A Neural System for Managing and Manipulating Directional
Filtering in Image Processing
XIUJUAN GUO1,2
Department of Computer Science
Jilin Architectural and Civil Engineering Institute
27 Honqi Street, Changchun, Jilin Province, China
2
College of Geoexploration Science and Technology
Jilin University, Changchun, China
1
Abstract - Identifying linear information in some selective directions is required sometimes in image
processing. Directional filters can be made in some specific directions so that the interested linear features in
these directions are enhanced. Traditionally the results of directional filtering in these directions are presented
in separate images, which is inefficient in revealing the relationships between linear features in these directions.
In this paper, we propose a new approach that uses a neural system to coordinate all the results from directional
filtering in several directions to produce an integrated image that presents the enhanced linear features in one or
more of these directions. The synthetic and real examples presented in this paper prove that the combination of
neural system and directional filtering is useful for linear feature enhancement.
Key-Words: - Directional filtering, Neural system, Image processing, Linear enhancement
1 Introduction
Various derivative operators are used for linear
enhancement [1]. Directional filters are used to
enhance linear features in a specific direction and
normally based on selective horizontal derivatives.
In theory, the linear features perpendicular to the
filter movement are enhanced and those parallel to
the filter movement are minimized. To illustrate
linear features in several directions of an image, one
needs to make several independent images resulted
from each of these directions. This is inefficient in
revealing the relationships between linear features in
these directions. In this paper, we propose a new
method that coordinates all the results of directional
filtering in several directions using a neural system to
produce an integrated image that combines the
enhanced features from one or more interested
directions. A synthetic example and a real image are
used to test the usefulness of this new technique.
2 Concept of the Neural System for
Coordinating Directional Filtering
Directional filtering can be achieved using space
domain filters, Fourier transformation, Randon
transformation and other methods [1,2,3,4].
Although higher order derivatives can be used as
filters, in this paper we only use the first order
derivatives for performing directional filtering.
Assuming directional filtering has been applied to
an original file (P0) in n directions and the
corresponding files are P1, P2, , Pn, we can design
a neural system to handle and further manipulate
these files. By changing some parameters of such a
neural system, an appropriate image that presents the
enhanced information in one or more specific
directions can be generated. The individual files of
directional filtering in the n different directions form
the input vector (P) of the neural system, and a set of
adjustable factors form the weight vectors (W) of the
system shown in Figure 1. This system also has a
zero bias and a set of adjustable transfer functions.
The output is controlled by adjusting either the
weight vectors or transfer functions, or both.
For example, if the directional filtering is applied
in the x-axis and its conjugated y-axis separately, the
result from the x-axis filtering will enhance the linear
features in the y-axis and minimize the linear features
in the x-axis. In contrary, the result from the y-axis
filtering will enhance the linear features in the x-axis
and minimize the linear features in the y-axis. If we
use the results from both the x-axis (P1) and y-axis
(P2) filtering as the neural inputs, and factors 0 and 1
as the neural weights, the corresponding input vector
will be P = [P1, P2], and the weight vectors will be
W0 = [0, 0]; W1 = [0, 1]; W2 = [1, 0]; W3 = [1, 1]. If
the transfer function is set as F = 1, these weight
vectors will in turn produce a black image, the y-axis
enhanced image, the x-axis enhanced image and the
enhanced image in both the x and y axes,
respectively. The last image will show the combined
linear features in both the conjugated x and y
directions, which is called the conjugate directional
filtering (CDF) [5].
W11
P1

P2
Pn
W1n
F
B
1
Figure 1. Schematic diagram of a neural system
3 A Synthetic Example
We first use a synthetic example to demonstrate the
usefulness of this neural system for handling
directional filtering.
The synthetic model is
comprised of two centred squares and a ‘plus’ sign in
the middle of the squares (Fig. 2a). All objects in
Figure 2a are clearly shown because the differences
between adjacent objects are profound. The model
used to test this system has the same structure as
shown in Figure 2a, but the central ‘plus’ is hidden
because the contrast between the ‘plus’ and the inner
square is very weak (Fig. 2b). After applying
directional filtering in both the horizontal and
vertical directions, the resulted data are used as the
a
b
d
e
input vector (P = [Px, Py]) to the neural system.
Selection of W1 = [1, 0] and F = 1 produces an image
that shows the result of horizontal filtering. Vertical
edges of the hidden ‘plus’ and the inner square are
enhanced whereas the horizontal linear features are
removed except at the ends of the horizontal line of
the ‘plus’ sign (Fig. 2c). Another selection of W2 =
[0, 1] and F = 1 allows the horizontal edges of the
hidden ‘plus’ and inner square to be presented only
(Fig. 2d). The other selection of W3 = [1, 1] and F =
1 generates an image that combines conjugate linear
features in both the vertical and horizontal directions
together. Both of the square and the hidden ‘plus’
sign are clearly outlined (Fig. 2e).
c
f
Figure 2. Structure of the synthetic model (a), original image of the model (b), image resulted from
W1 = [1, 0] and F = 1 (c), image resulted from W2 = [0, 1] and F = 1 (d), image resulted from
W3 = [1, 1] and F = 1 (e), and image resulted from W3 = [1, 1] and F = abs (f).
It is noticeable that the boundaries of the same
object have different grey scales in Figure 2e. This
is because directional filtering always results in a
pair of ‘positive’ and ‘negative’ changes along the
parallel boundaries of a symmetric shape. The
combination of W3 = [1, 1] and F = abs makes all the
boundaries of an object equally grey-scaled (Fig. 2f).
4 An Aerial Photograph of City View
We use the neural system to handle directional
filtering of an aerial photograph of a city view to
demonstrate the usefulness of the system in the
enhancement of conjugated linear features. There
are buildings, streets, trees, cars on parking areas and
other objects on the image (Fig. 3a). Most of these
objects are arranged in either the north-south
direction or the east-west direction, or in the both
directions. Figure 3b shows the result after applying
conventional Laplacian filtering. As expected, the
linear features are outlined on this Laplacian image,
a
but most background information or low-frequency
components, is removed.
Directional filtering is applied to this photograph
in the conjugate EW and NS directions, and
consequently the two processed files are used as the
neural input vector (P = [PEW, PNS]). In order to
integrate these two files into one image that also
retains the original information, a weight vector W =
[1,1,1] is used to manipulate the input. The first
factor indicates that the original file is counted as
one unit in the output whereas the transfer function is
fixed as F = 1.
The input produces an image that shows more
features in both the EW and NS directions (Fig. 3c)
with background information retained.
We may feel that the background information is
too strong in Figure 3c. A weight vector of W =
[1,3,3] can increase directionally filtered information
by 3 times more than the background, which is
shown in Figure 3d.
b
c
d
Figure 3. Original street photograph (a), Laplacian image (b), image resulted from W = [1,1, 1] (c), and
image resulted from W = [1,3, 3] (d).
5 Conclusion
The synthetic and real examples presented above
prove that the neural system is a useful tool to handle
directional filtering for linear feature enhancement.
The adjustable weight vectors in the neural system
allow not only the result from a specific direction to
be displayed independently, but also the results from
more directions to be further enhanced and
integrated into one image, for which the traditional
directional filtering is not able to achieve. The
selective transfer functions in the neural system also
allow further image manipulation to be performed
for some specific purposes.
References
[1] B Jahne, Digital image processing: concepts,
algorithms and scientific applications, Springer,
1997.
[2] S R Deans, Radon transform and some of its
applications, John Wiley and Sons, 1983.
[3] G Beylkin, Discrete Radon transform: IEEE
Trans. Acoustics, Speech and Signal Processing,
Vol.35, pp.162-172, 1987.
[4] J G Proakis and D G Manolakis, Digital signal
processing: principles, algorithms and applications,
Prentic-Hall International, 1996.
[5] W. Guo, D. Li, & A. Watson, Conjugated linear
feature enhancement by conjugate directional
filtering, IASTED international conference on
visualization, imaging and image processing,
Marbella, Spain, 2001.