Sie sind auf Seite 1von 6

(IJCNS) International Journal of Computer and Network Security, 35

Vol. 2, No. 1, January 2010

Analyzing Roberts Edge Detector for Digital


Images Corrupted with Noise
Raman Maini1, Himanshu Aggarwal2
1
Punjabi University, University College of Engineering,
Patiala-147002, Punjab, India
research_raman@yahoo.com
2
Punjabi University, University College of Engineering,
Patiala-147002, Punjab, India
himagrawal@rediffmail.com

Not all edges involve a step change in intensity. Effects


Abstract: Edge detection is difficult in noisy images, since both such as refraction or poor focus can result in objects with
the noise and the edges contain high-frequency content.
boundaries defined by a gradual change in intensity. The
Attempts to reduce the noise result in blurred and distorted
edges. Operators used on noisy images are typically larger in operator[3]-[4] needs to be chosen to be responsive to such a
scope, so they can average enough data to discount localized gradual change in those cases.
noisy pixels. This results in less accurate localization of the
detected edges. So, it is crucial to have a good understanding of This paper evaluates the performance of Roberts Edge
edge detection algorithms. This paper evaluates the Detector for detection of edges in digital images corrupted
performance of Roberts Edge Detector for detection of edges in with different kinds of noise. Different kinds of noise are
digital images corrupted with different kinds of noise. Different studied in order to evaluate the performance of the Roberts
kinds of noise are studied in order to evaluate the performance Edge Detector. Further, the various standard test Images are
of the Roberts Edge Detector. Further, the various standard test examined to validate our results. The software is developed
Images are examined to validate our results. The software is using MATLAB 7.5.0. It has been observed that the Roberts
developed using MATLAB 7.5.0. It has been observed that the Edge Detector works effectively for the digital images
Roberts Edge Detector works effectively for the digital images corrupted with Poisson Noise where as its performances
corrupted with Poisson Noise where as its performances reduces
reduces sharply for other kinds of noise in digital images [5]
sharply for other kinds of noise in digital images.
- [6].
Keywords: - Digital Images, Noise, Mask, Edge Detection
2. Noise
1. Introduction Noise is considered to be any measurement that is not part
of the phenomena of interest. Images are prone to different
Edge detection refers to the process of identifying and
types of noises. Departure of ideal signal is generally
locating sharp discontinuities in an image. The
referred to as noise. Noise arises as a result of unmodelled or
discontinuities are abrupt changes in pixel intensity which
unmodellable processes going on in the production and
characterize boundaries of objects in a scene. Classical
capture of real signal. It is not part of the ideal signal and
methods of edge detection involve convolving the image
may be caused by a wide range of sources, e. g., and
with an operator (a 2-D filter), which is constructed to be
variation in the detector sensitivity, environmental
sensitive to large gradients in the image while returning
variations, the discrete nature of radiation, transmission or
values of zero in uniform regions. There are an extremely
quantization errors, etc. It is also possible to treat irrelevant
large number of edge detection operators available, each
scene details as if they are image noises, e.g., surface
designed to be sensitive to certain types of edges. Variables
reflectance textures. The characteristics of noise [7] depend
involved in the selection of an edge detection operator
on its source, as does the operator which reduces its effects.
include: Edge orientation, Noise environment, Edge
Many image processing packages contains operators to
structure. The geometry of the operator determines a
artificially add noise to an image. Deliberately corrupting an
characteristic direction in which it is most sensitive to
image with noise allows us to test the resistance of an image
edges. Operators can be optimized to look for horizontal,
processing operator to noise and assess the performance of
vertical, or diagonal edges [1].
various noise filters. Noise is generally grouped into two
Edge detection is difficult in noisy images, since both the categories-independent noise and image data dependent
noise and the edges contain high-frequency content. noise.
Attempts to reduce the noise result in blurred and distorted
edges [2]. Operators used on noisy images are typically 2.1 Image Data Independent Noise
larger in scope, so they can average enough data to discount
This type of noise can be described by an additive noise
localized noisy pixels. This results in less accurate
model, where the recorded image, i(m, n) is the sum of the
localization of the detected edges.
true image t(m, n) and the noise n(m, n):
36 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 1, January 2010

i (m, n ) = t (m, n) + n(m, n) 2.2.3 Salt & Pepper Noise


(1) This type of noise is also caused by errors in data
The noise n(m,n) is often zero-mean and described by its transmission and is a special case of data drop-out noise
variance, σ n2 . In fact, the impact of the noise on the image when in some cases single, single pixels are set alternatively
to zero or to the maximum value, giving the image a salt
is often described by the signal to noise ratio (SNR), which and pepper like appearance[10]. Unaffected pixels always
may be given by remain unchanged. The noise is usually quantified by the
σt σ i2 percentage of pixels which are corrupted.
SNR = = −1
σn σ n2
(2)
σ t2 andσ i2 are the variances of the true image and the
recorded image, respectively. In many cases, additive noise
is evenly

distributed over the frequency domain (white noise),


whereas an image contains mostly low frequency
information. Therefore, such a noise is dominant for high
frequencies and is generally referred as Gaussian noise. Figure 2. Salt and Pepper noise Distribution Function

2.2.4 Poisson Noise


2.2 Image Data Dependent Noise
The Poisson Noise generates a Poisson-distributed
pseudorandom sequence whose values are the number of
2.2.1 Detector Noise discrete events that occur in a given interval. The following
Another kind of Gaussian noise, which occurs in all equation defines the probability density function of the
recorded images to a certain extent, is detector noise. This poisson noise:
kind of noise is due to the discrete nature of radiation, i.e.,
the fact hat each imaging system is recording an image by
counting photons. Allowing some assumptions (which are
valid for many applications) this noise can be modeled with (3)
an independent, additive model, where the noise has a zero-
mean Gaussian distribution described by its standard where is the mean.
deviation (σ), or variance[8]. This means that each pixel in
the noisy image is the sum of the true pixel value and a The following equations define the mean value, µ, and the
random, Gaussian distributed noise value. standard deviation value, , of the pseudorandom sequence:

µ = E{x} =
(4)

= [E{(x – µ)2}]1/2 =
(5)

Poisson noise is the result of the Poisson process. You


can use the Poisson process to describe the probability of a
certain number of events happening in a given period of
time. For example, you can use the Poisson process to
Figure 1. Gaussian Noise Distribution Function describe the nuclear decay of atoms and the number of
messages a transmitting station receives.
This type of noise is caused by the nonlinear response of
2.2.2 Speckle Noise
the image detectors and recorders. Here the image data
Another common form of noise is data drop-out noise dependent (signal dependent) term arises because detection
commonly referred to as Speckle noise. This noise is, in and recording processes involve random electron emission
fact, caused by errors in data transmission [9]. The having a Poisson distribution with a mean response value
corrupted pixels are either set to the maximum value, which [11]. Since the mean and variance of a Poisson distribution
is something like a snow in image or have single bits flipped are equal, the signal dependent term has a standard
over. deviation if it is assumed that the noise has a unity variance.
(IJCNS) International Journal of Computer and Network Security, 37
Vol. 2, No. 1, January 2010

3. Roberts Edge Detector and its edge was detected using Roberts Edge Detector. The
Edge image is used as a reference for the purpose of
A variety of Edge Detectors are available for detecting the
comparison in subsequent studies. Then the test image was
edges in digital images. However, each detector has its own
got corrupted with four different kinds of noise, generated
advantages and disadvantages. The basic idea behind edge
using in MATLAB [16] environment with default values.
detection is to find places in an image where the intensity
Then for each of the four noisy images, the performance of
changes rapidly. Based on this idea, an edge detector may
the Roberts Edge Detector was examined practically.
either be based on the technique of locating the places where
Predefined default threshold values used by Roberts Edge
the first derivative of the intensity is greater in magnitude
Detector for different test images corrupted with different
than a specified threshold or it may be based on the criterion
kind of noise are shown in Table-1. The results of edge
to find places where the second derivative of the intensity
detection are shown in Fig. 4.
has a zero crossing[12]-[14].
It has been observed that the Roberts Edge Detector works
The Roberts Cross operator performs a simple, quick to
well both with the Gaussian as well as Poisson noise
compute, 2-D spatial gradient measurement on an image.
corrupted images. Further, it has been observed that out of
Pixel values at each point in the output represent the
these two results, the performance of the said detector is
estimated absolute magnitude of the spatial gradient of the
much superior in Poisson noise corrupted image as
input image at that point.
compared to Gaussian noise [15] corrupted image. However,
The operator consists of a pair of 2×2 convolution kernels as
its performance decrease drastically for Salt & Pepper as
shown in Figure. One kernel is simply the other rotated by
well as Speckle noise corrupted images. Roberts edge
90°. This is very similar to the Sobel operator.
detector by inherent does the averaging of neighboring
pixels. Since the Salt & Pepper noise and speckle pixel
+1 0 0 +1 values are often very different from the surrounding values,
they tend to distort the pixel average calculated by the
0 -1 -1 0 averaging of neighboring pixels significantly. Therefore the
average value calculated will be significantly different from
Gx Gy
the true value. So, performance of Roberts’s edge detector
Figure 3. Roberts edge detecting templates decreases sharply for salt & pepper and speckle type [16]-
[17] of noise.
These kernels are designed to respond maximally to edges For Poisson noise, distribution for the values of an each
running at 45° to the pixel grid, one kernel for each of the pixel is determined by the nature of light itself. Light isn’t a
two perpendicular orientations. The kernels can be applied continuous quantity, but occurs in discrete photons. These
separately to the input image, to produce separate photons don’t arrive in a steady stream, but sometime vary
measurements of the gradient component in each orientation over time. Think of it like a flow of cars on a road-
(call these Gx and Gy). These can then be combined sometimes they bunch together, sometimes they spread out,
together to find the absolute magnitude of the gradient at but in general there is an overall average flow. Therefore,
each point and the orientation of that gradient. corrupted pixel come together and can be better smoothed by
averaging. So Roberts’s edge detector which by its inherent
The gradient magnitude is given by: property does the averaging of neighboring pixels values
reduces this kind of noise accurately.
In Gaussian Noise, each pixel in the noisy image is the
sum of the true pixel value and a random, Gaussian
G = Gx 2 + Gy 2 distributed noise value. So image corrupted by this type of
(6) noise is smoothed as a whole by Roberts edge detector but
although typically, an approximate magnitude is computed loses sharp image characteristics to large extent as noise in
using: present in each pixel value but not so well as Poisson noise
corrupted image because Poisson noise distribution is for
discrete values, not continuous ones which suits the Roberts
G = Gx + Gy
edge characteristic of averaging of neighborhood pixels
(7) well. Therefore, Roberts edge detector performance is better
for images corrupted with Poisson type of noise as compared
which is much faster to compute. The angle of orientation of to Gaussian noise.
the edge giving rise to the spatial gradient (relative to the
pixel grid orientation) is given by: In order to validate our results about the performance of
Roberts Edge Detector, six different standard test images,
θ = arctan(Gy / Gx ) − 3Π / 4 each corrupted with Poisson noise are considered.
(8) The performance of Roberts Edge Detector is again
examined both for the original as well as noise corrupted
4. Results and Discussion images. Predefined default threshold values used by Roberts
Edge Detector for different test images are given in Table-2.
First, in order to evaluate the performance of the Roberts The results are shown in Fig.5 and fig. 6. From the results,
Edged Detector, a standard test image of a coin was taken it has again been observed that the performance of the
38 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 1, January 2010

Roberts Edge Detector is found to be satisfactory for all the Table 2: Predefined default threshold values used by
six test images corrupted with Poisson noise. Roberts Edge Detector for Different Images

Table1: Predefined default threshold values used by Roberts


Edge Detector for different test images corrupted with different S.No Image Threshold
Kind of noise value
S. No. Image Original Poisson 1 Original 0.0805
image noise
corrupted
image 2 Gaussian noise 0.1380
1 Coin 0.0805 0.1036 corrupted image

3 Pepper & Salt noise 0.1854


2 Cameraman 0.1441 0.1533
corrupted image
3 Circuit 0.0825 0.0926 4 Speckle noise 0.1880
corruted image
4 Cell 0.0539 0.0750
5 Poisson noise 0.1036
corrupted image
5 MRI 0.0842 0.0876

6 Tire 0.1243 0.1282

7 Tree 0.0909 0.0973

Original Image Edge Image using Roberts Mask Gaussian noise corrupted image Pepper & Salt noise corrupted image

Speckle noise corrupted image Poisson noise corrupted image Edge Image using Roberts Mask Edge Image using Roberts Mask
from Gaussian noise corrupted image from Pepper & Salt noise corrupted image

Edge Image using Roberts Mask Edge Image using Roberts Mask
from Speckle noise corrupted image from Poisson noise corrupted image

Figure 4. Performance of Roberts Edge Detector for an image corrupted with different types of noise
(IJCNS) International Journal of Computer and Network Security, 39
Vol. 2, No. 1, January 2010

Original Image Edge Image using Roberts Mask Poisson noise corrupted image Edge Image using Roberts Mask
from Poisson noise corrupted image

Original Image Edge Image using Roberts Mask Poisson noise corrupted image Edge Image using Roberts Mask
from Poisson noise corrupted image

Original Image Edge Image using Roberts Mask Poisson noise corrupted image Edge Image using Roberts Mask
from Poisson noise corrupted image

Figure 5. Performance of Roberts Edge Detector for different images corrupted with Poisson Noise

Original Image Edge Image using Roberts Mask Poisson noise corrupted image Edge Image using Roberts Mask
from Poisson noise corrupted image

Original Image Edge Image using Roberts Mask Poisson noise corrupted image Edge Image using Roberts Mask
from Poisson noise corrupted image

Original Image Edge Image using Roberts Mask Poisson noise corrupted image Edge Image using Roberts Mask
from Poisson noise corrupted image

Figure 6. Performance of Roberts Edge Detector for different images corrupted with Poisson Noise
40 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 1, January 2010

5. Conclusion Trans. Information Theory, vol. 45, pp. 1043-1051, Apr.


1999.
In this paper, an attempt is made to evaluate the
[15] P. Bao and L. Zhang, “Noise Reduction for
performance of the Roberts Edged Detector for noisy
Magnetic Resonance Image via Adaptive Multiscale
images. Experimental results have demonstrated that the
Products. Thresholding,” IEEE Trans. Medical Imaging,
Roberts Edged Detector works quite well for digital images
vol. 22, pp. 1089-1099, Sept. 2003.
corrupted with Poisson Noise whereas its performance
[16] Canny, J., “A Computational Approach to Edge
decreases sharply for other kinds of noise. Hence, this type
Detection”, IEEE Trans. Pattern Analysis and Machine
of detector can not be used in practical images which are
Intelligence, 8:679-714, November 1986
generally corrupted with many kinds of noise. However,
[17] Maar, D., Hildreth E., “Theory of edge detection”,
these can be used successfully in conjunction with suitable
Proceedings Royal Soc. London, vol. 207, 187-217,
digital filter to reduce the effect of noise substantially before
1980R.C. Gonzalez, R.E. Woods and S.L. Eddins:
applying the Roberts Edged Detector. The work is under
‘Digital Image Processing using MATLAB’, Pearson
further progress to develop two dimensional digital FIR
Education Inc., 2004.
filter based on two dimensional weighted least squares
method to reduce the effect of different kind of noise in
digital images and study its performance in conjunction with Authors Profile
Roberts Edge Detector to find edges in noisy images.
Raman Maini received
B.Tech(Computer Science &
References Engineering) from Beant College of
[1] R. Gonzalez and R. Woods Digital Image Processing, Engineering, Gurdaspur, Punjab,
Addison Wesley, 1992, pp 414 - 428. India in 1999 and M.Tech( Computer
[2] R. Boyle and R. Thomas Computer Vision: A First Science & Engineering) from PAU,
Course, Blackwell Scientific Publications, 1988, pp 48 - Ludhiana, India , in 2002. He got
50. Merit certificate in his M.Tech thesis
[3] E. Davies Machine Vision: Theory, Algorithms and at PAU He is currently working as an
Practicalities, Academic Press, 1990, Chap. 5. Assistant Professor in Computer Engineering at University
[4] D. Vernon Machine Vision, Prentice-Hall, 1991, College of Engineering, Punjabi University, Patiala, India.
Chap. 5. He is a life member of ISTE (Indian Society of Technical
[5] B. Chanda and D.D. Majumdar: ‘Digital Image Education), India, IETE (Institution of Electronics &
Processing and Anylysis’, PHI, New Delhi, December, Telecommunication Engineers), India. His current area of
2002. research is Computer Vision (Specialty Noise Reduction in
[6] S. Ando: ‘Consistent Gradient Operators’, IEEE Medical Images, Edge Detection and Image Enhancement)
Transaction on Pattern Analysis and Machine
Intelligence, vol. 22, no. 3, pp.252-265, Mar. 2000. Dr. Himanshu Aggarwal, is Reader
[7] J. Koplowitz and V. Greco: ‘On the Edge Location Error in Computer Engineering at University
for Local Maximum and Zero-Crossing Edge Detectors’, College of Engineering, Punjabi
IEEE Trans. Pattern Analysis and Machine Intelligence, University, Patiala. He had completed
vol.16, no. 12, pp. 1207-1212, Dec. 1994. his Bachelor’s degree in Computer
[8] M. Sharifi, M. Fathy and, M.T. Mahmoudi: ‘A Science from Punjabi University
Classified and Comparative Study of Edge Detection Patiala in 1993. He did his M.E. in
Algorithms’, in proc. IEEE Computer Society Computer Science in 1999 from
International Conference on Information Technology: Thapar Institute of Engineering &
Coding and Computing, 2002. Technology, Patiala. He had completed his Ph.D. in
[9] L. Gagnon and F.D. Smaili: ‘Speckle Noise Reduction Computer Engineering from Punjabi University Patiala in
of Airborne SAR Images with Symmetric Daubechies 2007.He has more than 16 years of teaching. He is an active
Wavelets’, SPIE Proc. #2759, pp. 1424, 1996 researcher who has supervised 15 M.Tech. Dissertations and
[10] Z. Wang and D. Hang, “ Progressive Switching guiding Ph.D. to 6 scholars and has contributed more than
Median Filter for the Removal of Impulse Noise from 40 articles in International and National Conferences and 17
Highly Corrupted Images,” IEEE Trans. on Circuits and papers in research Journals. His areas of interest are
Systems-II: Analog and Digital Signal processing, vol. Information Systems, ERP and Parallel Computing
46, no. 1, pp. 78-80 Jan. 1999.
[11] K. Timmermann and R. Novak: ‘Multiscale modeling
and estimation of Poisson processes with applications to
photon-limited imaging’, IEEE Trans. Information
Theory., vol. 45 , no. 3, pp. 846-852., 1999
[12] H. Chidiac and D. Ziou:, ’Classification of
ImageEdges’, Vision Interface’99, Troise-Rivieres,.
Canada,pp. 17-24, 1999.
[14] B.M. Sadler and A. Swami, “Analysis of Multiscale
Products for Step Detection and Estimation,” IEEE

Das könnte Ihnen auch gefallen