2008 10th Intl. Conf.

on Control, Automation, Robotics and Vision

Hanoi, Vietnam, 17–20 December 2008

An Approach of Canny Edge Detection with Virtual

Hexagonal Image Structure
Xiangjian He, Wenjing Jia and Qiang Wu

Centre for Innovation in IT Services and Applications (iNEXT)

University of Technology, Sydney, Australia
{sean,wejia, wuq}@it.uts.edu.au

Abstract— Edge detection plays an important role in the areas simple pixel interpolation algorithms for simulation of
of image processing, multimedia and computer vision. Gradient- hexagonal structure. The computation cost for edge detection
based edge detection is a straightforward method to identify the using these methods is expensive because its edge values are
edge points in the original grey-level image. It is intuitive that, in based on the intensity values of all sub-pixels (see Section 2
the human vision system, the edge points always appear where for detail) used for simulation.
the gradient magnitude assumes a maximum. Hexagonal
structure is an image structure alternative to traditional square In this paper, a scheme to simulate a hexagonal grid on a
image structure. The geometrical arrangement of pixels on a regular rectangular grid device as presented in [10] is adopted.
hexagonal structure can be described as a collection of hexagonal In this scheme, each of the original square pixels and
pixels. Because all the existing hardware for capturing image and simulated hexagonal pixels is regarded as a collection of sub-
for displaying image are produced based on square structure, an pixels. The light intensities of all sub-pixels constituting a
approach that uses bilinear interpolation and tri-linear hexagonal pixel are computed using a bi-linear interpolation
interpolation is applied for conversion between square and technique. On the other hand, the light intensities of all sub-
hexagonal structures. Based on this approach, an edge detection pixels constituting a square pixel are computed using a tri-
method is proposed. This method performs Gaussian filtering to linear interpolation technique. We will use experimental
suppress image noise and computes gradients on the hexagonal results to show the improvement in edge detection in terms of
structure. The pixel edge strengths on the square structure are
accuracy and efficiency.
then estimated before Canny’ edge detector is applied to
determine the final edge map. The experimental results show that The contributions of this paper are listed in the following:
the proposed method improves the edge detection accuracy and
efficiency. • A fast edge detection algorithm on a hexagonal
image structure is proposed. Unlike the work
Keywords—Edge detection, image interpolation, hexagonal done in [11], the algorithm does not require the
structure, Canny edges, Gaussian filtering computation of light intensities at individual sub-
pixels and hence greatly increases the
I. INTRODUCTION computation speed. The proposed approach is
During the last three decades, many algorithms have been different from the work in [8] in the way that it is
developed for edge detection [1-4], among which the most based on a new hexagonal structure on which
popular one is Canny's computational approach [1] by images have much higher resolution than the
formulating the task as a numerical optimization problem. hexagonal structure introduced in [8]. This results
in higher accuracy of edge detection.
Since 1996, there have been many papers on edge
detection based on hexagonal image structure where an image • The work presented in [10] has shown the
is represented by a collection of hexagons of the same size. superior performance of using bi-linear and tri-
The importance of the hexagonal representation is that it linear interpolation approach compared with other
possesses special computational features that are pertinent to interpolation methods. This paper further
the vision process. Although hexagonal image structure has demonstrates the advantages of the linear
numerous advantages such as higher degree of circular interpolation algorithms using the edge detection
symmetry, uniform connectivity, greater angular resolution, outcomes for image processing.
and a reduced need of storage and computation in image
The rest of this paper is organized as follows. In Section II,
processing operations (see [5-7]), a problem that limits the its
we briefly review the simulation of hexagonal structure
use is the lack of hardware for capturing and displaying
together with bilinear interpolation and tri-linear interpolation
hexagonal-based images [8].
schemes as shown in [10]. In Section III, an edge detection
The edge detection methods as presented in [9] are based scheme is presented. The experimental results are
on a mimic hexagonal structure that uses four square pixels to demonstrated in Section IV. We conclude in Section V.
represent a single hexagonal pixel and hence has a loss of
image resolution. The methods as shown in [8] are based on

978-1-4244-2287-6/08/$25.00 
c 2008 IEEE 879 ICARCV 2008

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on September 26,2023 at 08:55:03 UTC from IEEE Xplore. Restrictions apply.
 II. CONVERSION BETWEEN HEXAGONAL AND SQUARE For the reference sub-pixel (denoted by X) of a given
STRUCTURES hexagonal pixel, there exist four square pixels denoted by A,
B, C and D, as shown in Figure 3, lying on two consecutive
Because there has been no hardware available for image
rows and columns such that, point X falls onto the rectangle
display and capture on hexagonal structure, in this section, we
with vertices at A, B, C and D.
review a software approach to the construction of virtual
hexagonal structure [10]. To construct hexagonal pixels, each
square pixel was first separated into 7×7 small pixels, called
sub-pixels. Each virtual hexagonal pixel was formed by 56
sub-pixels as shown in Figure 1. Figure 1 shows a collection
of seven hexagonal pixels.

Figure 3. A hexagonal pixel X located on a rectangle formed from square

pixels A, B, C and D

Let us denote the coordinates of A, B, C and X by (Ax, Ay),

(Bx, By), (Cx, Cy) and (Xx, Xy) respectively. Let

| Ax − X x | | Ay − X y |
α= ,β = . (1)
| Ax − Bx | | Ay − C y |
Then, it is easy to derive that
X = (1 − α )(1 − β ) A + α (1 − β ) B
(2)
+ (1 − α ) β C + αβ D.
Let f be the image brightness function that maps a pixel
(either square pixel or hexagonal pixel) to its light intensity
Figure 1. A cluster of seven hexagonal pixels value. Then the intensity value assigned to X using the bilinear
interpolation method as shown in [10] is computed as
A. Conversion from Square Structure to Hexagonal Structure f (X ) =
Using Bi-linear Interpolation (1 − α )(1 − β ) ⋅ f ( A) + α (1 − β ) ⋅ f ( B) (3)
The light intensity of any given virtual hexagonal pixel is
+ (1 − α ) β ⋅ f (C ) + αβ ⋅ f ( D ).
assigned to be the light intensity of the sub-pixel that is located
at the fourth row and the middle column of the 56 sub-pixels
forming the given hexagonal pixel. This sub-pixel is called the B. Conversion from Hexagonal Structure to Square Structure
reference (or central) sub-pixel of the corresponding virtual Using Tri-linear Interpolation
hexagonal pixel and it is close to the real centre of the
hexagonal pixel as represented by sub-pixels Pi (i=0,1,2,…,6)
as shown in Figure 2. The light intensity of each square pixel is assigned to be
the light intensity of its central sub-pixel. As shown in Figure
2, each given sub-pixel (and hence each central sub-pixel of a
square pixel) is located on a triangle form by three reference
sub-pixels corresponding to three hexagonal pixels that are
connected each other. Let (x, y) be the coordinates of the given
central sub-pixel, and (x1, y1), (x2, y2) and (x3, y3) be the
coordinates of the three reference sub-pixels respectively. Let

ª1 1 1º
ª y − y3 x3 − x 2 º
A = «« x1 x2 x 3 »»䋬 B = « 2 䋬
¬ y 3 − y1 x1 − x 3 »¼
¬« y1 y2 y 3 ¼»
ª x − x3 º (4)
C=« »;
¬ y − y3 ¼
ª k1 º 1
Figure 2. Reference sub-pixels of virtual hexagonal pixels «k » = ( A ) * B * C ; k1 + k 2 + k 3 = 1.
¬ 2¼

880

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on September 26,2023 at 08:55:03 UTC from IEEE Xplore. Restrictions apply.
 Let ij represent the intensity value of the given central sub- C. Edge Refining
pixel, and ij1, ij2, and ij3 be the intensities of the three After the edge detection step shown in the previous
reference sub-pixels respectively. Then ij is computed from ij1, subsection, all hexagonal pixels have been assigned the
ij2,and ij3 by intensity values that show the edge hexagonal pixels and their
ϕ = k1ϕ1 + k 2ϕ 2 + k 3ϕ 3 . (5) strengths. An edge map on the original square structure can
hence be obtained by simply computing the intensity value of
every square pixel using the tri-linear interpolation as shown in
Thereafter, all pixels in the square structure have been re- Subsection B of Section 2. This edge map shows the square
assigned intensity values. edge pixels and their strengths. We can then follow the
remaining steps of Canny’s method as shown in [1,9] to obtain
III. EDGE DETECTION
the final edge map by using one lower threshold and one higher
The edge detection approach goes through three steps: threshold.
noise filtering using a Gaussian filter, edge detection using
Sobel operator and edge refining using thresholds. IV. EXPERIMENTAL RESULTS
A. Noise Filtering To study the effect of new edge detection method on the
virtual hexagonal structure, and compare with the results
Before the edge map of an image is found, it is common obtained in [11] based on simple image interpolation and the
that image noise is removed (or suppressed) by applying a results on the square structure, 8-bit grey level Lena andn
filter that blurs or smoothes the image. Mary images of size 256×256 are chosen as our sample image
One commonly used filter is implemented by convolution (see Figure 4).
of the original image function with a Gaussian kernel as
defined in Equation (6) below. Let f : ℜ2 → ℜ be the original
brightness function of an image which maps the coordinates of
a pixel to a value in light intensity. Let a0 be the reference
pixel. Then, for a given reference pixel a0 , its new intensity
value, denoted by h( a0 ) , is computed by

n −1 ( ai − a0 ) 2
−
−1
h( a 0 ) = k ¦ f ( a )e i
2σ 2
, (6)
i =0
where k is the normalization constant and is defined as Figure 4. Original Lena image (left) and Mary image (right)

n −1 ( ai − a0 ) 2 Three different edge maps are produced in order to

−
k=¦ e 2σ 2
, (7) demonstrate the performance in accuracy and efficiency
i =0 improved by the proposed edge detection method. The first
and ai (i=0, 1, …n-1) are the n neighbouring pixels of a0. edge map is obtained on square structure. The second one is
obtained using the simple interpolation method as shown in
Considering about 99.5% of energy is found in the central [11] on hexagonal structure. The third one is obtained based
area of “Mexico cap” (the curve of Gaussian function with on the linear interpolation algorithms on hexagonal structure.
parameter ı) within the radius of 3ı, in order to increase the All edge maps are generated with ı=1 and n=49 for Gaussian
computation speed, Equation (6) in this paper is computed filtering. The same lower and higher thresholds are used to
only over a small area surrounding each reference pixel and locate the exact edge points for all edge maps. The higher
covering the disk with centre at the reference pixel and radius threshold is 0.125 and the lower threshold is 0.05.
of 3ı. In this paper, ı is set to be 1. Therefore, the convolution
window is set to be a 49 pixel block on the virtual hexagonal Figure 5(b) shows improved edge maps with clearer edges
structure for the weighted average computation using and less noise points compared with the maps in Figure 5(a).
Equations (6) and (7), centred at the reference pixel. Hence, Overall, Figure 5(b) has fewer blurred edge segments than
for Equations (6) and (7) above, n = 49. The computation to Figure 5(a). This can be seen from the edges at the hair areas.
include the 49 hexagonal pixels for convolution using (6) and On the other hand, those significant edge points on Figure 5(a)
(7) on the hexagonal structure is achieved using the intensities can also be seen in Figure 5(b). This is because an area
of the reference sub-pixels of the hexagonal pixels. consisting of 49 hexagonal pixels is bigger than an area
consisting of 49 square pixels, i.e., the area involved in
B. Edge Detection convolution for Gaussian filtering on hexagonal structure is
In order to implement edge detection on the virtual bigger than on square structure. Therefore, Gaussian
hexagonal structure, a Sobel operator as defined in [8] is convolution on hexagonal structure suppresses image noise
applied in this paper. Using the Sobel operator, edge strength better while keeping important image information.
and direction at each hexagonal pixel (i.e., its reference sub- Moreover, the edge maps on Figure 5(c) look almost the
pixel) can be calculated. same as those in Figure 5(b). However, the computation for
Figure 5(c) is 40 times faster than Figure 5(b) because it no

881

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on September 26,2023 at 08:55:03 UTC from IEEE Xplore. Restrictions apply.
longer request to compute light intensities of all sub-pixels. V. CONCLUSIONS
When converting images from square structure to hexagonal In this paper, an edge detection method has been presented.
structure, we need only compute light intensities for one out of The use of linear interpolation algorithms combined with the
every 56 sub-pixels (i.e., the reference sub-pixels). Therefore, advantages of hexagonal image structure has achieved
the edge detection based on linear interpolation algorithms is encouraging and promising edge detection performance. We
more efficient than simple interpolation. have shown that the use of linear interpolation results in more
efficient edge detection and hexagonal structure leads to more
accurate and less noise edge maps compared with the state-of-
the-art Canny edge detector.

REFERENCES
[1] J. F. Canny, A computational approach to edge detection, IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol.8 1986,
pp.679-698.
[2] G. Economou, S. Fotopoulos, and M. Vemis, A novel edge detector
(a) Edge maps on square structure
based on nonlinear local operations, Proc. IEEE International
Symposium on Circuits and Systems, London, 1994, pp.293-296.
[3] D. Marr and E. Hidreth, Theory of edge detection, Proc. Royal Soc.,
1980, pp.187-217.
[4] R. M. Haralick, Edge and region analysis for digital image data,
Computer Vision Graphics and Image Processing, 1980, pp.60-73.
[5] R. Staunton, The design of hexagonal sampling structures for image
digitization and their use with local operators, Image and Vision
Computing, Vol.7, No.3, 1989, pp. 162-166.
[6] B.K.P. Horn, Robot Vision. MIT Press, New York, 1986.
[7] C.A. Wuthrich and P. Stucki, An algorithmic comparison between
square- and hexagonal-based grids, CVGIP: Graphical Models and
Image Processing, 53(4), 1991, pp. 324-339.
[8] Xiangjian He, Wenjing Jia, Namho Hur, Qiang Wu, Jinwoong Kim and
(b) Edge maps based on simple interpolation on hexagonal structure Tom Hintz, Bi-lateral Edge Detection on a Virtual Hexagonal Structure,
Lecture Notes in Computer Science (ISVC2006), LCNS, Springer, 2006,
Vol.4292, pp.1092-1101.
[9] Qiang Wu, Xiangjian He and Tom Hintz, Bilateral Filtering Based
Edge Detection on Hexagonal Architecture, Proc. 2005 IEEE
International Conference on Acoustics, Speech, and Signal Processing,
Philadelphia, PA, USA, Volume II, 2005, pp.713-716.
[10] Xiangjian He, Jianmin Li and Tom Hintz, Comparison of Image
Conversions between Square Structure and Hexagonal Structure,
Lecture Notes in Computer Science (ACIVS07), J. Blanc-Talon et al.
(Eds.): ACIVS 2007, LNCS 4678, pp. 262–273, 2007.
[11] Xiangjian He, Wenjing Jia, Jianmin Li, Qiang Wu and Tom Hintz, An
Approach to Edge Detection on a Virtual Hexagonal Structure, Digital
Image Computing Techniques and Applications (DICTA2007), pp.340-
345.
(c) Edge maps based on linear interpolation on hexagonal structure

Figure 5. Edge detection results with Gaussian filtering

882

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on September 26,2023 at 08:55:03 UTC from IEEE Xplore. Restrictions apply.