Вы находитесь на странице: 1из 6

3rd International Conference on Advanced Technologies

for Signal and Image Processing - ATSIP'2017


May 22-24, 2017, Fez, Morroco

A 2D- Fractional Derivative Mask for Image


Feature Edge Detection
Meriem HACINI Akram HACINI Fella HACHOUF
Laboratoire d’Analyse and Herman AKDAG Laboratoire d’Automatique et de Robotique,
des Signaux et Systémes LASS, Laboratoire d’Informatique Avancée Université des Fréres Mentouri,
Université Mohamed Boudiaf, de Saint-Denis (LIASD), France Constantine, Algérie
M’sila, Algérie akram.hacini@etud.univ-paris8.fr,
Email: meriem.hacini@yahoo.fr herman.akdag@ai.univ-paris8.fr

Abstract—Feature extraction is a classic problem of ma- established to define the fractional differentiation. These
chine vision and image processing. Edges are often detected methods include DuttaRoy  s method [6], Carlson s method
using integer-order differential operators. In this paper, a one- [7], P odlubry  s method [8] and Charef  s method [9], among
dimensional digital fractional-order Charef differentiator (1D-
FCD) is introduced and extended to 2D by a multi-directional them the most famous definitions of Riemann − Liouville,
operator. The obtained 2D- fractional differentiation (2D-FCD) is Grümwald − Letnikov and Caputo definitions[10- 12].
a new edge detection operation. The computed multi-directional They have been applied in multidisciplinary domain, such as
mask coefficients are computed in a way that image details are economic, automatic, control, physics, mechanics and signal
detected and preserved. processing. In recent days, fractional theory has been already
Experiments on texture images have demonstrated the efficiency
of the proposed filter compared to existing techniques. utilized for a multitude of problem in image processing,
Index Terms—Edge Detection, Fractional-Order Derivative,
such as image de-noising [13-14], image segmentation [15],
Mask. image enhancement [16]. However, this concept has been
rarely used in features extraction as edges. In this work, we
aimed on the edge and feature detection. Usual integer-order
I. I NTRODUCTION
differential operators are essential tool in edge detection,
Most of all basic image operations is edge detection. Edges but they damage the texture details too. To deal with this
are useful in image analysis and interpretation in computer problem, a 2D-Fractional Differentiation, 2D − F CD mask
vision. The definition of edge is tributary on the scope. is proposed. The coefficients of the proposed mask are multi
Although, it is generally defined as a line between related directional. They are computed based on Charef method.
regions with relatively different features according to some The obtained results have demonstrated that the 2D-FCD has
useful characteristics such as gray level or luminance, color the capacity to detect effectively image edges.
and texture, etc,... The outline of this paper are as follows. In Section. 2,
Edge detection is the fact to define what are pixels belonging detailed implementation of 1D-FCD is introduced. In section.
edge pixels. There exist a multitude of edge detection 3, a 2D- formulation of the proposed operator is defined
methods formulated over the last three decades; where we generalizing the FCD operator from one dimension to two
can distinguished those based on the first-order derivative: dimensions. Then a multi-directional mask is constructed
Roberts operator [1], Sobel operator [2], P rewit operator [3], to improve the proposed operator. Numerical examples are
the second-order derivative: Laplace Operator [4], assuming presented in Section. 4, and a conclusion is given in Section. 5.
that, image edges and noise both of maps edges and noise
belong to high frequency components. In order to deal with
noise and its effects. Edge detection algorithm have been II. 1D-D IGITAL FRACTIONAL - ORDER C HAREF
widely developed. Hence noise and edge detection accuracy DIFFERENTIATOR
are conflicting. How to balance between edge detection and
noise is challenging. A new solution was presented has been Basically, fractional derivative is a derivative of any depen-
presented in this work. dent function f (t) with real order ν. In general, derivative
Since three decades, fractional calculus has received a has conventionally an integer order. Fractional derivative are
growing interest and has become an important branch newly used in a multitude of applications such as filter design
of mathematics analysis [5]. Recently, fractional calculus [17], image processing [18] and soon. Several definitions of
reached a considerable recognition by the researchers. the fractional models have been presented by researchers [1-3],
A multitude of mathematical formulations have been [6-12]. Due to less time complexity, Charef [9] definition has
been used in this paper. The mathematical Charef fractional
978-1-5386-0551-6/17/$31.00 2017
c IEEE derivative formulation is:

1
The ν − th order Charef derivative of a function f (t) is must beings exposed. Assume the filtering window’s size
defined as: p = 2m + 1, and m is a positive integer. Using formula (4),
for ν > 0 as: the νth−order derivative of the filtering window central point
is estimated by:
∂ ν f (t)
f ν (t) = Dν f (t) = ν (ν) (ν) (ν) (ν)
∂tν wm+1 = w0 , w1 , w2 , ..., , wm+1 . (6)
∞ N   k
1   2 gi βi
= lim+ ν [KD δ(k) + H(k) However, we find some problems when we apply the wm+1
(ν)
h→0 h h αi αi
k=0 i=0 in 2D- image processing. From equations (7) and (8) we can
N   k−1
2 gi βi observe that edge feature are extracted along horizontal and
− H(k − 1)]f ((n − k)h) (1) vertical directions with the same coefficients. Thus, wm+1
(ν)
i=0
h α i α i
is anisotropic it can be used to calculate the fractional-order
Where n = t/h is the step size, gi , αi and βi are the derivative in x or y directions. So, it is not effective because
approximation parameters defined as follows [9]: some edge features will be lost. Figure. 1(b) illustrate the result
For i = 1, 2, 3, ..., N obtained using the mask given by equation (6). It can be seen
⎧ ν+i
[ 10ν(1−ν) ] that the major features are extracted, but some details are not

⎪ αi = 1 + 2.10

⎪ distinguished with a clear appearance of false contours.

⎪ βi = αi − 2

⎪ Let f (x, y) denotes the gray value of the position (x, y) in the

⎪ N
⎨ i−j+ν
[1 − 10[ 10ν(1−ν) ] ] image f .
(2) Using the 2D-FCD mask described on equation (6), the ν -

⎪ gi = −KD
j=0

⎪ N order derivatives of f (x, y) in the x− and y− directions can

⎪ ν+i
[ 10ν(1−ν) ] i−j


⎪ 10 [1 − 10[ 10ν(1−ν) ] ] be calculated by

⎩ j=0
i=j ∂ ν f (x, y)
≈ w0 (ν)f (x, y) + w1 (ν)f (x − 1, y)+
1
(80ν+1) ∂xν
Where KD = 10−ν and N = integer[ 2 (1−ν) ], N w2 (ν)f (x − 2, y) + ... + wL−1 (ν)f (x − L + 1, y) (7)
is the number of approximation parameters for a better
approximation.
∂ ν f (x, y)
KD is a constant ν−order dependent. ≈ w0 (ν)f (x, y) + w1 (ν)f (x, y − 1)+
In practice, the Dirac function δ(k) is given by ∂y ν
w2 (ν)f (x, y − 2) + ... + wL−1 (ν)f (x, y − L + 1) (8)

1 if k == 0
δ(k) = H  (k) = (3)
0 if k =
 0.
B. The improved 2D-IFCD mask
Where H(k) is the Heaviside function.
Considering numerical calculation, we rewrite the discrete In this section, a multi-directional mask is proposed. The
approximation at point t of one-dimensional left-sided mask is rotational invariant. If the image is processed with a
fractional differentiation as: nonlinear filter, the values of its pixels are convolved with a
pxq size mask using:

∂ ν f (t) ν 1  a
 b

= D f (t) = wk (ν)f (t − kh) (4) ν
∂tν hν
k=0
g (x, y) = W ν (p, q)f (x + p, y + q) (9)
p=−a q=−b
Where h is the sampling step, t = kh, for (k = 0, 1, ..., L)
in the interval [0, t], equal to 1 in our case. wk (ν) are the Where f (x, y) is the value of a pixel, and W ν (p, q) is the val-
approximation coefficients. ues of the mask. To ensure that the mask wmxn has a specific
central point, m and n must be odd. Then, a = (m − 1)/2 and
N   k b = (n − 1)/2 and the distance between the mask center and
2 gi βi
wk (ν) = KD δ(k) + H(k) the margin of the image of the image is less than (n − 1)/2
i=0
h α i α i
pixels. Furthermore, an improved fractional-order differential
N   k−1
2 gi βi masks 2D − IF CD in the x and y directions, have been
− H(k − 1). (5) proposed (Figure.2). It is constructed to extract texture features
i=0
h α i α i
and image details. The proposed masks described in Figure.2
III. 2D-FCD- BASED EDGE DETECTION METHOD are complementary. They are selected in this way to better scan
vertical, horizontal, left and right diagonal pixels surrounding
A. Generalization
the central pixel. The fact that the proposed mask can consider
This section, introduced the extension of the fractional more neighboring pixels, is an excellent tool for feature and
Charef differentiation from 1D to 2D. With a view to image details extraction, which can preserve texture and edge
explain more clearly the process, some fundamental definitions details.

2
(a) Original image (b) 2D-FCD
Fig. 1. The edge extraction result of the proposed operator on two different scene.

an IntelCorei5 of 2.2GHz, 4GB memory, using 64bits


W indows10.
In order to illustrate the performance of the Improved 2D-
IFCD mask, gray-scale artificial and real images have been
chosen.The computational CPU Time of our algorithm are also
compared.

A. Comparison of the Improved 2D-IFCD with integral dif-


ferential operator mask
Most typical differential operator masks such as the first-
(a) (b) order Sobel, P rewit and Roberts operators, and the second-
order Laplace operator, are in fact integral differential oper-
Fig. 2. Fractional-order differential masks in the (a): x- and (b): y- directions
ators. They may perform well in the high frequency regions
but their performances will deteriorate when applied to rich
Feature extraction by the proposed mask are denoted as texture patterns in smooth regions.
Fxν and Fyν respectively. They are the result obtained from Figure. 4 illustrate obtained results by all the operators. Results
the correlation of the Wxν and Wyν masks and image. Hence, of theLaplace and Sobel are shown in Figure.4(b) and (c),
calculate the gradient magnitude of the image, which is an respectively. The results show that texture in some areas is
important quantity in edge detection. removed (Figure.4(b)). In conjunction, an excessive enhance-
As for classical integer order differentiation, the fractional- ment of the high frequency components appears, Hence, wide
order magnitude measure can be computed as: white edges in these regions are created Figure.4(c).
1
Different from the integral differential operators, the proposed
∇ν f = mag(Δν f ) = [(Fxν )2 + (Fyν )2 ] 2 (10) method takes into consideration texture details. It avoids false
Where mag denotes the magnitude. This quantity gives the detection and it emphasis the high frequencies Figure.4(d).
maximum rate of increase or decrease of f (x, y) per unit In conclusion the 2D − IF CD operator outperforms the
distance in the direction of ∇ν f . traditional integral differential operators.
Extracted feature on image of Figure.1(a) by the proposed
mask are shown in Figure.3(c), for ν = 1.5. It can be seen B. Comparison with other fractional differential operators
that the proposed 2D − IF CD mask gives best results. Fractional differential methods are able to preserve
texture detail. However, traditional methods usually ignore
IV. E VALUATION OF THE I MPROVED 2D-IFCD OPERATOR
local texture features. The performance of the proposed
FOR AN EDGE FEATURE DETECTION
algorithm is compared with Grümwald − Letnikov, [11]
In this section, the obtained results obtained by the 2D − and Riemann − Liouville [10] method in Figure.5. As
IF CD have been compared to those of four other meth- shown in Figure.5 (b), while details are well extracted by
ods: Sobel operator [2], Laplace operator [4], Riemann − Grümwald − Letnikov method in all cases, the overall
Liouville operator [10 ] and Grümwald − Letnikov op- brightness of the images decreases. Figure.5(c) shows the
erator [11]. To ensure the quality of experiments, optimal edge detection results based on the Riemann − Liouville
parameters for the above selected methods were chosen. We fractional differential operator, the method cannot effectively
implemented the proposed algorithm using M atlab2015a on take into consideration the complexity and diversity of local

3
(a) Original image (b) 2D-FCD (c) Improved 2D-IFCD
Fig. 3. The edge extraction result of the proposed operator on two different scene.

texture patterns, thus leading to only marginal edges in of the main limitation of the previous fractional derivative
Figure.5(c). Our approach improves the self - correlation in method in terms of edge detection, texture details preservation,
images and therefore it can detect the local complex texture the 2D operator is finally improved by defining x− and y−
patterns. It is noticed that, the curved patterns of image objects multi-directional operators, and a 2D − IF CD based edge
as parrot, fruits and girl face are well detected (Figure.5(d)). extraction have been proposed. Numerical experiments have
Textures as wool, hair and feathers are more distinct and been carried out to evaluate the performance of 2D − IF CD-
brighter. Table.1. summarizes the average CPU process based edge extraction method. The obtained results have
time of different fractional methods. i.e., The proposed demonstrated its validity. In the feature, the proposed method
2D − IF CD mask presents the best processing time it is will be extended to an adaptive fractional-derivative operator.
able to yield the best result in less CPU time than the G − L Moreover, due to low computational complexity of the method,
and R − L operators. In summary, the proposed algorithm a future direction may be used of the proposed operator in
can effectively preserve texture details and brightness. The image processing.
2D − IF CD operator presents a lower CPU time and more
effective results than the other compared operators. R EFERENCES

Table1: The average CPU Process time of different fractional [1] L. G. Roberts and P. W. Daly, Machine Perception of Three-Dimensional
Solids, in: J.T. Tippett et al (Eds), Optical and Electro-optimal Information
differential methods. Processing, Mass: M. I. T. Press, Cambridge, pp. 159-197. Cambridge,
1965.
[2] T. E. Sobel, Camera Models and Machine Perception, PhD Dissertation,
Image Stanford University, pp. 159-197. Palo Alto, Californie, 1970.
Tested G-L 2D-IFCD [3] J. M. S. Prewitt, Object Enhancement and Extraction, in: B.S.Lipkin,
size T-L (sec)
image (sec) (sec) A. Rosenfeld (Eds), Picture Processing and Psychopicotorics, Academic
(pixels) Press, New York, pp. 75-149. Cambridge, 1970.
California 733x490 98.366 85.036 14.044 [4] E. R. Davies, Design of Optimal Gaussian Operator in Small Neighbor-
Parrot 256x256 8.824 8.065 4.565 hoods, Image Vision Comput. 5(3) pp. 199-205. 1987.
[5] K. B. Oldham and J. P. Spanier, The Fractional Calculus,Academic, New
Fruits 512x512 64.544 52.050 11.026 York, 1974.
Girl 512x768 129.071 121.386 15.066 [6] S. C. Dutta Roy, On the Realization of a constant-argument emittance or
fractional Operator, In: IEEE Trans. Circuit Theory 14(3), pp. 264-274.
1967.
V. C ONCLUSION [7] G. Carlson, C. Halijak, Approximation of Fractional Capacitors
(1/s)(1/n) by a Regular Newton Process, In: IEEE Trans. Circuit Theory
In this paper firstly introduces the one-dimension fractional - 11(2), pp. 210-213. 1964.
[8] I. Podlubny, I.Pertras, B. M. Vinagre, P .O’Leary, L. Dorcak Analogue
order Charef differentiator (1D−F CD). Secondly, the 1D− Realization of fractional-order Controllers., Nonlineaar Dyn. 29(1-4), pp.
F CD has been extended to the 2-D. Through the identification 281-296. 2002.

4
(a) (b) (c) (d)
Fig. 4. Comparison between integral differential operators and our operator: (a) original, (b) Laplacian, (c) Sobel, and (d) The Improved 2D-IFCD result.

[9] A. Charef, H. H. Sun, Y. Y. Onaral, B . Onaral, Fractal System as Publications. pp. 1-6. 2014.
Represented by Singularity Function, In: IEEE Trans. Autom. Control. [15] P. Ghamisi, M. S. Couceiro, J. A. Benediktsson and N. M. F. Ferreira,
37(9), pp. 1465-1470. 1992. An Efficient Method for Segmentation of Images Based on Fractional
[10] B. Riemann, Versuch einer Allgemeinen Auffassung der Integration und Calculus and Natural Selection , Expert Systems With Applications. 39,
Differentiation,Gesammelte Mathemtische Werke. pp. 331-344. 1876. pp. 12407-12417. 2012.
[11] K. S. Miller, Derivative of Noninteger Order, Math. Mag. 68(3), pp. [16] Gao. Chaobang, Jiliu. Zhou, Ziuqing. Zheng, Fangnian. Lang, Image
183-192. 1995. Enhancement based on Improved Fractional Differentiation, , J. Comput.
[12] M. Caputo, Linear Models of Dissipation Whose Q is Almost Frequency Inf. Syst. 7(1), pp. 257-264. 2011.
Independent-II, Geophys. J. R. Astron. Soc. 13(5), pp. 529-539. 1967. [17] B. Kuldeep, A. Kumar, GK. Singh, Design of Quadrature Mirror Filter
[13] Dali. Chen, Yangquan. Chen, Dingyu. Xue, Three Fractional-Order TV- Bank using Lagrange Multiplier Method Based On Fractional Derivative
L2 Models for Image De-noising, J. Comput. Inf. Syst. 9(12), pp. 4778- Constraints , Eng Sci Technol: Int. J. 18(2), pp. 235-243. 2015.
4780. 2013. [18] Z. Z. Yang, J. L. Zhou, M. Huang and X. Y. Yan, Edge Detection Based
[14] Meriem Hacini; Fella Hachouf; Khalifa Djemal, A new fractional- on Fractional Differential , Journal of Sichuan University. 40, pp. 152-
order variational model for speckled de-noising , Visual Information 157. 2008.
Processing (EUVIP), 2014 5th European Workshop on IEEE Conference

5
(a) (b) (c) (d)
Fig. 5. Results of different fractional differential operator methods: Four test images are selected: Parrot, Fruits, Cat and Girl.(a) original, (b) G-L , (c) R-L,
(d) 2D-IFCD result.

Вам также может понравиться