Improved Edge Detection using Variable Thresholding Technique and Convolution of Gabor with Gaussian Filters

Signal & Image Processing: An International Journal (SIPIJ) Vol.13, No.5, October 2022
DOI: 10.5121/sipij.2022.13501 1
IMPROVED EDGE DETECTION USING
VARIABLE THRESHOLDING TECHNIQUE
AND CONVOLUTION OF GABOR
WITH GAUSSIAN FILTERS
Isaack Adidas Kamanga
Department of Electronics and Telecommunications engineering,
Dar es Salaam Institute of Technology (DIT), Dar es Salaam, Tanzania
ABSTRACT
Medical Field, Robotic vision, Pattern recognition, Hurdle detection, and smart city are examples of areas
that require image processing to achieve automation. Detecting an edge is an important stage in any
computer vision application. The performance of the edge detecting algorithm is largely affected by the
noise present in an image. An Image with a low signal-to-noise ratio (SNR), imposes a challenge to locate
its edges. To improve the observable image boundaries, an adaptive filtering technique is proposed in this
article. The proposed algorithm uses convolution of Gabor filter with Gaussian (GoG) operator to clean
the noise before non-Maxima suppression. Furthermore, using variable hysteresis thresholding can further
improve edge locating. The implementation of the algorithm was done by Python and Matlab. The obtained
results were compared to a number of reviewed algorithms such as the Canny method, Laplacian of
Gaussian, The Marr-Hildreth method, Sobel operator, and the Haar wavelet-based method. Three
performance factors were used; PNSR, MSE, and processing time. The simulation result shows that the
proposed method has higher PNSR, lower MSE, and shorter processing time when compared to the Canny
detector, the Marr-Hildreth, Haar wavelet-based, Laplacian of Gaussian, and the Sobel operator methods.
The higher PNSR, lower MSE, and shorter processing time mean improved edge details of the processed
image.
KEYWORDS
Edge, Edge locating, Filtering, Gabor filter, Python programming.
1. INTRODUCTION
Image processing is a kind of signal processing in which the input is the image and the output can
be either an image or a collection of features or characteristics of an input image. Image
processing is one of the main steps in the medical field, robotic vision, pattern recognition,
Hurdle detection, and smart city [1], [2], [3]. We may perform image processing so as to enhance
it for human interpretation, to enable information to be extracted for machine vision and
interpretation [4], [5]. Image processing is also done to restore the image details and achieve
pattern recognition [6]. Image processing can either be analog or digital. Digital image processing
involves the use of computer-written algorithms to process it to get an enhanced version of it to
extract important features for a given application. Digital image processing involves several
stages. The stages are image acquisition, enhancement, transformation, morphological
processing, segmentation, and finally representation and description [7]. The acquisition is
converting light to digital images. Enhancement means improving an image by suppressing the
noise and making the important features more observable. Transformation is the representation

2
of an image into a format such as grayscale for ease of further processing. Morphological
processing is the process of extracting image components necessary for shape representation.
Segmentation involves separating a part or parts of an image that is useful for further processing.
Segmentation is a difficult process in digital image processing. To successfully separate these
parts from the rest of the image, we need to clearly define the boundaries of an image. An image
boundary is often known as an edge [8]. An edge is a boundary separating two objects in an
image. It is obtained by determining the continuous line of pixels separating two different
brightness [9]. According to [10] and [11], An edge is a curve of pixel positions where a sudden
change of brightness occurs.
To locate an edge, a number of methods can be used. Robert operator, Sobel operator, Prewitt,
and Canny are the basic methods for edge detection. Out of these methods, Canny is optimal.
However, Canny faces four main challenges. First, the Gaussian filtering applied in the Canny
method may blur the weak edges out causing them to be undetected. Secondly, Canny uses the
same-sized sigma leading to missing weak edges. Thirdly, the Canny method involves very
complex computations, this wastes time, therefore, is not a good approach in the application that
requires real-time responses such as robotic vision systems and car plate number identification in
a smart city. The Canny approach has another problem, the pixels at corners may appear in the
wrong directions for their neighbors leading to the missing junction. This problem has been
solved by the SUSAN edge detector [12]. The remaining issues present a research opportunity to
enhance the algorithm.
In this paper, a fast algorithm based on the variable thresholding technique, the application of the
convolution of Gaussian with Gabor filters, and the use of different-sized Gabor filters before and
after non-Maxima suppression is proposed. The main target is to overcome the two challenges of
the Canny method. Firstly, is the possibility of losing the weak edges by using the same-sized
Gaussian filter. Secondary, the long computation speed. The computational speed is improved by
using variable-sized sigma and threshold depending on the brightness and noise level of an image
pixel. The implementation is done using Python.
1.1. Problem Statement
Based on the literature review, the Canny method is the most effective method for edge detection,
however, it has the following significant flaws. First, the Canny method’s Gaussian filtering may
blur the weak edges out, making them go undetected. Additionally, Canny employs the same-
sized filter, which could result in the loss of weak edges. Last but not least, the Canny method
requires time-consuming, extremely complex computations, making it a poor choice for real-time
applications like smart city car plate recognition. To overcome these challenges, this paper
proposes a fast algorithm based on the application of different-sized Gabor filters before and after
non-Maxima suppression. Also, the algorithm adjusts the threshold and sigma size in accordance
with the brightness and noise level of an image part to speed up computation. Python is used to
carry out the implementation.
This article is organized into five sections. Section 1 gives the introduction and the contributions
of the proposed method. Also, section 1 defines the keywords used in the study and the problem
statement. Section 2 is a literature review giving a survey on scholarly related works. Section 3 is
about the methodology. Section 4 gives the experimental results and discussion. Finally, section 5
provides the conclusion. The end of the paper also contains an acknowledgment of key
contributors toward successfully of this work and a list of references used.

3
1.2. The Keywords
An edge is a line joining the pixels separating two different light intensities in an image. Edge
locating of detection involves determining and drawing a line to connect all pixels that sits on the
points of a sudden change of light intensity in an image. In order to locate an edge, the first and
second derivatives of an image may be computed first. Filtering is a necessary part of edge
detection. Filtering involves the suppression of noise to clearly define the areas with different
light intensities. In this work, a Gabor filter is engaged. The Gabor filter was invented by
Hungarian engineer Denis Gabor. Gabor filter is a band pass linear filter which is orientation
sensitive. The filter examines whether the image contains any particular frequency content in
particular directions in a focused area close to the point or region of analysis. Python is a general-
purpose and interpreted high-level language for computer programming. Python is chosen
because it has simple codes and can be executed much faster than other languages’ codes. The
syntax of Python is simple meaning that it can easily be learned, read, and developed.
2. LITERATURE REVIEW
The basic edge detection methods can be based on the first derivative or second derivative of an
image function [13]. The operators are of two types, the gradient-based ones, and the Gaussian
based. Gradient-based computes the first derivatives while the Gaussian based computes the
second derivatives [14], [15]. Examples of Gradient-based operators are Sobel, Prewitt, and
Robert operators. Examples of Gaussian-based operators are the Laplacian of Gaussian and the
Canny method. Out of these methods, Canny is optimal but involves complex computation. Table
1 summarizes the advantages and limitations of each of these basic algorithms.
2.1. Related Works
The canny proposal came to solve challenges imposed by other operators in Table 1. Despite the
success of Canny methods, it has limitations hindering its application in many real-time
applications. A number of studies have been done to solve the limitations of the Canny approach.
Few works are reviewed in the following sub-sections.
2.1.1. Bilateral Detector
The bilateral edge detector has been proposed by Chandra and Abin [16]. Traditional bilateral
filtering is a nonlinear method for smoothing the image while retaining its edges. Using domain
and range kernels, the bilateral filtering operation is applied to a pixel's spatial neighbourhood. A
Gaussian function serves as the domain kernel, which smoothens the image by giving each pixel
a weight based on how far it is from the central pixel at a given location. The range kernel takes
into account the radiometric distance of the pixels from the central pixel to preserve the edge. In
order to achieve adaptive smoothing of the image, the intensity variations are taken into
consideration. This is unlike the Canny method which uses the same sized filter and sigma. The
bilateral detector is not an iterative method and hence reduces the computation time when
compared to the Canny operator.

4
Table 1. Summary of the advantages and limitations of the basic edge detection algorithms
S/N Operator Advantages Limitations
1 Robert Simple than all other operators, small
sized mask (2x2 matrix) hence has the
lowest computation time, Easy to
search smooth edges
It is very sensitive to noise, it is not
very accurate in edge detection
2 Sobel Uses a small-sized mask (3x3 matrix)
hence fast than Prewitt and Canny, it
is easy to search smooth edges
Inaccurate for weak edges, highly
sensitive to noise, may fail to
preserve diagonal direction points.
4 Laplacian
of
Gaussian
It can detect edges and their various
orientations easily, and it has fixed
characteristics in all directions
It is sensitive to noise, localization
mail fails at error curved edges, and
It generates noisy responses that do
not correspond to edges, so-called
“false edges”
S/N Operator Advantages Limitations
5 Canny Suitable for step edge, non-sensitive
to noise, and characterized by good
localization
Has complex computation, takes a
long time to complete computations,
and uses the same sized filter in
strong and weak edges
2.1.2. Haar Wavelet-based Algorithm
By computing the sums and differences of adjacent elements, the Haar wavelet analyses can data.
In this proposed method, the Canny detector is implemented with adaptive parameters using the
Haar wavelet. These variables include sigma and others. It has been noted that canny detectors
outperform conventional canny detectors when adaptive parameters are used.
2.1.3. Soft Morphology-based Detectors
The algorithm for mathematical morphology edge detection is presented by Shang and Jiang in
[17]. In order to measure and extract the shape of the corresponding image for the purposes of
image analysis and identification, Mathematical Morphology is a novel type of nonlinear
algorithm. The basic operations of erosion and dilation, as well as open and closed operations,
complete the type of interaction, making it more suitable for processing and analyzing visual
information. A modified soft morphological edge detection algorithm that can accurately detect
image edges while limiting the impact of noise and having strong robustness has been proposed
after analysis and comparison of the conventional morphological edge detection algorithm.
However, the computation speed is low and sensitive to noise hence the high probability of
missing the weak edges.
2.1.4. SUZAN Algorithm
Brady and Smith invented the SUSAN algorithm as an improvement for the Canny detector. The
algorithm shows better performance on low-contrast input images [18]. The study [19] proposes
an adjustment of this algorithm to improve its performance in high-contrast input images.
Another advantage of this detector is that its sensitivity to the noisy image is very small giving
more edge details on images with low SNR when compared to another algorithm.
Basically, SUZAN solved the inability of the Canny operator to locate junction edge pixels. With
the Canny edge detector, the pixels at corners may appear in the wrong directions for their

5
neighbors leading to the missing junction. In this paper, the digital image will be represented by a
mathematical model to ease analysis.
However, the main weakness of this method is the high latency between taking the input,
processing, and provision of results.
2.1.5. Detectors based on Variable Sigma
Many scholars have proposed an improved version of the Canny detector based on using the
variable sigma, see [20], [21], [22]. These scholars suggest that if we engage variable sigma
during image smoothing it also affects the non-Maxima suppression process. The Canny detector
works by accepting the input image I and the sigma value to the Gaussian filter. The filter filters
out the noise to reduce its effect. The ordinary Canny detector uses a constant-sized sigma
throughout the smoothing of the image. The proposed approach determines the noise of a specific
part of the image and assigns an appropriate sigma. As a result, the part of the image with more
noise is applied with a large sigma while low noise parts are applied with a small valued sigma.
2.1.6. The Marr-Hildreth Algorithm
Another algorithm was suggested by [23] before the Canny edge detector. It is based on a
gradient operator. It is superior to the Gaussian operator because it uses both Laplacian and
Gaussian filters. Gaussian filter for blurring the image while Laplacian for detecting sharp edges.
Thus the output image is the convolution of the input image and the Laplacian of the Gaussian
operator (LoG).
If the Gaussian filter is given by
)
2
(
exp
2
1
)
,
( 2
2
2
2 

j
i
j
i
G
+
−
=
Whereby i and j are the image pixel coordinates and  is a sigma value then the LoG operator
is given by the equation below.
)
2
exp(
2
)
,
(
)
,
( 2
2
2
4
2
2
2
2
2
2
2


 j
i
j
i
j
i
G
j
j
i
G
i
LoG
+
−
−
+
=


+


=
The major weakness of this method is that it is sensitive to noise, localization mail fails at error
curved edges, and it generates noisy responses that do not correspond to edges.
2.2. Research Gap
With all the algorithms proposed to solve the limitations of the Canny detector, there is no single
algorithm solving all the limitations. As discussed in section 2.1, most of the algorithms solve
one or two limitations. There is no single proposed method for solving the iteration issue and the
use of a constant filter issue. This motivated the researcher to propose an algorithm that uses
different sigma and filter sizes in smoothing the digital image. The filtering is performed before
and after non-Maxima suppression.

6
3. METHODOLOGY
3.1. Edge Detection Process
The proposed algorithm has seven main stages. These stages are image acquisition, noise
suppression stage, gradient computation, non-Maxima suppression, hysteresis thresholding,
combining individual edges to form a single image, and output stage. The flow chart in Figure 1
enlightens these stages.
Fig 1. Edge detection process flow by the proposed algorithm
3.1.1. Image Acquisition
For easy processing, the input image is converted to a non-colored image called a grayscale
image. The intensity of the image at any position ranges from 0 to 255.
Mathematical representation of an image
An image is nothing but a matrix of pixels packed in rows and columns [10-12], [13] also see
[18] and [20]. The size of rows and columns depends on the resolution of an image. The
resolution of an image is the total number of pixels represented as ny
nx, . The digital image can
be represented as a matrix of pixels’ intensity at a location )
,
( yi
xi where by nx is the width
and ny is the height of the image. Let’s assume the input to the algorithm has been already
converted to a grayscale image n
j
n
i
i,j
I ,..
3
,
2
,
1
,
..
3
,
2
,
1
, =
= whereby j
i
I ,
represents the intensity of
the digital image at a point )
,
( j
i . Practically the image contains a random noise j
i
N , which affect
the pixel at the position )
,
( j
i . Thus, the gray level of the image’s intensity can be represented by
an additive mode shown as Equation.
j
i
j
i
j
i N
I
G ,
,
, +
=
(1)

7
3.1.2. Noise Suppression
First, image G is cleaned by the Gaussian of Gabor operator. Good results have been observed by
the use of this combination. Gaussian filter is a low pass filter for removing high-frequency
components in an image [13], [16]. Gabor filter is a band pass filter allowing the defined set of
frequencies. The Gabor filter was invented by Denis Gabor in 1971 [21]. At any specific location
and given orientation, Gabor filters are very sensitive to the boundaries of two different
intensities in an image [21].
Mathematically a 1-D Gabor filter can be viewed as a product of Gaussian function and
sinusoidal function defined by Equation 2.
))
2
(
exp(
)
2
2
2
exp(
)
,
,
;
( 



 +
−
=
f
x
i
x
x
f
x
x
g (2)
1
;
sin
cos −
=
+
= i
x
i
x
ix
e
Whereby g is a Gabor function in the x-axis, x
 standard deviation (sigma) in the x-axis
direction, f and  are the frequency and phase offset of the sinusoidal wave.
An image is a 2-D function hence we rewrite equation (2) into equation (3).
))
2
(
exp(
)
2
2
2
2
2
exp(
)
,
,
,
,
;
,
( 






 +
+
−
=
i
x
i
i
y
s
x
s
y
x
g (3)
Equation 3 has both real and imaginary parts:
Real part )
2
cos(
)
2
2
2
2
2
exp(
)
,
,
,
,
;
,
( 






 +
+
−
=
i
x
i
y
s
x
s
y
x
g
Imaginary )
2
sin(
)
2
2
2
2
2
exp(
)
,
,
,
,
;
,
( 






 +
+
−
=
i
x
i
y
s
x
s
y
x
g
Whereby 
 sin
cos y
x
i
x +
= , 
 cos
sin y
x
i
y +
−
=
Fig. 2 (a) and (b) show a plot of 1-D and 2-D Gabor filters respectively.
Figure 2. MATLAB plots of (a) 1-D and (b) 2-D Gabor filters
To filter the unwanted information, the grayscale image in its mathematical model is convoluted
with the Gabor filter. The blurring of the image by the filter is specified by sigma  [6].

8
Suppose the input image to the filter is )
,
( y
x
I and the output image is )
,
( y
x
F , Figure 3 is the
suggested smoothing architecture based on Equation (3).
Figure 3. Noise smoothing architecture based on Gabor filter
3.1.3. Gradient Computation
The main concern is computing the gradient and finding the points pixel points with maximum
change in grayscale intensity [13], [17]. Sobel operator is engaged to compute the gradients of the
image pixels. The operator is applied in 2-D i.e. i and j directions. The Sobel operators in these
two directions are given by matrices i
M and j
M .












+
−
+
−
+
−
=
1
0
1
2
0
2
1
0
1
i
M












−
−
−
+
+
+
=
1
2
1
0
0
0
1
2
1
j
M
The matrices i
M and j
M are convolved with blurred image pixels )
,
( j
i
F i and j directions.
Equations (4) and (5) are the gradients at pixel point (i , j ).
)
,
(
* j
i
F
i
M
i
g = (4)
)
,
(
* j
i
F
j
M
j
g = (5)
Edge: The magnitude of the edge E is computed by applying Equation (6). look [17], [20], and
[21].
j
g
i
g
E 2
2 +
= (6)
The arc tan( i
g
j
g / ) give the direction of the computed gradient at pixel point (i , j ).
3.1.4. Non-Maxima Suppressions
A non-Maxima is a small gradient of grayscale intensity at a point compared to that in either
direction. Non-Maxima suppression means deleting this small gradient and preserve the larger
one. This process is very important in object identification [14]. This stage involves marking of

9
the local maxima as edges. The resulting gradient image contains non-maxima. According to [16]
and [20], these are deleted so as to create thin and clear edges.
To achieve the suppression, the algorithm performs two process by using a Python do-while loop
for all pixels. The proposed algorithm also adjusts the magnitude of the gradient for clear
observation of the output image.
Step1: Define the magnitude of gradient for clear observation.
Step2: Round off the gradient of pixel )
,
( j
i
I to nearest 45 degrees.
Step3: Compare the magnitude of its gradient )
,
( j
i
g to that at position )
1
,
1
( −
− j
i and that at
position )
1
,
1
( +
+ j
i .
Step4: If gradient magnitude at )
,
( j
i is larger than the left and right ones, preserve the edge at
)
,
( j
i .
Step5: Otherwise preserve the large gradient either left or right by discarding the one at )
,
( j
i .
Step6: Adjust the gradient to the defined fixed magnitude.
Step7: continue step1 to step4 for all pixels, i, j=1,2,3,4…. n
Hence when compared to the gradient in either direction, the non-Maxima gradient at a point has
a small grayscale intensity.
3.1.5. Hysteresis Thresholding
The proposed algorithm suggests using different sized threshold values at different parts of the
image depending on the noise level. An image may have banding, a fixed pattern, or random
noise [16]. This work only considers fixed pattern noise. The assumption is made such that noise
will tend to make an edge weak at the affected point, hence the algorithm lowers the threshold
value. This stage involves the deleting the edges that are not connected to the strong edges. The
results of the previous stage still contain edges created by noise. The proposed algorithm uses the
variable threshold values depending on the noise level kept in memory from stage 2 of image
blurring by the Gabor filter. Two threshold values are selected at every pixel point. The proposed
algorithm passes through the following steps to achieve hysteresis thresholding.
Step1: If the gradient at point (i , j ) > average of two threshold values, keep the gradient.
Step2: If the gradient at point (i , j ) < average of two threshold values, discard the gradient
Step3: If the gradient at point (i , j ) neither > nor < average of two threshold values, search in the
3*3 matrix region. If not found increase the region size to 5*5 matrices.
Step4: After step 3, repeat step 1 and step 2 to locate a maximum gradient.
3.1.6. Combining Individual Edges to form a Single Image
All gradients that pass a previous process are maintained in a matrix as an image with boundary
traces.
3.2. Algorithm Implementation
The pseudocode for the process summarized in Fig.1 is given below
1. Read colored image I
2. Convert I to grayscale

10
3. j=1
4 While j<512 do
5. For i=1 to 512
6. I(i,j)= 0.2989 * R + 0.5870 * G + 0.1140 * B//RGB are the red, green and blue pixel
intensities
7 i=i+1, j=j+1
8 end while
9 Smoothen the grayscale image
10 Compute gradients and suppress the weak gradients
11 Perform hysteresis thresholding
12 Get the edge image
3.2.1. Python Programming
Python is opted for because of its simple syntax, is based on object-oriented, and can speed
computation [24]. Python is easy to learn as a high-level programming language and runs faster
on hardware such as Raspberry pi. For this study, a converter tool called small Matlab and octave
to Python compiler was used to convert an m-file to corresponding Python code.
Part of Python code for image filtering
3.2.2. MATLAB2019a
Matlab is a numeric computation environment developed and supported by MathWorks.
Matlab2019 was selected due to its flexibility in code conversion and improved computer vision
toolbox.

11
3.3. Algorithm Testing
Two DIT laboratory images were used for testing the implementation. The results summarized in
Fig.4 was after several adjustments of sigma and threshold value.
Fig 4. Results of algorithm testing
4. EXPERIMENTAL RESULTS
4.1. Variables
Central spatial frequency: This is the frequency of the unidirectional sinusoidal signal. The range
is (0 - 0.5) cycle per pixels. In simulations, a maximum frequency of 0.5 cycle per pixels was
used. This help to avoid possible breakdowns of the filtered image. See [14] and [21]. Low and
high thresholds used were 5% and 30% respectively.
Other variables such as sigma, lambda, s are shown in Table 2

12
Table 2. Values of variables applied in simulation
S/N Variable Value Entry Required
value
1 Sigma  2.01 1/
2
 0.2475
2 Lambda  3.3 1/ 0.303
3 Spatial speed s 0.3 2
s 0.09
4.2. Datasets
Two standard images (Lena and sunflower) with a resolution of 512x512 were used to determine
the algorithm’s performance. PSNR, MSE, and processing time were recorded. Furthermore, to
compare the results with other algorithms, it was necessary to use standardized datasets [9], [18],
[25]. Two laboratory images taken at Dar es Salaam Institute of Technology (DIT) was used in
testing the algorithm only.
4.3. Results Discussion
The results obtained with the proposed algorithm as detailed in Table 3. Figure 4 shows filtered
images and corresponding edge detection by proposed algorithm.
Table 3. Simulation results
S/N Image
used
Size used MSE
(x10^3)
PSNR
1 Lena 512x512
standard
4.2398 12.06
2 Sun
Flower
Image
512x512
standard
3.8224 13.77
Figure 4. Original images and their corresponding filtered images and edge detection for Lena and
sunflower standard test images

13
4.4. Performance Comparison
The proposed algorithm in this work is compared against other methods reviewed earlier. Even
though not all works were compared but the best among them were selected. The proposed
algorithm was compared against Canny detector, The Marr-Hildreth algorithm, Haar wavelet-
based algorithm, Soft morphology-based detectors, Laplacian of Gaussian, and Sobel operator.
The performance comparison is done in terms of the following important terms in image
processing:
1. Mean Squared Error (MSE): Varies inversely proportional with the quality of the edge
detected [22].
2. Peak Signal to Noise Ratio (PSNR): Varies proportionally with the quality of edge
detected [22].
3. Processing time: This is a response time from the moment of accepting image to provision
of the edge image. Processing time of different algorithms is summarized by study
conducted by [25].
Table 4 shows the comparison results; the results show a significant improvement with the
proposed study.
From Table 4, the proposed method performs better in terms of PSNR and MSE and processing
time when compared to the rest. The testing was not done for other standard images. It was
conducted on 512x512-sized Lena and sunflower images. The reason is the unavailability of
PSNR and MSE values of many algorithms. Other standard images such as salt pepper, apple,
and Statue of liberty lack the values to compare with the proposed method [22], [25].
Table 4. Comparison between the proposed method and reviewed methods in terms of PSNR, MSE and
processing time values on Lena and sunflower standard 512x512 images
S/N Algorithm Testing image MSE (x10^3) PSNR(dB) Processing
time(s)
1 Proposed Lena 4.2398 12.0621 3.1312
Sunflower 3.8224 13.7701 N/A
2 Canny detector
[22],[23]
Lena 15.7000 6.2060 3.9411
Sunflower 87.54 8.7087 No data
3 The Marr-Hildreth
[22] ,[23] based on
LOG
Lena 19.53 5.2217 40.7277
4 Haar wavelet-based
[23],[24]
Lena 6.2343 11.8700 No data
Sunflower No data No data No data
6 Laplacian of
Gaussian [22], [23]
Lena 19.53 5.2217 40.7277
7 Sobel operator
[22],[23]
Lena 19.42 5.2476 29.6719
5. CONCLUSION
This article proposes an algorithm based on the Gaussian of Gabor filter (GoG) operator. The
implementation has been done in Python and some parts were tested by Matlab2019a. Two
standard images of Lena and sunflower were used to compute PSNR and MSE (summarized in

14
Table 3). The results in Table 3 were compared to the PSNR and MSE obtained by other
reviewed edge detection algorithms. Moreover, the processing time of the algorithms in these two
images was recorded and compared to that of the proposed method. The comparison of which has
been presented in Table 4.
Therefore, the proposed approach attains better edge detection with a thick edge or the
boundaries of an image compared to the reviewed edge detecting algorithm.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.
ACKNOWLEDGMENTS
The author acknowledges the valuable comments of Prof. Dr Dening Jiang of Tianjin University
of Technology and Education and Dr Mbazigwa Mkiramweni of DIT for their valuable inputs in
this work. The author is also grateful to the anonymous reviewers whose comments improved the
presentation of the article contents.
REFERENCES
[1] Suman Shrestha, (2014) “Image denoising using new adaptive Based median filter”, Signal & Image
Processing: An International Journal (SIPIJ) Vol.5, No.4, pp1-13.
[2] P. Srujana, J. Priyanka, V. Y. S. S. S. Patnaikuni and N. Vejendla, "Edge Detection with different
Parameters in Digital Image Processing using GUI," 2021 5th International Conference on
Computing Methodologies and Communication (ICCMC), 2021, pp. 795-802, doi:
10.1109/ICCMC51019.2021.9418327
[3] G. M. H. Amer and A. M. Abushaala, "Edge detection methods," 2015 2nd World Symposium on
Web Applications and Networking (WSWAN), 2015, pp. 1-7, doi: 10.1109/WSWAN.2015.7210349.
[4] S. R. Joshi and R. Koju, "Study and comparison of edge detection algorithms," 2012 Third Asian
Himalayas International Conference on Internet, 2012, pp. 1-5, doi: 10.1109/AHICI.2012.6408439
[5] Ajay Kumar Boyat & Brijendra Kumar Joshi, (2015) “A review paper: noise models in digital Image
processing”, Signal & Image Processing : An International Journal (SIPIJ) Vol.6, No.2, pp63-75.
[6] Isaack Kamanga, "An Adaptive Approach to Improve Canny Method for Edge Detection",
International Journal of Science and Research (IJSR), Volume 6 Issue 6, June 2017, pp. 164-168
[7] Chandra Sekhar Seelamantula and Abin Jose and,“Bilateral Edge Detectors",ICASSP,IEEE,pp:1449-
1453,2013.
[8] Meiling Huang, Yuelei Liu, Yongmei Yang, Edge detection of ore and rock on the surface of
explosion pile based on improved Canny operator, Alexandria Engineering Journal, Volume 61, Issue
12, 2022, Pages 10769-10777, ISSN 1110-0168, https://doi.org/10.1016/j.aej.2022.04.019
[9] Isaack Adidas Kamanga, Dening Jiang, "Securing Iris Recognition System for Access Control Based
on Image Processing", International Journal of Science and Research (IJSR), Volume 6 Issue 10,
October 2017, pp. 1131-1140,
[10] J. F. Canny, A Computational Approach to Edge Detection, IEEE Trans. on PAMI, vol 8, No.6, pp.
679-698 (1986).
[11] W. Rong, Z. Li, W. Zhang and L. Sun, "An improved Canny edge detection algorithm," 2014 IEEE
International Conference on Mechatronics and Automation, 2014, pp. 577-582, doi:
10.1109/ICMA.2014.6885761.
[12] Ewaryst, R. SUSAN edge detector reinterpreted, simplified and modified, Proceedings of the 2007
International Workshop on Multidimensional (nD) Systems. nDS 2007 Aveiro, Portugal, June 27-29,
2007. [Piscataway, NJ] : IEEE, cop. 2007, 1-4244-1112-2/07
[13] Shashidhar Ram Joshi,Roshan koju," Study and Comparison of Edge Detection Algorithms",
IEEE,2012.
[14] Shang Junna, Jiang Feng," An Algorithm of Edge Detection Based on Soft
Morphology",IEEE,2012,pp: 166-169.
[15] Shokhan M. H.(2014). An efficient approach for improving canny edge detection algorithm.
International Journal of Advances in Engineering & Technology. Issue 1 volume 7.

15
[16] Chao Gao, Hongjun Zhu, Yongcai Guo, Analysis and improvement of SUSAN algorithm, Signal
Processing, Volume 92, Issue 10,2012,Pages 2552-2559, ISSN 0165-1684.
[17] M. Sharifi, M. Fathy and M. T. Mahmoudi, "A classified and comparative study of edge detection
algorithms," Proceedings. International Conference on Information Technology: Coding and
Computing, 2002, pp. 117-120, doi: 10.1109/ITCC.2002.1000371.
[18] Lejiang Guo,Yahui Hu Ze Hu, Xuanlai Tang “The Edge Detection Operators and Their Application
in License Plate Recognition, IEEE TRANSATION 2010, 20978-1-4244-5392-4/10/.
[19] Jianjum Zhao, Heng Yu, Xiaoguang Gu and Sheng Wang. “The Edge Detection of River model
Based on Self-adaptive Canny Algorithm And Connected Domain Segmentation” IEEE,Proceedings
of the 8th World Congress on Intelligent Control and Automation July 6-9 2010, Jinan, China, 978-1-
4244-6712-9/10/$26.00 ©2010 IEEE.
[20] Mee-Li Chiang, Siong-Hoe Lau “Automatic Multiple Faces Tracking and Detection using Improved
Edge Detector Algorithm” IEEE 7th International Conference on IT in Asia (CITA),2011 ,978-1-
61284-130-4/11/
[21] K. R. Namuduri, R. Mehrotra and N. Ranganathan, "Edge detection models based on Gabor filters,"
Proceedings., 11th IAPR International Conference on Pattern Recognition. Vol. III. Conference C:
Image, Speech and Signal Analysis, 1992, pp. 729-732, doi: 10.1109/ICPR.1992.202090.
[22] Dakshayani, V.; Locharla, G.R.; Pławiak, P.; Datti, V.; Karri, C. Design of a Gabor Filter-Based
Image Denoising Hardware Model. Electronics 2022, 11, 1063. https://
doi.org/10.3390/electronics11071063
[23] D. Poobathy, R. Manicka Chezian,"Edge Detection Operators: Peak Signal to Noise Ratio Based
Comparison", IJIGSP, vol.6, no.10, pp.55-61, 2014.DOI: 10.5815/ijigsp.2014.10.07
[24] Maini R, Aggarwal H. Study and comparison of various image edge detection techniques. IJIP. 2009
Mar; 3(1):1–12.
[25] https://www.python.org/doc/essays/blurb/ Visted on 12th June, 2022.
[26] Raver, C. (2020). LENA Data. Databrary. Retrieved August 24, 2022 from
http://doi.org/10.17910/b7.1102
AUTHORS
Isaack Adidas Kamanga is a research assistance at Dar es Salaam Institute of
Technology with a research base in computer vision, Machine learning, Artificial
intelligence and Signal and information processing. Specifically, in application of image
processing in medical diagnosis automation, Fault identification in systems, pests and
diseases identification in crops, features extraction from image. Application of sound
spectrum analysis in recognition. FPGA and VLSI technologies and Nano-electronics.
He completed MEng. at Tianjin University of Technology and Education in 20018, Tianjin China P.R.
Completed B.S.in Telecommunications Engineering from University of Dar Es Salaam in 2012. He is a
registered professional engineer by Tanzania Engineers Registration Board (ERB) Tanzania

Improved Edge Detection using Variable Thresholding Technique and Convolution of Gabor with Gaussian Filters

More Related Content

Similar to Improved Edge Detection using Variable Thresholding Technique and Convolution of Gabor with Gaussian Filters (20)

Recently uploaded (20)

Improved Edge Detection using Variable Thresholding Technique and Convolution of Gabor with Gaussian Filters