) 3 X 3 model, 0 < @ < tan-' (1/3) (c) 3 X 3
model, ten- (1/3) < @ < n/4

Sobel mag. -0
0
L-

Sobel sq.root
Template matching
Prewitt sq.rooi
Roberts sq. root

Robwts mag,
P

:.=[
0
0.0
, , , wr ,
0.0 .N) A0 .60 .80 1.0
actual edge orientation, 0. rod. actual edge arientation,+,rad
Fig. 5. Edge gradient amplitude response as a function of actual edge Ag. 6. Detected edge orientation as a function of actual edge orienta-
orientation for 2 X 2 and 3 X 3 operators. tion for 2 X 2 and 3 X 3 operators.

mask functions have been proposed. Analysis and comparison Simple geometric calculations canbe performed to deter-
of these mask functions is deferred to Section VII. The inter- mine the edge gradient and detected edge orientation response
vening sections describe design and evaluation methods using as a function of actual edge orientation for the edge models of
the edge detectors introduced in this section as examples. Fig. 4 [61. Results of these calculations are presented in Figs.
5 and 6 . In Fig. 5 , the edge gradient response is normalized to
III. EDGEDETECTORSENSITIVITYANALYSIS unity for a vertical edge. 'The curves indicate that the Prewitt
Desirable properties of any edge detector are an amplitude and Sobel square root differential operators and the template
response invariance to edge orientation and a lack of bias in matching operators all possess an amplitude response relatively
orientation measurement. Edge detector sensitivity is con- invariant to actual edge orientation. TheSobeloperator
sidered here for ideal, noise-free edges. The effect of noise is provides the most linear response between actual and detected
covered in thefollowing sections. edge orientation.
Fig. 4 contains models of ideal step edges passing through Another desirable edge detector property is a rapidly declin-
enhancement masks. Pixel amplitudes under the mask vary as ing edge gradient response as the detector mask moves away
a function of edge orientation as a result of the inherent averag- from a central edge. Fig. 7 contains edge models of vertical
ing associated with discretization of the sampled image array. and diagonal edges displaced from the mask center. Geometric
 756 PROCEEDINGS OF THE IEEE, VOL. 67, NO. 5 , MAY 1979

I I+ horizontal
displacement

Fig. 9. Typical
conditional
probability
density functions of edge
enhancement.

images,however, threshold selection becomes tradeoff

a
between the missing of valid edges and the creation of noise-
horizontol induced false edges. This section considers a statistical design
procedureforthreshold selection. A deterministic approach
is discussed in the following section.
Edge detection can be regarded as a hypothesis testing prob-
@) lem t o determine if an image region contains an edge or con-
Fig. 7. Edge models for edge displacement sensitivity analysis. (a) 2X tains no edge. Let P (edge) and P (noedge) denote theu priori
2 model. @) 3 X 3 model. probabilities of these events. Then, the edge detection process
can be characterized by the probability of correct edge detec-
tion

PD = P(A 2 t ledge) =I t
m

p(Aledge) dA (9)

and the probability of false edge detection

OD

PF = P ( A 2 tlnoedge) =J] p(A1noedge) d A (10)

edge displacement, d where ( t ) is the edgedecision threshold andp(A1edge) and

(8) p(A [noedge) are the conditional probability densities of the
edge enhanced field A ( j , k). Fig. 9 contains a sketch of
typical densities. . The probability ofedge misclassification
error can be expressed as
P(error) = [ 1 - P(edge) + [PFIP(no-edge). (1 1)
This enor will be minimum if the threshold is chosen such that
the edge hypothesis is accepted if

0.0 .so 1.0 LS 2.0

edge displacement,d
and the noedge hypothesis is accepted otherwise. Equation
(b)
(12) defines the well-knownmaximum likelihood ratio test
Fig. 8. Edge gradient amplitude response as a function of edge displace-
ment for 2 X 2 and 3 X 3 operators.(a)verticaledge. associated with the Bayes minimum error decision rule of
@) diagonal
edge. classical detection theory [ 7 1. Decision strategies also exist
for minimizing the cost-weighted error (Bayes minimum cost
calculations yield a relationship between the edge detector test), minimizing PF fora fmed acceptable PD (Neyman-
amplitude response and displacement distance. Fig. 8 contains Pearson test), andminimizing the maximumpossible error
plots of the displacement sensitivity.' All of the operators, with (minimax test) 171.
the exception of the Kirsch operator, exhibit a desired monoton- Application of a statistical decision rule t o edge detection to
ically decreasing responseas a function of edge displacement. determine the threshold value requires knowledge of the con-
ditional densities of the edge enhanced image and the a priori
IV. STATISTICAL DESIGNPROCEDURE edge probabilities. The a priori probabilities can be estimated
Thethreshold value ( t ) on an enhancement/thresholding from images of the class under analysis. Alternatively, the a
type ofedge detector controls the sensitivity of the edge priori probability ratio can be regarded as a sensitivity control
detector. For noise-free images, the threshold can be chosen factor for the edge detector. The conditional densities can be
such that all discontinuities of a minimum contrast level are found in terms of a probability density model for the original
detected as edges, and all others are called noedges. With noisy image plus knowledge of the edge detector operator law. The
 ABDOU AND PRATT: ENHANCEMENT/THRESHOLDING
DETECTORSEDGE 757

TABLE I TABLE I1
MEANVECTOR AND COVARIANCEMATRIXOF DIFFERENTIAL M E A N VECTOR AND COVARIANCE MATRIX OF TEMPLATE
MATCHING
OPERATORS
GRADIENT OPERATORS FOR A VERTICALEDGE

Operator qg Kg
%
Vertical Diagonal 3
- ' 6 4 0 - 4 4 - 4 04-
e No-Edge
Operator Edge Kg
2 4 6 4 0 - 4 4 - 4 0
0 0 -
-2 4
h
-3 -6
-2 -4
0 0 .

Prewitt
- 2, -4

4l
3
common case of additive, independent, white Gaussian noise 0 0
will be given as an example of the design technique. -3 -8
Let f be a Q X 1 vector of pixel amplitudes lying under an 5-Level h
-4 -12
edge detector mask ( Q = 9 for a 3 X 3 mask). Vector f is
assumed to be composed of samples of an ideal edge region s -3 -8 .
plus a Gaussian noise component n of zero mean and common 0 0
variance uz . Thus 3
f =s+n. (13)
The probability density off can be written in general form as

P c f ) = Q ( f , s f , Kr>

where a f and K f are the mean vector and covariance matrix of

f, respectively. For the model under consideration, the noise
is independent and the edge s is deterministic. Hence, a f can
be determined fromthe edge model andKf = $1 where f h an 12'
identity matrix.
8 56 120 56 -8 -12
-12
-12 -8
Since the gradient operation of the edge detector is linear,
the gradient outputs Gi(j, k) are Gaussian random variables. 0

h 02

LGN(j, k) J
be the vector formed by each of the N gradient operators. represent general terms of the mean and covariance matrices of
Then, g , respectively. In (17), Mthe k ( j ) denotesscanned
column
j t h term of the k th convolution mask. To complete the analysis,
Pk)= 9k,?g, K g ) (' 6, it is necessary to examine specific edge detectors.
Tables
where covariance
matrices
I and
and
I1vectors
list mean of
the two-componentRoberts,Sobel,and Prewitt differential
*gW =
j=1
MkW E (17a)
ponent
operators for vertical and diagonal edges and the eightcom-
template matching operators
conditional densities pand ( A ledge)
afor vertical edge. The
p(A1noedge) and the
Q corresponding detection probabilities havebeenderived [6]
Kg(k, I ) = 0' M k ( j ) Ml(i) (17b)
for
the
differential
operators. In the case of
the template
j=1 matching operators, the conditional densities are difficult to
 PROCEEDINGS OF THE IEEE, VOL. 67, NO. 5, MAY 1979

, Sobel diogcnol 8 Prewittvertical , Kirsch

vertical 3-levelg 5-level
0.8

0.6

0.2

0.0
0.0
-

0.2
I
0.4 0.6
I
0.8
I
1.0
0.21
0.oL
0.0 0.2 0.4 0.6 0.a 1.0

PF
(b)
10
. - 1.0
PO
0.8 - 0.8

0.6 -
0.6

0.4 - Robertsvertical
0.4

0.2

0.0 a2 0.4 0.6 0.8 1.0 0.0

0.0 a2 a4 a6 a8 0.1

Fig. 10. Probability of detection versus probability of false detection

for 2 X 2 and 3 X 3 operators. (a) Differential operators, SNR = 10.
(b) Template operators, SNR = 10. (c) Differential operators, SNR =
1. (d) Template operators, SNR = 1.

state explicitly, but the error probabilities can be determined
numerically.
The detection performance of edge detectors can be readily
compared by a parametric plot of the correct detection prob-
ability Po versus false detection probability PF in terms of the
detection threshold ( t ) . Fig. 10 presents such plots for square
root differential operators and templatematchingoperators
for verticalanddiagonaledgesand a signal-to-noiseratio’
(SNR) of 1.0 and 10.0. From these curves, it is apparent that
the Sobel and Prewitt3 X 3 operators are superior to the
Roberts 2 X 2 operators. The Prewitt operator is better than
the Sobel operator for avertical edge. But, for a diagonal edge,
the Sobel operator is superior. In the case of template match-
ing operators, the 3-level and 5-level operators exhibit almost
identical performancesuperior to the Kirschandcompass
gradient operators. Finally, the Sobel and Prewitt differential
operators perform slightly better than the 3-level and 5-level
template matchingoperators.
v. PATTERN CLASSIFICATION DESIGNPROCEDURE
There are two difficulties with the statistical design proce-
dure described in the previous section: reliability of the
stochastic edge model and analytic problems associated with
‘Signal-to-noise ratio is defined as SNR = (h/u)* where h is the edge
contrast and u is the noise standard deviation.
complex edge models such as non-Gaussian signaldependent
noise. Thepattern classificationdesign procedure described
in this section avoids these difficulties.
Edge detection can be viewed as a classical pattern recogui-
tion or classification problem. A pattern consisting of the
pixels encompassed by an edge detection operatoris classified
as a region containing an edge or noedge on the basis of an
extracted region feature, the amplitude A ( j , k) of the edge en-
hancement. Classification can be accomplished by the linear
discriminant function method [ 81 in which the edge hypothesis
is selected if
w*x>o
and rejected if
WTX <0 (18b)
where w( 1) and w(2) are weighting factors of the weight vector
w = [w( l), w(2)l and x = [ A , 11‘. The weight factors are
related to the decision threshold by
Components of w can be determined by a training procedure
using a set of prototype pixel regions containing edges or no-
edges. Let } {xml , x m 2 , * * , x z ~ )
{XI, x’, * * * ,x ~ and -
represent sets of M edge and M noedge prototypes,respectively.
Then,equations(18a) and (18b) canbe reformulated as a
 ABDOU AND PRATT: ENHANCEMENT/THRESHOLDING
DETECTORSEDGE 159

TABLE I11
THRESHOLD
LEVELAND ERRORPROBABILITIES FOR Ho-KASHYAP DESIGN
PROCEDURE

Vertical Edge, SNR = 1 Vertical Edge, SNR = 10 Diagonal Edge, SNR = 1 Disgonal Edge, SNR = 10

experiment theory experiment theory experiment theory experiment theory

Operator r' PO PF PDPF PF PF 7 PO 7 PD PF PD PF r' PD PF PD PF

Roberts
SquareRoot 1.36 52.0 37.6 55.89 39.87 0.67 91.2 11.6 89.16 10.47 1.74 55.60 46.80 55.10 46.88 0.78 74.80 20.80 77.79 22.11

Roberts
Magnitude 1.22 52.0 38.4 55.16 39.30 0.62 90.8 7.6 89.20 9.85 2.24 54.40 46.40 ' 53.85 45.70 0.97 75.60 19.60 76.81 23.30
Sobel
SquareRoot 1.18 60.8 41.2 60.01 39.54 0.66 92.0 92.34
9.6 5.69 1.14 63.20 39.20 60.40 37.63 0.63 90.80 94.65
9.20 5.28

I I 1 I
~~

I
Rewitt
Square Root
Compass
I 1.16 59.5
36.6 60.80 38.40 0.66 93.0 3r8 91.20 4.80 1.19 61.20 39.60 59.27 38.71 0.64 90.00 8.40 93.07 6.421

Gradient 1.52 57.6 45.2 61.27 46.56 0.73 85.2 12.8 88.58 13.55 1.51 57.60 46.80 61.80 47.20 0.71 80.80 14.00 90.00 15.30
I Kirsch I 1.43 56.0 38.4 53.08 34.08 I 0.69 89.2
_ _ _
9.2
_ _
89.78
~ ~~ ~ ~
5.76
~
I 1.45~~~~~
54.40 36.00
~~
52.40 32.40
_____ ~~
1 0.79 ~
82.80 3.60 82.50 2.301
3-Level 1.16 60.8 38.4 59.02 36.92 0.65 6.4
89.6 92.64 1.16
3.79 59.20 38.40 58.70 36.50 0.61 89.60 8.40 94.60 5.60
5-Level I 1.24 58.0 37.6 58.12 36.09 I 0.66 90.8 6.8 92.45 4.90 1 1.22 60.40 39.20 59.30 37.40 I 0.65 90.00 8.40 93.10 5.40

search for a vectorw for which

where
1 i I N=even
integer

If there exists some w satisfying (20), then the prototype data

is said to be linearly separable, otherwise it is linearly non-
separable. In the latter case, a weight vector w is sought to
minimize some misclassification error criterion function. One (dl (e)
common choice is the mean-square miscldication distance Fig. 11. Types of edge, detection errors. (a) Vertical edge test image.
@) Ideal. (c) Fragmented. (d) Offset. (e) Smeared.
J(w)=(Xw-b)T(Xw-b) (22)
note that in most cases the optimum threshold converged to
where b 2 0 is a vector ofpositivebias constants. In this a value for which the error probabilities were approximately
study,the Ho-Kashyap algorithm [4, p. 151, 91 hasbeen equal (PF 1 - Po). This is the same result that is obtained
employed to find the optimum weight vector. Reasons for its by the Bayes minimum error design procedure if edges and no
selection and details of its applications to the edge detection edges are equally probable. Thus in the Gaussiannoisecase,
design problem are found in reference [ 61. similar design results are obtainedwiththe statistical and
An experiment has been performed to evaluate the pattern pattern classifcation design approaches.
classificationedge detector design procedure. In this experi-
ment, sets of 20 edge prototypes and 20 no-edge prototypes VI. FIGURE OF MERITCOMPARISON
have been generated for vertical and diagonal edges embedded The probabilities of correct detection and false detection,
in independent Gaussian noise at signal-to-noise ratios of 1.0 obtained analytically or experimentally,are useful performance
to 10.0. This prototype data has then been used to determine indicators for edge detectors. However, these detection prob-
the optimumthreshold. After thetraining phase was completed, ability functions do not distinguish between the various types
the edge detectors were tested with250 other prototypes. or errors than can be introduced by an edge detector as in-
Optimum thresholds and detection probabilities are tabulated dicated in the example of Fig. 11. Fram and Deutsch [ 101,
in Table 111 for variousedge detectors. It is interesting to [ 1 1I have suggested an edge detector f i of merit based on
a combination of experimentally measured correct and false
' For purposes of comparison, the decision threshold is normalized as detection probabilites. The simple mean-square distance fire
t = t/AM where A M denotes the maximum value of the edge enhance- of merit introduced by Pratt [ 1, p. 4951 has been chosen here
~

meat in the absence of noise. for evaluation of enhancement/thresholdingedge detectors.

 760 PROCEEDINGS OF THE IEEE, VOL. 67, NO. 5, MAY 1979

Fig. 12. Figure of merit test image geometry fordiagonal edge.

100 - 100 -

60 - 80 -
'C6 0
at
-
Compass gradient

Rbberts sq. root

I

0 1 I I I I I I I I I I I I I I I I I 0 I 1 I I I I I l l 1 I I I I 1 1 1 1
1.0 2.0 5.0 10.0 20.0 500 100.0 1.0 2 .o 5.0 10.0 20.0 50.0 100.6
signal-to-noiseratio signal-to-noiseratio
(a) (b)

compass gradient

I I I I 1 l 1 1 1 0 I I I I I I
I I I II I I I I l l
5.0 10.0 20.0 5ao 100.0 10
. 5
.
0 2.0 10.0 20.0 50.0 lOQ0
signal-to-noise ratio signal-to-noise ratio
(C) (dl
Rg. 13. Rguff of merit as a function of signal-to-noise ratio for 2 X 2 and 3 X 3 operators. (a) Differential operators, vertical edge. (b) Tem-
plate operators, vertical edge. (c) Differential o p e r a t a , diagonal edge. (d) Template operators, diagonal edge.

Pratt's f i e of meritmeasurementprocedure utilizes a

square array of pixels with a vertically oriented ramp edge in
its center asshown in Fig. 1 l(a). The edge parameters and
noise level can be varied to generate test edges which are then
processed by an edge detector to produce binary edge maps.
The figure of merit is defined as

F= l ? 1
(23)
max (11,ZA} i= 1 + ad2 ( i )
where ZI and IA are the number of ideal and actual edge points,
d ( i ) is the pixel miss distance of the ithedge detected, and a is 10
. 2.0 90 no 20.0 50.0 loo0
a scaling constant chosen to be a = to provide a relative
signal-to-noise ratio
penalty betweensmeared edges andisolated, but offset, edges.
This technique can be extended to diagonal edgesas indicated Fig. 14. Figure of merit comparison between differenQal and template
in Fig. 12. Edge pixels are only counted in a centrally located matching operators.
 762 PROCEEDINGS OF THE IEEE. VOL. 67, NO. 5, MAY 1979

Fig. 17. Examples of edge detection for airport picture.(a)Original.

(b) 3 X 3 mask. Rewitt differential mask operator. (c) 3 X 3 mask,
3-level template matching operator. (d) 7 X 7 mask, 34evel template
matching operator.

a slightlyhigher f i e of merit than the 3-level template

matching operator for vertical edges. For diagonal edges, the
reverse is true.

w. EXTENDEDM A S K OPERATORS

The design techniques and evaluation methods presented in

the previous sections can be applied to larger size masks with
a variety of weighting factors. Increasing mask size decreases
noise sensitivity because of the inherent noise averaging per-
formed by the operator. However, there is aperformance
penalty associated with large size masks; displaced edges within
the mask contribute to the edge gradient, and therefore, there
is an increased potential for false edge detection. The counter-
vailing effects of noise averaging and edge displacement sensi-
tivity can be examinedjointly be thefigure of merit introduced
in Section VI.
Quantitative studies presented in reference [6] indicate that
extended mask size edge detectors are superior to 3 X 3 pixel
edge detectors at low signal-to-noise ratio. Also, it has been
found 161 that amplitude weighting of masks to decrease the
influence of pixels distant from the mask center is beneficial
in terms of increasing the figure of merit. Fig. 18. Examples of edge detection for TANK picture.(a)Original.
@) 3 X 3 mask, Prewitt defferential operator. (c) 3 X 3 mask, 34evel
template matching operator. (d) 7 X 7 mask, 34evel template match-
W I . EDGEDETECTIONEXAMPLES ing operator.
An important test of the imageedge detector design and
evaluationprocedurespresented previously is their value in boundary points, and have produced relatively few extraneous
extracting subjectivelyrelevant object boundaries in real edges. The edge broadening effect obtained with the 7 X 7
images. Figs. 16 to 18 contain examples of edge detection in 3-level mask is readily apparent from Figs. 16-1 8(d).
three different types ofimages using the 3 X 3 Prewitt dif-
ferential operator and the 3 X 3 and 7 X 7 3-level template IX. SUMMARY
matching operators. In order to provide a comparison between This paper M presented quantitative design andperformance
the three operators, the detector thresholdshave been selected evaluation techniques for the enhancement/thresholdingclass
so thatthe number ofedges detected is the same for all of edge detectors. Two designtechniques have beenintroduced.
operators. In one approach, edge detection is viewed as a statistical deci-
There is very little difference betw.een the edge maps of Figs. sion process, and in the other approach, edge detection is con-
16-18 (b) and (c) produced by the 3 X 3 Prewitt and 3-level sidered as a deterministic patternrecognition classification
operators. Both operators have detected mostof the object task. It has been shown that the two design approaches are
ABDOU AND PRATT: ENHANCEMENTITHRESHOLDING EDGE DETECTORS
consistent. The performance evaluation methods developed
include: a) deterministic measurement of the edge gradient
amplitude; b) comparison of the probabitities of correct and
false edge detection; and c) figure of merit computation..
The design and evaluation techniques have been applied to a
variety of 2 X 2, 3 X 3, and extended size edge detectors. On
the basisof testing, the following conclusions havebeen
formulated.
1) The 3 X 3 differential edge detectors perform appreciably
better than the2 X 2 differential edge detectors.
2) The 3 X 3 Prewitt and Sobel differential edge detectors
are the best of the 3 X 3 pixel differential class of edge
detectors.
3) The 3-leveledge detector is the best of the 3 X 3 pixel
template matching class of edge detectors.
4) The 3 X 3 pixel 3-level template matching edge detector
and the 3 X 3 pixel Sobel and Prewitt differential edge detectors
perform almost equally well as a function of edge orientation
and signal-to-noise ratio. It should be noted that differential
edge detectors require fewer operationsthantemplate edge
detectors.
5) Extended mask size edge detectors are superior to 3 X 3
edge detectors at low signal-to-noiseratios.
6 ) Amplitudeweightingofmasksof extended sizeedge
detectors is beneficial.
763
REFERENCES
W. K. Pratt, Lligital Image Processing. New York, Wiley-Inter-
science, 1978.
L. G. Roberts, “Machine perception of three-dimensional solids,”
in Optical and Electro-OpticalInformationProcesdng, J. T.
Tippett etal., Eds. Cambridge, MA: M.I.T. Press, 1965, pp.
159-197.
J. M. S. Prewitt, “Object enhancement and extraction,” in Picture
Processing and F‘sychopictorics, B. S. Lipkin and A. Rosenfeld,
E&. New York: Academic Press, 1970.
R. 0.Dudsand P. E. Hart, Pattern ChsPification and Scene
Analysis. New York: Wiley, 1973.
R. Kirsch, “Computer determination of the constituent structure
of biological images,” Comput. Biomed Res., vol. 4, no. 3, pp.
315-328, 1971.
I. Abdou,, “Quantitativemethods of edge detection,” Image
RocessingInstitute, Univ. SouthernCalifornia, Los Angeles,
USCIPI Rep. 830, 1973.
K. Fukunaga, Introduction to Statistical PatternRecognition.
New York: Academic Press, 1972.
H. C. Andrews, Introduction to Mathematical Techniques in
Pattem Recognition. New York: Wiley-Interscience, 1972.
Y.-C. Ho and R. L. Kashyap, “An algorithm for linear inequalities
and its applications,”IEEE Trans. Electron. Comput.,vol. EC-14,
no. 5, pp. 683-688,Oct. 1965.
J. R. Fram and E. S. Deutsch, “On the evaluation of edge detector
schemes and their comparison with human performance,” IEEE
Trans. Comput., vol. C-24, no. 6,pp. 616-628,June 1975.
E. S. Deutschand J. R. Fram, “A quantitativestudyofthe
orientation bias of some edge detector schemes,” IEEE T n v n
Comput., vol. C-27,no. 3,pp. 205-213,Mar. 1978.