IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 18, NO.

5, MAY 1996 48 1

Image Filtering, Edge Detection,

and Edge Tracing Using Fuzzy Reasoning
Todd Law, Student Member, /€€E, Hidenori Itoh, and Hirohisa Seki, Member, /€€E

Abstract-We characterize the problem of detecting edges in images as a fuzzy reasoning problem. The edge detection problem is
divided into three staged: filtering, detection, and tracing. Images are filtered by applying fuzzy reasoning based on local pixel
characteristics to control the degree of Gaussian smoothing. Filtered images are then subjected to a simple edge detection
algorithm which evaluates the edge fuzzy membership value for each pixel, based on local image characteristics. Finally, pixels
having high edge membership are traced and assembled into structures, again using fuzzy reasoning to guide the tracing process.
The filtering, detection, and tracing algorithms are tested on several test images. Comparisori is made with a standard edge
detection technique.

Index Terms-Fuzzy reasoning, filtering, edge detection, tracing, joins.

1 INTRODUCTION

I MAGES of real scenes very frequently contain data which

is incomplete and ambiguous. In particular, the problem
of determining what is and what is not an edge in an image
is confounded by the fact that edges are often partially hid-
den (incomplete) or distorted (ambiguous) by various ef-
fects such as uneven lighting. Furthermore, images fre-
quently contain data with edge-like characteristics, such as
high gradient, but a confident classification of such data can
best be resolved when high-level constraints are imposed
on the interpretation of an image.
Human beings are able to make some sense of even un-
familiar objects, which necessarily have an imperfect high
level representation. We suggest that this is possible be- Fig. 1. Some common naive idealizations of edges.
cause our visual systems have built up much experience
viewing different objects under a wide variety of condi- In reality however, edges may deviate from these cases
tions. For edge detection, it appears that the eye-brain is in any number of manners. Fig. 2 shows some examples of
familiar with different types of edges, and how these edges how (continuous) edges may manifest themselves in the
appear under different conditions. real world. Edges may be
To perceive unfamiliar objects, or to perceive familiar ob-
1) gradual in transition,
jects in imperfect images, it appears that humans are apply-
2) between areas of non-uniform intensity,
ing heuristic algorithms to understand such images. Al-
3) between areas of similar intensity,
though these algorithms may be "implemented in the wet-
4) noisy, or
ware of our visual systems, we believe that it is possible to
5) any combination of these problems.
characterize an equivalent process systematically.We submit
6) If these deviations from the ideal are severe enough,
that fuzzy reasoning is a suitable framework for expressing
edge data may become nearly indistinguishable from
heuristic processes applied to incomplete and imperfect data.
noise.
To extract rules for detecting edges, it is helpful to have
some familiarity with different kinds of edges in order to Furthermore, effective edge detection must go beyond
construct a suitable characterization. Typically, researchers local shape characteristics to reflect structural constraints.
have characterized edges as a step function or as a slope Edges appear not only as a local phenomenon, but as part
between two flat regions (see Fig. 1). of larger structures which can also be characterized heuris-
tically. Again, we argue that it is possible to express struc-
tural constraints heuristically in a systematic way.
e The authors are with the ltoh Laboratory, Department of Artificial Intelli- We employ fuzzy reasoning for these tasks because the
gence and Computer Science, Nagoya Institute of Technology, Gokiso-cho,
Showa-ku, Nagoya 466, Japan. E-mail: [email protected]. nature of the data is indeterminate at a low-level stage of
Manuscript received June 27,1994; revised Feb. 5,1996. Recommended for accep-
processing. Fuzzy reasoning, as outlined by Zadeh [30],
tance by K. Boyer. with the power to model and respond usefully to approxi-
For information on obtaining reprints of this article, please send e-mail to: mate situations, is ideally suited to these problems.
[email protected], and reference IEEECS Log Number P96014.

0162-8828/96$05.00 01996 IEEE

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 482 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL 18, NO 5, MAY 1996

gradient in a local region und a pixel Due to the pres-

ence of noise in an image, s usually quite difficult to ac-
curately determine the direction of the gradient. Simply
choosing the direction of greatest ascent is unsatisfactory,
because such a method is vulnerable to minute noise varia-
tions, which have a particularly strong effect in small-scale
regions of a few pixels Furthermore, this method is also
unsatisfactory because there may be several directions with
the same gradient.
Consequently, we offer a new notion of gradient, which
is determined by the characteristics of two regions one ei-
ther side of a dividing line (see Fig. 3) in the digitized im-
age. First, to ignore noise effects as much as possible, we
consider the average value of pixels in each region, ignor-
Fig. 2. Cross-sections of some nonideal (continuous) edges. ing pixels directly on the dividing line. The difference in the
two averages gives us an idea of steepness between the two
In the following sections, we attempt to characterize real regions. Second, to settle the issue of several directions pos-
edges, then develop algorithms based on fuzzy reasoning sibly having the same steepness between their respective
to enhance edges (filtering), detect edge points, and assem- regions, we also measure the symmetry of both regions
ble edge points into edge structures. across the line of the candidate gradient In Fig. 3, it can be
seen that choosing a direction with high reflective symme-
try means choosing the tangent to the surface of an edge. In
2 FUZZY GAUSSIAN FILTERING other words, we claim that the direction of significant gra-
dient is a measure of high steepness between regions, and
Image blurring/sharpening is a common technique in the
high reflective symmetry across the line of gradient. We
edge detection process, with the objective of removing
express this using fuzzy rules, as expressed in Table 1 The
noise and enhancing edges Furthermore, good filtering
exact method of fuzzy calculation is specified in Appendix A.
preserves shape for high curvature edge structures such as
The method of symmetry evaluation is defined precisely in
at corners and T-junctions Typical algorithms adjust the
Section 3.1. Using this method to select the gradient direction,
value of each pixel according to a weighted average of the
the gradient magnitude is simply the difference in the aver-
pixels in a surrounding neighborhood For example, Gra-
age values of pixels on either side of the dividing line.
ham [7] used an algorithm which blurred images across
edges but not along edges, in order to preserve edge infor-
plane In direction
mation Recent versions of such filtering are by Perona and direction with highest of consideration
degree of gradienl
Malik [21] who formulate smoothing as a diffusion process, magnitude and symmetry
is selected
and Nitzberg et a1 [18]who extend this idea to include
preservation of cornerness (high curvature edges)
For any algorithm, it is difficult to know where to blur
and where to sharpen without some idea of where edges
may be located Overblurring can eliminate edge informa-
tion, while oversharpening can produce edges in unex-
pected places Many algorithms have arbitrary parameters,
which must be judiciously selected for a given image, and
may perform poorly with a different image
Consequently, we seek a blurring/sharpening algorithm
which is immune to the particular information content in an
image The algorithm should reduce noise as much as pos-
sible, while not assuming the existence of edges, and si-
multaneously preserving dominant shape
Below, we present a technique for blurring and sharp-
ening images with these objectives in mind Our technique
is a form of controllable Gaussian filtering, where the size,
Fig 3. Evaluating gradient.
shape, and direction of the filter are selected to achieve the
goals of reducing noise while bringing out edges This
technique, which we refer to as fuzzy Gaussian filtering is TABLE 1
documented in [13] DETERMINING
GRADIENTWITH FUZZY RULES
radient
2.1 Determining Gradient small
small
Before we can select a suitable filter shape, it is necessary to large medium
have an accurate idea of the direction and magnitude of the

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 LAW ET AL.: IMAGE FILTERING, EDGE DETECTION, AND EDGE TRACING U$;ING FUZZY REASONING 483

2.2 Controllable Gaussian Filtering 1

(3)
Regular Gaussian filtering can be expressed as the convolu-
tion of the original image I with a Gaussian shaped kernel W
Rotating W to the angle of the gradient, we can define a
P=I* W (1) new coordinate axis system U and v, with W separated as
where W has the (discrete) form

where o, and o, control the squashing in the directions

oeffectively determinesthe breadth of W, and the value of W along and across the edge respectively. We control the two
is constant for a given radius from the center of the function. os using fuzzy control based on local shape information.
Controllable Gaussian filtering permits the size of the We determine the nature of the shape in the local area from
filter, dictated by 0, and the shape of the filter, which we two parameters: gradient magnitude, and straightness.
refer to as squashing (see Section 2.3.11, to be determined These two parameters are determined as follows.
dynamically according to local data characteristics. As mentioned above, the gradient magnitude is deter-
mined by the difference in average values of pixels on ei-
2.3 Filter Dimensions ther side of a dividing line.
Controlling the value of o effectively determines the degree The straightness is determined by comparing pixels
of smoothing of the image. We wish to smooth the pixel in translated along the direction of the edge. Similar to the
consideration with pixels having similar values, i.e., in the determining of symmetry above, we expect that a pixel
same region, but must avoid smoothing with pixels of translated in the direction of a straight edge will line up
greatly differing values which may belong to different re- with a pixel of similar value. Performing such comparisons
gions. In other words, we wish to smooth away small dis- for each of the pixels lying within the evaluation region and
tortions caused by noise, but preserve dominant shape, i.e. then averaging the deviation from the expected values pro-
edges, locally. duces a statistical measure of straightness.
Allowing o to fluctuate over a useful range (set between The reason for choosing these two parameters are as fol-
0.1 and 5.25), we fixed the size of our filter to 7 x 7 pixels. lows. High gradient magnitude indicates the strong possibil-
However, we must recognize that pixels don’t always lie in ity of an edge, where smoothing along the edge should be
even approximately uniform regions, but are often situated done as much as possible, but smoothing across the edge
in regions of changing intensity, due to shape, lighting con- should be limited to preserve this dominant shape. However,
ditions, or other optical effects. Therefore, we loosen the if the edge segment xs curved, sharply angled, or at a triple
restriction of smoothing with similarly valued pixels to a point, then smoothing along the edge will only smear mean-
restriction of smoothing with pixels having similar dope. ingful edge information, and should also be limited. If the
Consequently, we choose to make our filter directional to gradient magnitude is low, this indicates the region is flat,
emphasize smoothing across the dominant slope and de- and smoothing should be in all directions as much as possi-
emphasize along the dominant gradient. We do this by de- ble. To increase the degree of smoothing in a particular di-
forming the filter, or squashing, as explained in the next rection, we simply increase the value of o in that direction,
subsection. where the U direction is along the edge, and the v direction is
Our algorithm considers concentric square regions across the edge. The idea of controlling o, and o, according
around the pixel in question, and measures the gradient for to the gradient magnitude and straightness are expressed as
all opposing pairs of pixels on the perimeter square. This fuzzy rules in Table 2. The calcuation of the two os is deter-
mined according to the procedure in Appendix A.

4
gradient measurement technique is illustrated in Fig. 3.

2.3.I Filter Shape Squashing TABLE 2

FUZZY RULESFOR BLURRING
INTENSITY
Within the selected region, we choose to shape the Gaus-
sian smoothing function according to local data. If the data
small large large
has meaningful shape information in the direction of the small high large large
gradient (i.e., edge or slope point), then it is preferable to large small small
smooth only perpendicular to the gradient so as to preserve lar e small lar e
that information. Our Gaussian smoothing function should
In Table 2, o,and o,are relative values, with the final 0,
be squashed in the direction of the gradient, so that
and onbeing set in (direct) proportion to the filter dimen-
smoothing occurs along the edge or slope, not across it. If
sion, the size of surrounding square area selected as ex-
the data has meaningful shape information both in the di-
plained above.
rection of the gradient, and perpendicular to the gradient
(i.e., corner or T-junction), then it is preferable to tightly 2.4 Experimental Evaluation
limit smoothing in both directions; the Gaussian function
should be squashed in both in the direction of the gradient We stated that the objective of our filter was to smooth in
and perpendicular to the gradient. Rewriting W in polar accordance with the particular information in an image. To
coordinates, evaluate success in achieving these objectives, we created

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 Image added error error after Gauss error after fuzzy Gauss
per pixel SNR per pixel SNR i m p per pixel SNR imp. (dB)
Fig. 4 10.0 15.5 71 ( ~ = 0 5 ) 2.45 2.40 10.4
Fig 5-1 10.0 15 5 9.1( g = 04) 0.58 0.61 84
Fig. 5-2 10 0 15.5 8 0 ( g =0 5) 1 40 2.64 91
Fig. 5-3 10.0 15.5 7.9( 0 = 0.5) I .40 2 78 85

having an average value of 10.0 to each pixel (Fig. 4b), cor- (all (hi) (Cl) (dl)

ues of 0 and evaluated the result (Fig. 4d) when compared (a2J (C2) (d2)

(a3) (b3) (c3)

Fig. 5. Smoothing images with high curvature e

ages, (b) with random added noise, (c) fuzzy Ga
best Gaussian smoothing

The table shows that om algorithm is

ing noise consistently better than Gaussi
optimized value of 0 Furthermore, as it ca
Fig. 5, our filter clearly preserves sharpness
ners and triple points, while providing smoothing else-
where.
Lastly, we applied our filter to a real i
version of the well-known Lena image shown in Fig. 6
Since the image contains inherent noise, no comparison can
be made to an idealized image. However,
interesting regions of the image to see the
ter Fig. 6a shows the original image, and
three-dimensional close-up view of a 50 x 50
white-bordered square in Fig. 6a) Figs. 6c and
Fig 4 Smoothing edges of different sizes (a) original image, (b) with same after fuzzy Gaussian filtering. Note that the triple
added noise, (c) smoothed by fuzzy Gauss, (d) smoothed by regular
Gauss (o=0 5) point is left largely intact while flat regions are well

Second, we created three test images with high curvature

and triple points to evaluate the response of our filter
(similar images were used by Nitzberg et al. [18]to evaluate
the same objective) Fig. 5 shows the images for these ex- and the triple point is largely obscured.

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 LAW ET AL.: IMAGE FILTERING, EDGE DETECTION, AND EDGE TRACING USING FUZZY REASONING 485

gradient direction

'\\\,\yl -\uniformity
evaluation
regions
Fig. 7. Corners and triple points do not satisfy a refined edge definition

processing stage.
To identify the degree of membership in the fuzzy sets
edge, coyner, and triple, we examine several local shape pa-
rameters. We have found that gradient, symmetry, and
straightness are useful parameters for this purpose. An
ideal edge point along a straight segment will have high
gradient, high reflective symmetry, and a high level of
straightness in a local neighborhood of pixels. An ideal cor-
ner point will have high gradient, high reflective symmetry,
but a low level of straightness in a local neighborhood of
pixels. An ideal triple point will have high gradient, low
symmetry, and a low level of straightness in a local neigh-
borhood of pixels. These ideas are summarized in Table 4.
TABLE 4
EDGEFUZZYMEMBERSHIP
RULES
Fig. 6. Filtering applied to Lena image (a) original image, (b) partial 3D grad- sym- straight-
view, (c) fuzzy filtered, (d) partial 3D view, (e) Gaussian filtered, IS = 0.4, ient metry ness
(f) partial 3D view, (9) Gaussian filtered, IS=1 .O, (h) partial 3D view. low low low
low low high
low high low
3 DETECTION low high high
high low low high
3.1 Edges, Corners, Triples high low high high
high high low
The edge detection problem necessitates an adequate char- high high high
acterization of an edge. Intuitively an edge corresponds to
significant intensity change in an image. This leads to at- Gradient (magnitude) was determined as in the previous
tempts to detect edges by locating local gradient maxima. section, i.e., as a difference in averages of neighboring re-
Unfortunately, because of various image phenomena, such gions, with the gradient direction selected for the highest
as noise or distortion from various sources, this is a poor degree of edgeness, as defined in Section 2.1.
definition for meaningful edge detection. To effectively Symmetry y was determined by reflecting pixels across a
characterize edges, we require a more effective definition line in the direction of the gradient passing through the
which identifies the various attributes of an edge. center pixel, and then comparing pixel values (see Fig. 8).
A more refined definition for an edge would appear to Comparisons are made for all possible pixel pairs which fall
be a point of high gradient between two uniformly flat re- in the region of evaluation, which is an n x n square. The
gions. Such a definition excludes many undesirable would- sum total of differences for all pairs of compared values is
be edge-points (such as noise), but may also exclude several then used as a measure of symmetry, y. To make y a posi-
points which are important in edge structures. Namely, tive value which increases with symmetry, we subtract it
such points occur as sharp corners and triple points. In from the largest possible value.
other words, corners and triple points may be points of nl2 n12
high gradient, but may not satisfy the condition of uniform (5)
flatness on either side of the edge point, as shown in Fig. 7
(continuous case shown).
where plly is the value of the reflected pixel across the line
Consequently, it is useful to detect corners and triple
points separately from edge points which occur along rela- in the direction of the gradient.
tively straight segments. Yet, at such an early stage of proc- Straightness t is defined similarily to symmetry, except
essing, we cannot definitely say whether a point is an edge, that pixels are translated rather than reflected (see Fig. 9).
a corner, or a triple point. Therefore, we chose to assign Pixels are translated a random distance in the edge direc-
fuzzy membership values to each pixel for each category. tion, within a limited range so as to fall within the region of
This allows us to delay confident identification until a later evaluation. Again, comparisons are made for all possible

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 486 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL 18, NO 5, MAY 1996

pixel pairs which fall in the region of evaluation. The sum 3.2 Experimentation
total of differences for all pairs of compared values is then
We used the impleme
used as a measure of straightness, t. To make t a positive
relations of Table 4, gen
value which increases with symmetry, we subtract it from
represent edgeness, corne
the largest possible value.

(6)

where plI, is the value of the pixel translated in the direction

of the edge.
reflected

<\
pixel plane in direction images. Fig. 20 shows the algor1
of consideration
version of the well-known Lena i
pixel

/
region for
symmetry
evaluation
dividing line for
symmetry evaluation
(=gradient direction)

Fig. 8. Determining symmetry

edgeness plane, (c) cornerness plane, (d) triple plane

ness enhanced for viewing purposes. In the e

(Fig. lob), darker regions indicate higher e
values. Similarly in the cornerness plane (Fi

in similarity between the two planes.

sible set of conditions results in high cornerness, c

Fig 9. Determining straightness.

edgeiiess conditions are much more common. The corner-

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 LAW ET AL.: IMAGE FILTERING, EDGE DETECTION, AND EDGE TRACING USING FUZZY REASONING 487

ness plane is also less pronounced than the tripleness plane between two segments at a triple point (one end
simply because corners tend to occur less frequently than point, one mid-point)
triple points in real images. which are illustrated in Fig. 12.

AND JOINS
4 TRACING
The algorithm of the previous sections produces four
planes of information, one for each of edgeness, cornerness,
tripleness, and direction. In this section, we present an al-
gorithm for tracing edge segments and for assembling edge
segments into larger structures based on the information in
these four planes.

4.1 Tracing Edge Skeleton

\ . A' triple
point point
/' /
To trace the basic outline of edge segments, we chose a
standard text book algorithm, as outlined in 1261. Briefly,
(C) (4
this algorithm examines the edgeness at each pixel, and the Fig. 12. Different kinds of joins: (a) two aligned edges, (b) two edges at
edgeness of two neighboring pixels on either side of the line a corner, (c) three edges at a triple point, (d) two edges at a triple point
in the direction of edge at that point. If the edgeness of the (mid-point join).
center pixel exceeds the edgeness of the two neigboring
pixels, we mark the pixel as a crisp edge point. Neigh- Of course, we cannot expect the ideal situation of perfect
bouring pixels are determined by the direction of the edge, alignment between edge segments, corners, and triple
as determined in the previous section. In a 3 x 3 block of points. Once again, fuzzy logic is well-suited for evaluating
pixels, there are only four possible pairs of (opposing) such approximate situations. As before, we outline the
neigbouring pixels, which divide the unit circle into 45" qualifications for a good join.
sections, as shown in Fig. 11. A simple edge-to-edge join (Fig. 12a) requires edge seg-
ments to be well aligned, to have end points relatively close
together, and to have pixels of at least medium edgeness in
between. A corner join of two edges (Fig. 12b) requires edge
segments to intersect close to a strong corner point, and to
have end points relatively close together. A 3-edge triple join
(Fig. 12c) requires three edge segments to intersect close to-
gether to a strong triple point, and to have end points rela-
I /I 1'. I
tively close together. A 2-edge triple join (Fig. 12d) requires
the end point of one edge segment to be relatively close to
Fig. 1 1 . Neighbor pixels assigned to 45' segments. any point on another edge segment, and to have a strong
triple point between the end point and the edge segment.
This algorithm only compares relative edgeness, which Once again, we express this fuzzy reasoning in tabular form
we use as an initial guide for finding edge traces. Further- (Tables 5 to 8). In Tables 5 to 8, alignment refers to the differ-
more, we remove very weak edge points by thresholding, ence in angle between two end points, and proximity refers
in order to simplify the tracing process. to the Euclidean distance between two end points.
After this detection process, we also remove any spuri-
TABLE 5
ous patterns which can affect the tracing process, namely FUZZY RULESFOR ALIGNED-EDGEJOIN
any T-, Y-, or cross-shaped patterns. Besides simplifyingthe
tracing process, this cleaning of the edges also removes any
I alignment proximity interim I join-
edgeness I ness
triple points, which we choose to detect explicitly, and use low I low
for assembly of edge segments, as explained in the next high
section.
low medium
high
4.2 Joining Edge Segments high medium
Finally, we wish to assemble the traced edges into larger high medium
structures. Corners and triple points provide natural cues
for such joinery, although we also consider joins which do TABLE 6
not occur at corners and triple points. Altogether, we con- FUZZY RULESFOR CPRNER JOIN
sider four possible kinds of joins,
between two aligned segments (two end points) intersection
between two segments at a corner (two end points) low high low
between three segments at a triple point (three end high low low
points)

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL 18, NO 5, MAY 1996

TABLE 7 In Fig. 14, the results of appl

FUZZY RULESFOR %EDGE TRIPLE JOIN
I tripleness at proxlmlty I ioinness
intersection shown, where the parameter o is set to 1.5. in Fi
low low I low
Canny’s edge detector is appl
low high low
high low IOW (specifymg the fraction of pixel
high high high
border one pixel in width has been adde

intersection
low low I low
low high low
high low low
high high high

We then implemented this fuzzy reasoning process in pro-

gram form. To simplify the joining process, we only con-
sider endpoints which were inside a common 10 x 10
(a) (b)
square. In the case of 3-edge triple joins, such joins can be
detected as a combination of any two of the 2-point joins Fig. 14. Other edge detection methods (a) Laplacian of
(straight, corner, or 2-point triple join), and thus do not re- Canny’s edge detector.
quire a separate implementation.
Again using the Lena image as an example, results from
In practice, triple joins appear the most common,
applying our algorithm are shown in Fig. 13. Fig. 13a shows
lowed by straight joins, and corner joins. For example,
the results of the tracing algorithm, with simple threshold-
Lena image in Fig. 13c contains 74 triple joins, 37 stra
ing of edgepoints with very weak edgeness. Fig. 13b shows
joins,and 19 corner joins. Since triple joins can appear as a
the same image with short edges (less than three pixels)
Combination of corner joins and straight joins, however,
removed. Fig. 13c shows the same image with straight, cor-
these numbers can vary depending on the order in which
ner, and triple joins completed.Fig. 13d shows only segments
they are detected.
which have been connected by at least one kind of join.
We noticed that points at the end of edge segments often
contain the weakest edge information of the entire edge

spurious joins can appe

function to refine the en
specifying that
1) the edge-direction a

If an end point did not sat

m o v e d . The removal p r o c e
of four pixels, until a poi
the end of each segment
joining process.
Figs. 15, 16, 17, and 18 sh
three-stage algorithm (filt
ing/joins) to some other test
age, Figs. 16, 17, and 18 are
while Fig. 19 is 512 by 512 pi
level representation

Fig 13 Final results of edge detector (a) basic edge skeleton, (b) with
short lines (< 3 pixels) removed, (c) with joins completed, (d) with un-
connected lines removed

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 ~-
LAW ET AL.: IMAGE FILTERING, EDGE DETECTION, AND EDGE TRACING USING FUZZY REASONING 489

brim of the hat in the Lena picture. During the tuning of the
joining program, we noticed a tendency for parallel edges
to cross over and fuse to their neighboring edge. These
cases are especially difficult to resolve since they are weak,
yet closely located distinct sets of information. The weak-
ness of the data produces gaps in the tracing of such edges,
and makes joining a sensitive process.
This dilemma suggests that a specialized detector for
nearby parallel edges would be useful. Furthermore, our
joining process results in some joins which are not intuitive,
as can be seen in the eyes and hair of the Lena image. The
decision of whether to join or not to join is linked to the
Fig. 15. (a) original image, (b) after filtering, detection,tracing, and joins. broader object recognition problem, and the problem of
recognizing unfamiliar patterns in images.

5 RELATEDRESEARCH
Fuzzily controlled filters are related to the steerable filters
proposed by Freeman [5], [61. Other fuzzy research applied
to image processing include Pan and Kink 1191 who have
used the concept of fuzzy sets in order to enhance the con-
trast between successive regions in x-ray images. In re-
search by Bezdek [ll, fuzzy versions of c-means clustering
have been developed, and [29] shows an application to as-
tronomical images, while Jung et al. [ l l ] show an applica-
(a) (b)
tion for classifying images of tire treads. Bezdek 121 also
Fig. 16. (a) original image, (b) after filtering, detection, tracing, and joins. describes fuzzy learning vector quantization clustering algo-
rithms with an application to image segmentation. Kim et al.
[12] apply fuzzy logic towards edge relaxation in the prob-
lem of connecting detected edge segments. A nonfuzzy tech-
nique for dealing with imperfect and incomplete data in im-
ages have been developed by Rao and Nevatia [23]. Since our
research also uses simple textural statistical information
(treated in a fuzzy way), it can be related to the various sta-
tistical work summarized by Reed and Hans du Buf [24].
Regarding edge detection in general, a large number of
papers have been published. A discussion of edge detection
methods can be found in Heijden [9], which includes refer-
ences to several prominent and recent papers in the field.
(a) (b) A fuzzy edge detection technique was also introduced
by Tao and Thompson [27]. In this technique, sixteen possi-
Fig. 17. (a) original image, (b)after filtering, detection, tracing, and joins. ble edge structures in a 3 x 3 area are considered, and fuzzy
edge membership is determined by fuzzy if-then rules. Af-
ter redundant edge pixels are discarded, remaining edge
pixels are thresholded based on a noise factor. Filtering and
enhancement possibilities are not considered.
A fuzzy filtering and edge detection technique was in-
troduced by Tyan and Wang [28]. Tyan and Wang use
fuzzy rules for contrast enhancement, requiring arbitrary
definitions of dark and bright pixels, which the authors
admit require adjustment by a human operator. Images are
further enhanced by an interesting fuzzy low-pass filter.
The edge detection scheme also relies on arbritary dark and
bright definitions (again requiring user adjustment), and
operates on a 2 x 2 area. A single test image is presented.
Fig. 18. (a) original image, (b) after filtering, detection, tracing, and joins. Neither of the techniques in [271 and [28] acknowledge
the possibility of various edge shapes such as corners and
triple points, leaving gray-level differences among immedi-
In many cases, close parallel edges appear due to shad-
ate neighbors as the only determining factor in edge extrac-
ows or subtle object shape, as seen in the thickness of the
tion. Furthermore, in both techniques edge direction is ig-

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 490 I€€€ TRANSACTIONS ON P A I T E R N ANALYSIS AND MACHINE INTELLIGENCE, VOL 18, N O 5 MAY 1996

nored completely Also, both techniques operate on very

small regions, ignoring the possibility of broader edges in
the process The relation between neighboring edge points
and connecting edge segments are also not considered. small and large respectiv
This research has unique merit for the following reasons. B', we obtain fuzzy mem
Our techniques acknowledge the existence of imperfect fuzzy sets low and high respecti
edges and of ambiguous edges, and then elaborates a lin- The result C is determined b
guistic framework for characterizing these ambiguihes and
imperfections These characteristics are then approximated
from local statistical parameters, and fuzzy rules based on
heuristics are introduced to evaluate these parameters. This
procedure is performed for image filtering, edge detection,
and edge tracing Our research is the first to attack the edge fuzzy membership
detection problem with a fuzzy reasoning approach, and is
the only research to apply fuzzy reasoning to multistage
(1 e , filtering, detection, and tracing) edge detection method of simplified fuzzy reasonin
to this example gives
6 DISCUSSION/CONCLUSIONS U,, =

In this research, fuzzy algorithms have been developed for

the purpose of filtering images, detecting edge points, and
assembling edge structures These algorithms have been
implemented and tested on a variety of real images with
positive results. Comparisons were made with standard which is a simultaneous versi
image processing techniques. producing a final value for t
Current trends in image analysis show a shift towards constants- (determined empirically) for Slow C and Fast C
top-down image processing techniques. (See for example, respectively, and each term in the numerator corresponds
141.) While our algorithm begins with bottom-level detail, to a rule in Table Al.
we think of it as a bottom-middle algorithm, since it takes
into account intermediate-level structural characteristics. APPENDIXB
Furthermore, we feel that our algorithm appropriately in-
terprets low level results at an intermediate level, without Calculation of signal-to-noise rati
premature assignment. Based on experience with common SNR in this research was perfor
edge structures, our algorithm begins to assemble edge in- is the original image, and g(n,
formation in a meaningful way then the normalized mean squa

APPENDIX
A
In this paper, the calculation of a fuzzy variable is as fol-
lows. A set of fuzzy rules are expressed as in Table A l . After the image has been filtered, the processed image
TABLE AI
p(n,, E,) has NMSE of
FUZZYRULES
EXAMPLE
result C
small
small The SNR improvement due to processing is given by
large slow
slow
SNR improvement = 10 log,,
First the fuzzy membership values, k, for A and B are de-
termined as shown graphically in Fig. Al.
which simplifies to
" P
~ P.,
" ....._.....................
O K
SNR improvement = 10log,,,

2 pa2....... -.-.................
x
4 00
To calculate the SNR alone, if f(n,, n2)is the original signal,
3 0.0 A' A and v(n,,n,) is the added noise signal

Fig A I . Determining fuzzy membership values

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.
 LAW ET AL.: IMAGE FILTERING, EDGE DETECTION, AND EDGE TRACING USING FUZZY REASONING 49 1

ACKNOWLEDGMENT [25] R. Rodieck, “Quantitative Analysis of Cat Retinal Ganglion Cell Re-
sponse to Visual Stimuli,” Vision Research, vol. 5, no. 11/12, -pp.
_ 583-
601,1965.
This research was by the Japanese Ministry Of [261 R.J. Schalkoff, Digital Image Processing and Computer Vision-An
Education. The authors wish to thank the developers of the Introduction to Theory and Implementations, Chap. 6, -pp.
~ - 267-270.
Vista [22]program library which was used for this research. New York: John Wiley & Sons, 1989.
[27] C. Tao, W. Thompson, and J. Taur, “A Fuzzy If-Then Approach to
Edge Detection,” Puoc. Second IEEE lnt’l Conf. Fuzzy Systems, vol.
REFERENCES 11, pp. 1,356-1,360, San Francisco, 1993.
1281 C. Tyan and P. Wang, ”Image Processing-Enhancement, Filter-
J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algo- ing and Edge Detection Using the Fuzzy-Logic Approach,” Proc
rithms. Plenum, 1981. Second Int’l Conf Fuzzy Systems, vu1 I, pp 600-605, San Francisco,
J.C. Bezdek, ”Prototype Generating Clustering Algorithms,” Proc. 1993
Fifth Int’l Fuzzy Systems World Congress, pp. XXXVI-XLIII, Seoul, [291 X L Xie and G. Bmi, ”A Validity Measure for Fuzzy Clustering,”
1993. I E E E Trans Pattern Analysis and Machine Intelligence, vol. 13, no. 8,
J.F. Canny, ”Finding Edges and Lines in Images,” Technical Re- pp 841-847,Aug 1991
port AI-TR 720, Massachusetts Artificial Intelligence Laboratory, 1301 L.A. Zadeh, ”The Concept of a Linguistic Variable and Its Appli-
June, 1983. cation to Approximate Reasoning-I, 11,111,” Information Science, 8,
P. Cavanaugh, ”What’s Up in Top-Down Processing?” Representa- 9,1975
tions of Vision: Trends and Tacit Assumptions in Vision Research, A.
Gorea. ed., pp. 295-304,1991, Todd W. Law (S’83) received the BEng
W., Freeman, ”Steerable Filters and Analysis of Image Structure,” (Honours) degree in electrical engineering from
PhD thesis, MIT, Cambridge, Mass., 1992. McGill University in Montreal, Canada, in 1987.
W. Freeman and T. Adelson, ”Steerable Filters,” Proc. Third Int’l From 1987 to 1990, he worked at Northern
Conf. Computer Vision, Osaka, 1990. 1-elecom Limited. In 1992, he received the MSc
R. Graham, “Snow Removal: A Noise-Stripping Process for TV degree from the Nagoya Institute of Technology,
Signals,” IRE Trans. Information Theory, vol. 8, pp. 129-144,1962, and is currently a doctoral candidate at that
W. Grimson, ”A Computer Implementation of a Theory of Hu- institution Mr. Law is a member of the Japanese
man Stereo Vision,” AI Memo 565, Artificial Intelligence Labora- Society for Information Processing.
tory, MIT, Cambridge, Mass., 1980.
F. van der Heijden, “Edge and Line Feature Extraction Based on
Covariance Models,” IEEE Trans. Pattern Analysis and Machine ln-
telligence, vol. 17, no. 1, pp. 69-77, Jan. 1995. Hidenori ltoh received a BSc degree from Fu-
G. Kanisza, Organization in Vision. New York: Praeger, 1979. kui University in 1969, and an MSEE degree in
S. Jung, S. Bae, Y. Kang, G. Park, and G. Kim, “Tire Tread Pattern 1971 and a PhD degree in 1974 from Nagoya
Recognition Using Hierarchical Fuzzy Pattern Matching AD- Ilniversity. From 1974 to 1985, he worked at
proarh,” Proc. Fiftvh Int’l Fuzzy Systems World Congress, P.; 1y7- Nippon Telephone and Telegraph laboratories,
130, Seoul, 1993. developing operating systems. From 1985 to
J.S.’Kim, H.S. Cho, and S.K. Kim, “An Edge Relaxation Method 1989, he worked at ICOT, the Fifth Generation
Based on Fuzzy Logic and Neural Network Theory,” Proc. Fifth Computer Project, developing knowledge base
Int’l Fuzzy Systems World Congress, pp. 143-146, Seoul, 1993. systems. Since 1989, he has been a professor
T. Law, H. Itoh, H. Seki, ”Filtering Images for Edge Detection at the Nagoya Institute of Technology, and is
Using Fuzzy Reasoning,” Proc. Third Int‘l Conf. Automation, Robot- currently head of the Department of Intelligence
ics, and Computer Vision, Singapore, 1994. and Computer Science. Dr. ltoh is a member of the Japanese Society
M. Levine, Vision in Man and Machine, Chap. 6, pp. 151-210. New for Information Processing
York: McGraw-Hill, 1985.
J.S. Lim, Two-Dimensional Signal and Image Processing, pp. 529-530.
Prentice Hall, 1990. Hirohisa Seki received the BEng, MEng, and
E.H. Mamdani, ”Advances in the Linguistic Synthesis of Fuzzy DrEng degrees from the University of Tokyo in
Controller,” lnt’l 1.Man-Machine Studies, vol. 8, no. 6, pp. 669-679, 1979, 1981, and 1991, respectively. He joined
1976. the Central Research Laboratory of Mitsubishi
D. Marr and E. Hildreth, ”Theory of Edge Detection,” Proc. Royal Electric Corporation in 1981 and engaged in the
Society, (London), ser. B, vol. 207, pp. 187-217,1980. research on artificial intelligence. From 1985 to
M. Nitzberg, D. Mumford, and T. Shiota, “Filtering Segmentation .I 989, he was with the Institute for New Genera-
and Depth,” Lecture Notes in Computer Science, G. Goos, J. Hart- tion Computer Technology (ICOT). Since 1992,
mannis, eds., Chaps. 1-3, pp. 1-49. Berlin: Springer-Verlag, 1991. he has been an associate professor in the De-
S.K. Pal, and R.A. King, ”On Edge Detection of X-Ray Images partment of Intelligence and Computer Science
Using Fuzzy Sets,” IEEE Trans. Pattern Analysis and Machine Intel- at Nagoya Institute of Technology. His current
ligence, vol. 5, no. 1, pp. 69-77,1983. research interests include logic programming, deductive databases
S.K. Pal, ”A Note on the Quantitative Measurement of Image and automated deduction. He is a member of ACM, IEEE Computer
Enhancement Through Fuzziness,” IEEE Trans. Pattern Analysis Society, IPSJ, and JSAI.
and Machine Intelligence, vol. 4, no. 2, pp. 204-208,1982,
P. Perona and J. Malik, ”Scale-Space and Edge Detection Using
Anisotropic Diffusion,” I E E E Trans. Pattern Analysis and Machine
Intelligence, vol. 12, no. 7, pp. 629-639, July, 1990.
A.R. Pope and D.G. Lowe, “Vista: A Software Environment for
Computer Vision Research,” Proc. IEEE CS Conf. Computer Vision
and Pattern Recognition, 1994.
K. Rao and R. Nevatia, ”Describing and Segmenting Scenes from
Imperfect and Incomplete Data,” CVGIP: Image Understanding,
vol. 57, no. 1, pp. 1-23, 1993.
T.R. Reed and J.M. Hans du Buf, “A Review of Recent Texture
Segmentation and Feature Extraction Techniques,” CVGIP Imape
Understanding, vol. 7, no. 3, pp. 359-372, May, 1993.

Authorized licensed use limited to: KTH THE ROYAL INSTITUTE OF TECHNOLOGY. Downloaded on September 24, 2009 at 12:06 from IEEE Xplore. Restrictions apply.