See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/277299989

ON THE USE OF GENERALIZED HOUGH TRANSFORM IN MATCHING DIGITAL

IMAGES

Conference Paper · March 1999

CITATIONS READS

0 801

1 author:

Mohamed Zahran
Benha University
24 PUBLICATIONS   93 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The Nile Delta - Contributed Book Project published by Springer International Publishing in 2017 View project

Integrating Modern Classifiers for Improved Building EXtraction from Aerial Imagery and LiDAR Data View project

All content following this page was uploaded by Mohamed Zahran on 27 May 2015.

The user has requested enhancement of the downloaded file.

 ON THE USE OF GENERALIZED HOUGH TRANSFORM
IN MATCHING DIGITAL IMAGES

M.I. ZAHRAN
Assistant Professor of Surveying and Photogrammetry
Faculty of Engineering at Shoubra, Benha University

ABSTRACT

Many photogrammetric procedures, such as orientation of stereopairs, aerial

triangulation, and DEM generation, depend on image matching. Most of the work
related to matching shapes in digital photogrammetry utilizes parameterized
shapes, where shapes are approximated by analytical functions or polygons. In
this paper a general matching algorithm is proposed based on the principles of
Generalized Hough Transform (GHT). It utilizes non-parameterized shapes and
does not require a polygonal approximation nor does it depend on curve
breakpoints. An approach is also proposed for rotation and scale invariant GHT
in order to enable dealing with stereopair images that have rotation and/or scale
change between them.

In the proposed matching algorithm edge segments are extracted in the

overlapping images using Canny edge detector and an edge tracking algorithm.
The matching procedure is applied for each edge segment in one image against
its matching candidates in the other image. A consistency check for the matching
results is performed utilizing the relationship among edge segments in both
images. Results showed the efficiency of the matching approach in dealing with
broken and occluded edge segments and in achieving good and robust results.

KEY WORDS Digital Photogrammetry, Image Matching, Hough Transform

1
1. INTRODUCTION

The Hough Transform (HT) is an elegant technique for detecting shapes. It is a

popular approach for shape and motion analysis of images which contain noise
or have missing or extraneous data [1,9,10,11,17]. The technique was first
proposed by Hough [8]. Duda and Hart [5] modified it for the detection of straight
lines. The classical HT is used when little is known about the location of of an
object while its shape can be described as a parameter curve [3]. The basic
strategy is to compute the possible loci of the curve reference points in a
parameter space from edge point data in an image space and increment the
parameter point in an accumulator array. The location of the curve is detected at
the peaks in the parameter space. The computational cost increases
exponentially with the number of parameters, limiting the method to curves with a
small number of parameters. However, reduction in the amount of computations
can be achieved if the gradient direction is integrated into the algorithm.
Ballard [2] generalized the Hough Transform to detect objects that have arbitrary
shapes. In the Generalized Hough Transform (GHT) the shape to be detected is
represented in tabular form, called R-table. In this table, boundary points of the
shape, indexed by their gradient direction are stored as a vector relative to an
arbitrary reference point.

2 THE CLASSICAL HOUGH TRANSFORM

2.1 Hough Technique for Straight Lines

The initial work for finding straight lines is presented in [8]. All points on the same
line must satisfy the equation y  mx  c for some particular values m and c.
Suppose that the input image has been processed by a local edge detector to
produce a set of edge points, {(x,y)}, subsets of which are candidate points to lie
on straight lines in the image. For an edge point ( x1, y1 ) on a line L, the equation
y1  mx1  c gives an infinite number of solutions for m and c. These solutions
together form a line in m-c space, or parameter space. Each of the other edge
points that lie on the line L in the image space would yield another line in the
parameter space. These lines will intersect at the point (m,c) in the parameter
space which corresponds to the line L in the image space.

Ballard and Brown [3] suggested the following algorithm to detect lines using the
Hough method: 1. Quantize a parameter space between possible minimum and
maximum values for m and c. 2. Form a 2-D accumulator array of possible m
and c values A ( mmin : mmax , c min : c max ) initialized to zero. 3. For each edge point
(x,y) increment all points (m,c) in the accumulator array that satisfy the equation
c=y-mx 4. Local maxima (peaks) in the accumulator array indicate straight lines
the image space. Each detected line has a number of points equal to the counts
in the corresponding peak.

2
Duda and Hart [5] modified the original Hough technique. In their approach, a
straight line is parameterized in the image space as xcos  +ysin  =r
where  and r are the parameters of the line in a polar representation (Fig.1).
This representation gives a range of  from 0 o to 180 o degrees eliminating the
problem of singularity which might be introduced by large slopes in the y=mx+c
formulation. The transformation from the image space (x,y) to the parameter
space (r,  ) can be implemented as suggested above. The only difference is
that image points are mapped into sinusoidal curves, instead of straight lines, in
the parameter space. Thus, if for a number of image points, the corresponding
sinusoidal curves in the parameter space intersect in a common point ( rj , j ),
then these points would lie on the straight line x cos  j  y sin  j  rj

2.2 Hough Technique for Analytical Curves

The technique explained in Section 2.1 can be generalized to work for any
parameterized curve f(x,a) = 0 where x represents image points and a is the
parameter vector [3]. Suppose that the object to be detected has circular
boundaries. In this case the object can be parameterized by
( x  a) 2  ( y  b) 2  r 2 . For each boundary point belong to this circular object the
resultant loci of parameters is a circular cone. This can be seen from the circle
equation by fixing x and y and changing the values a, b and r. These circular
cones intersect at a common point in the parameter space. The coordinates of
this point are the parameters (a,b,c) of the circular object. The Hough Transform
can be implemented to detect circular objects in an image as follows: 1. Quantize
the parameter space between possible minimum and maximum values for a, b
and r. 2. Form a 3-D accumulator array of possible a, b and r values
A( amin : amax ,bmin :bmax ,rmin :rmax ) initialized to zero. 3. For each edge point (x,y)
increment all points (a,b,r), in the accumulator array, which satisfy the equation
( x  a) 2  ( y  b) 2  r 2 4. Local maxima (peaks) in the accumulator array
correspond to circles in the image space.

The computational cost using Hough Transform grows with the complexity of the
curve. This is because the amount of computations as well as the size of the
accumulator array increase exponentially as the number of parameters. To have
more efficient and computationally economic Hough transform, gradient direction
of edge points should be used. In this case each edge point can be expressed
as ( x, y, y / ) or ( x, y,  ) where  is the slope given as an angle. In the case of
detecting straight lines, y /  tan  = m . This leads to m  y / and c  y  y / x
which is a fully determined case. In the case of detecting circular objects the
parameter locus is reduced from a circular cone to a straight line.

3
3. THE GENERALIZED HOUGH TRANSFORM (GHT)

In Section 2 it is shown how to use the HT technique to find straight lines as well
as parameterized shapes. Here the interest is to generalize the technique to
detect arbitrary shapes which can not be described analytically. Merlin and
Farber [13] were the first who showed how the HT could be generalized to detect
arbitrary shapes at a given orientation and a given scale. They developed their
algorithm without making use of gradient information of the boundary points of
the shape.

Ballard [2] extended the work of Merlin and Farber to formulate more efficient
GHT to detect arbitrary shapes by utilizing gradient information. A general shape
is defined as: F = {C,S,  } where C = (xC,yC) is a reference origin for the shape,
 is the shape orientation, and S = (Sx,Sy) are two orthogonal scale factors. This
definition is simplified by considering S as a scalar which lead to a four-
dimensional subspace. In developing GHT a tabular representation, called R-
table, for the shape is utilized. Each boundary point belong to the shape is
stored in this table as a vector relative to an arbitrary reference point, i.e., a
distance r, and a direction , of the line connecting the boundary point and the
reference point. The resulted list of vectors is organized so that they are indexed
by the gradient direction of the boundary points. Accordingly, when applying the
GHT each image point is only compared with a subset of R-table entries which
have similar edge direction. Fig.2 shows shape representation for R-table and
Table 1 illustrates the form of R-table.

4
 Table 1. The form of the R-table

Gradient Orientation Set of radii r, r = (r,)

1 r11, r21, ... , rn11
2 r12, r22, ... , rn22
. .
. .
. .

m r1m, r2m, ... , rnmm

R-table for certain shape can easily be constructed by examining its boundary
points. A reference point c for the shape is selected first. Then, for each
boundary point x, compute r = c - x and the gradient direction  (x). Store the
vector r as a function of  (x). For a shape that has known rotation and scale, the
reference point coordinates (xC,yC) are the only parameters. Therefore, an image
point p(x,y) with a gradient direction  constraints the locus of the reference
point to be at { x  rp ()cos[ p ()], y  rp ()sin[ p ()].

The GHT for finding arbitrary non-parameterized shapes is implemented in the

following steps [3]: 1. Form the R-table for the shape to be located. 2. Form an
accumulator array of possible reference points A ( x : x ,y :y )
C min C max C min C max

initialized to zero. 3. For each edge point, compute ( ) , calculate the possible
reference points for each table entry for ( ) [ xc  x  r()cos () ,
y c  y  r()sin () ], and then increment the accumulator array
A( xc , y c ):  A( xc , y c )  1. 4. Possible locations for the shape are given by
maxima in array A.

If the shape to be detected is scaled by a scale S and/or rotated by an angle  ,

the accumulator array should be expanded and modified to
A (x : x ,y :y , S : S ,  :  ) . And the possible reference points
C min C max C min C max min max min max

are computed as :

xc  x  r() S cos [() +  ] (1)

y c  y  r() S sin [() +  ] (2)

4. A PROPOSED MATCHING TECHNIQUE USING GHT

The development of the proposed matching system aims at achieving several

objectives. It is desired in this work to utilize shapes as feature primitives in the

5
matching process. Shapes are the most complex primitives to be used in feature-
based matching. Thus, unique feature representation as well as robust matching
results can be expected [7], [18], [20]. Moreover, the more complex the matching
primitives the smaller is the number of the possible candidates and the fewer
are the chances of having ambiguities in the matching process. An important
goal for this work is to have a matching system in which matching results are not
so sensitive to errors in the segmentation process or to partial variations
between the corresponding shapes in two conjugate images. These problems
are challenging to many shape matching methods.

Non-parameterized arbitrary shapes are the type of shapes that are used in the
proposed matching approach. Therefore the principles of Generalized Hough
Transform are utilized in the development of this approach. However the GHT
is required to be modified in order to serve the goal of matching overlapping
images. As described in section 3, GHT is used to detect an instance of a shape
in an image domain. This is accomplished by matching the shape points against
the qualified edge points in the image. This is less complicated than matching
the whole content in overlapping images. In the later case we deal with too
many arbitrary shapes. And it is expected that some of these shapes do have
similar geometry. Therefore matching each shape in one image against edge
points in the other image is not efficient in this case. More efficient approach is
to match shapes in the two images. And the interrelationships among these
shapes are utilized to achieve the consistency of the matching results.

4.1 Components of the Matching technique

The proposed algorithm consists of five components: edge detection using

Canny operator, extracting edge segments from both images utilizing an edge
tracking algorithm, forming R-tables for each edge segment, applying the
matching process on the extracted edge segments and finding the common
subsegment (s) of each matched pair of segments.

4.1.1 Edge detection

In order to extract the matching primitives, Canny’s edge detector is used to find
edges in the overlapping images [4,6]. The detector is implemented in the
following steps:
1. The image is convolved with the Gaussian function [12].
2. The first directional derivative of the convolved image is computed. This is
accomplished by using differential operators that approximate the first directional
derivative [6,14]. Upon that the gradient magnitude at each point of the
convoluted image is calculated.
3. A non-maximum suppression is performed in the direction of the gradient. This
is necessary to make sure that the gradient magnitude of each point is the
maximum in the direction of that gradient.
4. The resulting edge image is thresholded in order to eliminate false edges.

6
5. At last, a fine to coarse technique can be applied to mark additional edges.

4.1.2 Extracting edge segments

The detected edge points are aggregated into edge segments. An edge tracking
algorithm which utilizes both gradient magnitude and gradient orientation is
applied to achieve this goal. The algorithm is implemented as follows:
1. Two thresholds on edge gradient magnitude are specified in order to control
the tracking procedure. The high threshold is used to avoid having spurious
edges with high gradient magnitude as starting pixels. The low threshold is
selected such that spurious edges with low gradient are discarded
2. The edge image is searched for edge pixels with gradient magnitude above
the high threshold. These pixels form the starting points of the tracked edge
segment.
3. The following procedure takes place at every starting point:
. The gradient magnitude and orientation of the eight neighboring points around
each starting point are examined.
. If any of these neighboring pixels has a gradient magnitude above the low
threshold and a gradient orientation that does not differ significantly from the
orientation of the starting pixel, include this neighboring pixel in the tracked edge
segment.
. The process of checking neighboring pixels is repeated recursively until
reaching an edge pixel of a magnitude below the lower threshold and/or an
orientation that differs significantly from the orientation of the preceding pixel.

After edge tracking, edge segments are numbered and a record for each
segment is maintained, including the segment length and the coordinates and
gradient orientation of each segment point.
apply HoughLines()
4.1.3 Formation of R-tables
As explained in section 3, R-table lists each boundary point of the shape to be
represented as a vector relative to an arbitrary reference point, i.e., a distance, r,
and a direction  , of the line from the reference point to the boundary point. The
resulting list of points is arranged so that entries are indexed by the gradient
orientation of the boundary points.
A R-table is formed for each of the extracted edge segments in one image of
the stereo-pair, say the left image. The centroid of each segment can be taken
as a reference point. The centroid coordinates (xC,yC) can be calculated from the
coordinates of the segment points. Therefore, the vector from the centroid to
each boundary point j is of a distance rj  [( x j  xC ) 2  ( y j  y C ) 2 ]1/ 2 and a
direction  j  tan 1[( y j  y C ) / ( x j  xC )] .

4.1.4 The Matching process

The input data for the matching module are a R-table for each extracted edge
segment in the left image and a record for each extracted edge segment in the
right image, including the coordinates and gradient orientation of each segment

7
point. Utilizing these data, the matching process, for each edge segment in the
left image can be implemented as follows:
1. Compute an approximate location for the displaced centroid of the left edge
segment in the right image, using the knowledge about the overlap and image
resolution.
2. To find the corresponding segment in the right image all the right edge
segments whose centroids are located within a specified search window in the
right image are considered to be candidates. The size of the search window can
be computed using the elevation range of object space [15].
3. The matching process is applied for the left segment against each of the
matching candidates in the right image as follows:
a. Get the R-table for the left segment.
b. For each of the matching candidates in the right image form two 2-D array of
size m by n where m is the length of the left edge segment to be matched and n
is the length of the matching candidate in the right image. For each edge point
( xrj , y rj ) along the matching candidate and for each R-table entry  in the range
between (  ) and ( j  t) , where t is an angular tolerance accounting for
j

slight rotation between conjugate edge segments , compute the possible

reference points as follows:
r(ϕ) cos α(ϕ) (3)
r(ϕ) sin α(ϕ) (4)
where x C j and y C j are the possible locations of displaced centroid of the
left segment in the right image. ( xC j , y C j ) values are stored in the two arrays,
one array for the x-coordinates and the other for the y-coordinates as explained
in Section 4.2. The two storage arrays are searched for locations which have
similar x C j and y C j values. Each set of similar values forms a cluster.
c. If there is a cluster whose centroid location lies within the search window, a
matching score is computed as the ratio between the number of counts in the
cluster and the minimum length of the two segments.
d. If the matching score exceeds a selected threshold this will be an indication to
a match.
e. If there are no clusters in the accumulator array or there are clusters of scores
less than the selected threshold, the matching procedure is applied for the left
edge segment against each edge segment in the right image by repeating steps
b to d. This is the expected case of broken and partially occluded segments in
the right image where their centroids will be located outside the search window
(see Section 4.3).

4.1.5 Finding common subsegments

Although the reference points in the left image and their computed locations
establish the correspondence between the matched images, they are not
physical points. Thus it is important to find conjugate points along the matched
edge segments. This is accomplished by tracking the points contributing to the

8
detected peak during the matching process. Using these points, common
subsegments of each two matched edge segments can be found. This can be
carried out in two steps. First, contiguous tracked points along each of the two
matched pair of segments are connected so that two subsegments are formed.
Then the common parts of these two subsegments are extracted based on the
knowledge of the correspondence between the tracked points. This procedure
may not lead to reliable results when matching straight lines of different lengths.

4.2 The accumulation process

The quantization of the parameter space is a necessary step when implementing

the Hough Transform technique. However, the accuracy of the estimated
parameters depend on the level of quantization [16]. The proposed matching
scheme is implemented without the need for parameter space quantization. The
accumulator array is also extended to include the labels of points contributing to
the detected peaks. The accumulation process used is implemented as follows:
a. Two 2-D arrays of size m by n are formed and initialized to zero. The
dimensions m and n are the lengths of the left edge segment and the matching
candidate in the right image respectively.
b. The matching process described in Section 4.1.4, is applied for each segment
in the left image against each matching candidate. The computed possible
locations of the displaced centroid of the left segment are stored in the two
arrays, the x-coordinate in one array and y-coordinate in the other array. The
coordinates of storage location in the two arrays are the labels of the left and
right points contributing to the computation of the displaced centroid.
c. Peaks are found by counting the cells of similar coordinates in both arrays.
The peak center is determined by averaging the coordinates of the points
contributing to it. The spread of the peak is measured by averaging the
coordinate differences between the peak points and their center.

4.3 Broken and Occluded Edges

Owing to distortion, occlusion, and deficiency of the segmentation process, some

of the extracted edge segments in stereo-pair images are broken or occluded in
the overlapping images. If these segments are existed in the right image their
centroid will be located outside the computed search window. And thus there will
be no matching candidates for the corresponding left segment. In this case the
left segment is matched against all edge segments in the right image. If there is
a match, the displaced centroid of the left segment will be located inside the
search window. The accumulation process used in the proposed matching
system enables us to identify each set of broken parts in the right image and
assign them to their corresponding segment in the left image. For example,
suppose that there is a left edge segment labelled C which is broken into three
subsegments C1, C2 and C3 in the right image. As a result of applying the

9
matching scheme, segment C will match each part individually. And the centers
of the three resulting peaks will have the same coordinates. This observation can
be utilized to improve the results of the segmentation process performed before
applying the matching algorithm. For instance, if the C1, C2, and C3 are close
enough to each other, they can be grouped together.

On the other hand, if an edge segment in the right image is broken into
subsegments in the left image the case will be little different. By applying the
matching scheme, each of these subsegments match the right segment. More
importantly is that the coordinate difference between each two subsegment
centroids in the left image is equal to the difference between corresponding
displaced centroids in the right image. Again, these subsegments can be
grouped in one segment in case they are separated by few pixels.

4.4 Consistency Check

A global consistency check is accomplished by establishing the relationship

between the reference points of the edge segments in the left image and their
corresponding displaced locations in the right image. One way is to compute
relative orientation parameters between the stereo-pair images. Another way is
to apply the Generalized Hough Transform for the set of reference points in the
left image against the corresponding set in the right image. The left set consists
of all reference points which have matches in the right image. The right set
includes displaced reference points of all possible matches in the right image. In
this method, the spread of the cluster of the formed peak resulting from matching
the two sets is a measure of the global consistency. And wrong matches will
deviate significantly from the cluster center.

It is worth mentioning that conjugate reference points in the two sets deviate
from the cluster center depending on their disparity value. Larger disparities lead
to larger deviations. Therefore, before performing the consistency check. It is
important to eliminate the effect of the disparity especially for the reference
points of large disparity values. The knowledge about disparity values can be
obtained from the corresponding tracked common points.

5. A Proposed Rotation Invariant GHT

By following the procedure explained above in Section 3 for using GHT in

detecting rotated shapes, the accumulator array expands from a 2-D array to a
3-D array. And the complexity of the transform will increase to the complexity of
detecting a fixed shape times the number of quantized  values for rotated
shapes. Thus it is fruitful to obtain rotation invariant GHT [19].
Assume that the shape to be detected has a fixed scale, say S o , and is rotated
by an unknown angle  in the image space. In this case Equations (1) and (2)
are modified to

10
 xc  x  r() S cos [() + ] (5)
o

y c  y  r() S sin [() + ] (6)

o

Equations (5) and (6) can be rewritten as:

( xC  x) 2 / So  r 2 ()cos 2 [() + ]
2
(7)
( y C  y) 2 / So  r 2 ()sin2 [() + ]
2
(8)

By adding Equation (7) to Equation (8) we obtain

( xC  x) 2  ( y C  y) 2  r 2 ()So
2
(9)

In Equation (9) the values of r() , x and y are known whereas xC and y C are
unknowns. Thus the problem is to determine which location ( xC , y C ) in the
accumulator array A ( x : x ,y :y ) satisfy this equation. This can be
C min C max C min C max

resolved in an inverted procedure. We substitute the coordinates ( xC , y C ) of

each possible cell in the accumulator array in Eqn.9. At the locations ( xC , y C )
where ( xC  x) 2  ( y C  y) 2 is equal to r 2 ()So the accumulator array is
incremented by one. And maxima in the accumulator array indicate possible
locations for the shape. The idea is displayed graphically in Fig.3. Fig.3a shows
a simple shape S which has four points p1 through p4 of distances r1 through r4
respectively from the shape reference point C. Fig.3b illustrates how to compute
the reference point after rotating the shape. At each point of the rotated shape,
circles of radii r1 through r4 are drawn. The reference point of the rotated shape
is the point of intersection of these circles.

However, following this procedure requires tremendous amount of computations.

In addition, it will be difficult to find a unique solution. To reduce the search
space dramatically and introduce robustness to this procedure, one constraint is
imposed. It is the invariance of the difference between the angles  and  for
each shape boundary point with rotation. This property is illustrated in Fig.4. The
figure shows that  -  =  / -  / where  and  are the angles at the shape
boundary point before rotation and  / and  / are the corresponding angles
after rotation. Thus for each boundary point of the rotated shape, cells of the
accumulator array of distance r from the boundary point are checked first for the
fulfillment of this constraint. The qualified array cells only are incremented. It is
worth mentioning that the angle of rotation can be computed as  -  / at each
qualified cell. This angle can be used later as additional matching primitive.

11
6. A PROPOSED SCALE INVARIANT GHT

Similar to the procedure proposed in Sec.5, scale invariant GHT can be

developed. Here assume that the shape to be detected has a fixed orientation,
say  o , and an unknown scale S in the image space. In this case Equations (1)
and (2) are modified to

xc  x  r() S cos [() +  ] (10)

o
y c  y  r() S sin [() +  ] (11)
o

12
Equations (10) and (11) can be rewritten as:

( xC  x) / r()S  cos [() +  ] (12)

o
( y C  y) / r()S  sin [() +  ] (13)
o
By dividing Equation (13) by Equation (12) we obtain

( y C  y) / ( xC  x)  tan [() +  ] (14)

o

In Equation (14) x C and y C are unknowns and the task is to determine which
position (xC,yC) in the accumulator array satisfy this equation. This can be
resolved in a procedure similar to the one described in Section 5. We substitute
the coordinates (xC,yC) of each possible cell in the accumulator array in Equation
(14). At the locations where ( y C  y) / ( xC  x) is equal to tan [() +  o ] the
accumulator array is incremented by one. Possible locations for the shape are
given by the maxima in the accumulator array. The concept can be displayed
graphically in Fig.5. Fig.5a shows the shape S with its four boundary points.

13
Fig.5b depicts how to compute the reference point by the boundary points of the
scaled shape. At each point of the scaled shape, lines parallel to the vectors r1
through r4 are drawn. The reference point of the scaled shape is the point of
intersection of these lines.

Again this procedure requires large amount of computations and does not lead
necessarily to a well-defined solution. To reduce the computational cost
dramatically and strengthen the solution of the problem, an additional constraint
is introduced. It is the invariance of the angle  for each boundary point of the
shape with scale change. Thus for each boundary point of the scaled shape, the
cells of the accumulator array which satisfy Eqn.14 are incremented if they fulfill
this constraint.

7. RESULTS AND ANALYSIS

In order to evaluate the feasibility of the proposed matching algorithm, a software

prototype has been developed and experiments with real stereo-images are
performed. The images are digitized images depicting the campus of The Ohio
State University at a scale 1:4000. They reflect a typical urban area, with tall
buildings and man-made structures. They also have repetitive patterns and many
occluded areas. The size of each image is 512 pixels by 512 pixels. Two 250
pixel by 250 pixel patches, signifying most of the overlap area, are used. They
are named S1 and S2 and illustrated in Fig.6 and Fig.7 respectively.

Fig. 6. Patch S1; a 250x250-pixel patch (left) Fig. 7. Patch S1; a 250x250-pixel patch (right)

The edge segments are extracted by applying Canny’s edge detector and the
edge tracking algorithm to the patches S1 and S2 respectively. For each edge
segment, the coordinates and the gradient orientation of its points are recorded.
14
The centroid of each edge segment in each of the two image patches is
computed by averaging the coordinates of segment points. A R-table is
constructed for each edge segment in the patch S1 based on the computed
centroid.

Patches S1 and S2 are matched using the procedure described in Section 4.1.4.
The threshold for the matching score is selected as 0.5. This means that at least
50% of the points of each of two matched edge segments are contributing to the
detected peak. A consistency check is performed according to the procedure
described in Section 4.3. The reference points whose disparities are significantly
different from the mean disparity are adjusted for the disparity difference. As a
result of the consistency check, a peak in the parameter space is formed with a
computed dispersion of 1.5 pixels. Fig.8 and Fig.9 display the matched segments
in patches S1 and S2 respectively.

Fig. 8. Matched edge segments in patch S1 Fig. 9. Matched edge segments in patch S2

The following three matching cases are obtained:

1. There is a total of 54 unique matches. In this case an edge segment in S1
matches only one edge segment in S2.
2. There are four edge segments in S1 who has two matches in S2. It is noted
that each two matches in S2 are in fact subsegments of their corresponding
edge segment in S1.
3. Two edge segments in S1 match the same edge segment in S2. It is realized
that the two segments are subsegments of the S2 segment.

The dispersion of the associated peak is computed for each matched pair of
edge segments. It is found that they are almost equal to zero. This means that
there is no spread in the detected clusters in the parameter space.

15
The conjugate points along each pair of the matched edge segments in the test
pair are extracted using the procedure explained in 4.2. The locations of these
points in patches S1 and S2 are displayed in Fig.10 and Fig.11 respectively. The
disparity values are computed for the extracted conjugate points along each
matched pair of edge segments. It is found that these values are the same; i.e.,
the figural continuity is preserved after applying the matching procedure.

Fig.10. Location of conjugate points in patch S1 Fig.11. Location of conjugate points in patch S2

8. CONCLUSIONS

A technique for robust feature-based image matching is developed and tested. It

utilizes the principles of the Generalized Hough Transform (GHT) in matching
edge segments of arbitrary shapes. An approach is also proposed for rotation
and scale invariant GHT. The approach overcomes the shortcoming of the GHT
of being not rotation or scale invariant.

This research is a new attempt to match images in the field of digital

photogrammetry. The proposed algorithm use arbitrary non-parameterized
shapes as matching features. This is different from most of the work in the area
of image matching where analytical shapes or shapes approximated by polygons
are rather utilized. The proposed method makes use of the information provided
for each boundary point of the shape and does not depend on curve critical
points or breakpoints, which may not exist due to noise and occlusions.
Some advantages of the technique are: 1. Satisfaction of figural continuity
through matching edge segments, 2. The ability to match edge segments as
whole entities or in parts, 3. The capability of tracking common parts of matched
pairs of edge segments, and 4. Rotation and scale invariance properties.

16
The proposed technique is developed theoretically and then implemented as a
software prototype in order to assess its feasibility. The results are good and
robust. The ambiguity in the matching results is minimal. It is worth pointing that
the proposed matching technique becomes less efficient if there is large disparity
differences in the image space. This is because points deviate from the Hough
peak depending on their disparity values.

REFERENCES

[1] Adamos, C. and Faig, W., “Hough Transform in Digital Photogrammetry,”

Proceeding of International Society of Photogrammetry and Remote sensing,
Commission No.3, pp 250-254, (1992).

[2] Ballard, D.H., “Generalizing the Hough Transform to Detect Arbitrary

Shapes,” Pattern Recognition, Vol. 13, No. 2, 111-222, (1981).

[3] Ballard, D. H. and Brown, C. M., Computer Vision, Printice-Hall,

IncEnglewood Cliffs, New Jersey, (1982).

[4] Canny, J. F., “A Computational Approach to Edge Detection,” IEEE

Transactions on Pattern Analysis and Machine Intelligence, Vol. 8, No. 6, Nov.,
679-698, (1986).

[5] Duda, R.D., Hart, P.E., “Use of The Hough Transform to Detect Lines and
Curves in Pictures,” Communication of the ACM, Vol. 15, No.1, 11-15, (1972).

[6] Haralic, R. M. and Shapiro, L. G., “Computer and Robot Vision,” Addison-
Wesley Publishing Company, Inc, (1993).

[7] Hellwich, O. and Faig W., “Graph-Based Matching of Stereo Image Features,”
Proceeding of International Society of Photogrammetry and Remote sensing,
Commission 3, Washington, D.C, USA, 307-317, (1992).

[8] Hough, P.V.C., “Method and Means for Recognizing Complex Patterns,” U.S.
Patent No. 3069654, Dec 18, (1962).

[9] Illingworth, J. and Kittler, J., “A Survey of The Hough Transform,” Computer
Vision, Graphics and Image Processing, Vol.44, PP 87-116, (1988).

[10] Leavers, V. F., Shape Detection in Computer Vision Using the Hough
Transform, Spring-Verlag, New York/Berlin, (1993).

[11] Leavers, V. F., “Survey: Which Hough transform?,” CVGIP: Image

Understanding, Vol. 58, No.2, Sept., pp 250-264, (1993).

17
 [12] Marr, D. and Hildreth, E., “Theory of Edge Detection,” Proceedings of Royal
Society of London, Vol. 207, pp 187-217, (1980).

[13] Merlin, P.M. and Farber, D.J., “A Parallel Mechanism for Detecting Curves in
Pictures,” IEEE on Computers, C24, pp 96-98, (1975).

[14] Pal, N. R. and Pal, S. K., “A Review On Image Segmentation Techniques,”

Pattern Recognition, Vol. 26, No.9, pp 1277-1294, (1993).

[15] Schenk, T. and Li, J. C. and Toth, C., “Towards an Autonomous System for
Orienting Digital Stereopairs,” Photogrammetric Engineering and Remote
Sensing, Vol.57, No.8, August, pp 1075-1064, (1991).

[16] Van Veen, T. M. and Groen, F. C. A., “Discretization Errors in The Hough
Transform,” Pattern Recognition, Vol. 14, pp 137-145, (1981).

[17] Vosselman, G., Relational Matching, Springer-Verlag, Berline Heidelberg,

(1992).

[18] Wolsford, H. J., “On Curve Matching,” IEEE Transactions on Pattern

Recognition and Machine Intelligence, Vol. 12, No.5, May, pp 483-489, (1990).

[19] Zahran, M. I., Shape Matching Using A Modified Generalized Hough

Transform, Ph. D. Dissertation, Department of Geodetic Science and Surveying,
The Ohio State University, Columbus, Ohio, (1997).

[20] Zilberstein, O., Relational Matching for Stereopsis, Ph. D. Dissertation,

Department of Geodetic Science and Surveying, The Ohio State University,
Columbus, Ohio, (1991).

18

View publication stats