ISSN (Print) : 0974-6846

Indian Journal of Science and Technology, Vol 10(4), DOI: 10.17485/ijst/2017/v10i4/108963 January 2017 ISSN (Online) : 0974-5645

A Survey on Different Edge Detection

Techniques for Image Segmentation
K. Bala Krishnan* , Shiva Prakash Ranga and Nageswara Guptha
Department of ISE, Sri Venkateshwara College of Engineering, Bengaluru - 562157, Karnataka, India;
Bala.k.btech@gmail.com, shivaprakashranga@gmail.com, mnguptha@yahoo.com

Abstract
Objectives: To compare different edge detection techniques for the effective performance results with respect to complex
images. Methods: This paper compares and analyses the performance of diverse edge detection techniques for recent
digital image processing like canny optimal edge detection, sobel operator, prewitt operator and LoG. Findings: Canny
Optimal Edge detection (COE) technique provides better performance results under noise images or complex images when
compared to other techniques. In COE the result of PSNR is high which helps to identify the effective edges.

Keywords: Canny edge detection, Classical Operator Edge detection, Edge detection algorithm, Image segmentation.

1. Introduction edges are identified by diverse edge detection techniques

in several image processing applications such as object
In the current scenario, the digital images play vital recognition, motion analysis, pattern recognition,
role in different applications such as restorations and computer- guided surgery, finger print recognition,
enhancements, digital cinema, image transmission and automatic traffic controlling systems, anatomical
coding, color processing, remote sensing, robot vision, structure and image processing. Detecting the edges from
hybrid techniques, facsimile, pattern recognition, noisy images or corrupted images is difficult in nature. In
registration techniques, multi-dimensional image the past two decades’ several edge detection techniques or
processing, video processing, high resolution display, high algorithms are proposed, based on that the effective edges
quality color representation and super high definition are evaluated or analyzed. The ultimate reason behind in
applications. Image segmentation process is important these methods to restrict the false detection in the edges,
for these digital image processing applications because edge localization and computational time. In this, canny
the raw images are captured by the digital camera or optimal detection algorithm aims to discover the optimal
mobile camera with nonessential background scenes or edge which reduces the probability of detecting false
noise. The elimination of background images and noise edges, and gives sharp edges.
is quite important to get accurate results. This process The rest of this paper is organized as follows: Section
is a low level image engineering process which converts 2 discusses the purpose of providing image segmentation.
the raw images into segments are pixels or objects. These Section 3 focuses on diverse edge detection techniques
pixels are converted into vectors and analyzed or tested or algorithms and comparison between detection
with the any one of the image segmentation process. The methodologies. Section 4 presents the conclusion.
removal of noise from the images is performed using
de- noising techniques such as filtering, enhancement,
2. Image Segmentation
detection and localization for identifying the edges. These
edges are analyzed with the help of mid and high – level Image segmentation is the method to simplify the digital
image engineering processing methods. The adequate image into segments or pixels which are easier to analyze

* Author for correspondence

 A Survey on Different Edge Detection Techniques for Image Segmentation

and identify the effective edges in a complex image1. The • Filtering.

image engineering processes are sub divided into low- • Enhancement.
level, mid –level, high level which is shown in Figure 1. In • Detection.
low-level engineering process, the raw image is taken as • Localization.
an input and the noises are eliminated. These raw images Filtering is an initial step or technique for transforms
are transformed into pixels. The pixel is a collection of an image. It is used to highlight the features or remove
discrete cells in the particular image. The characteristics some features in the image.
of color and texture are found as similarity in a pixel. Image enhancement is the process of modify the
Generally, the raw images are taken as a colored image digital images into pixels for analysis. The following
which is needed to be modify into grayscale or black and operations are performed eliminate noise, sharpen an
white images, because the edges can be detected using the image, or brighten an image which is making it easier to
pixels. In the Mid –level, the output is presented in the identify key features of an image. It is used to measure the
form of attributes like edges, contours, and the identity gradient magnitude with respect to pixels.
of individual objects. In the High- level engineering Detection is the process of identify the adequate edges
process, involves making sense, of a recognized object in from the pixel.
the image analysis and it performs the cognitive functions Localization which is used to analyze the resolution
associated with computer vision. with respect to sub pixel.
These steps are followed one by one in order to predict
the edges effectively. Commonly segmentation process
uses the adaptive thresholding method3. But it is not
suitable for the complex or noise images.
In Feature Based Clustering, the images are
transformed into histogram and then clustering is
performed. The pixels are identified using Histogram and
detected by thresholding method. K-means algorithm is
used for clustering the texture images. The color images
Figure 1. Image engineering process.
or ordinary images of the pixels are clustered through
The De – noising techniques available in low level Fuzzy C method. The complex images are taken to the
image segmentation are2: fragmentation. The technique is successful proved for
• Region Based. identifying the color region.
• Edge Based. Model based segmentation is also known as Markov
• Threshold. Random Field (MRF). A Built in feature is available in the
• Feature Based Clustering. MRF for the color images. MRF is used to identifying the
• Model Based. edges accurately with respect to complex images.
In the Region based technique, pixels or segments are As shown in Table 1, edges are important feature in an
grouped together and the edges are identified. Each and image. Finding the effective edges with respect to noisy or
every pixel in the regions are identified and considered corrupted images are difficult due to lighting conditions,
to analyze the images. Finally, the pixels are transformed background, luminance and geometrical features, noise
in the form of vectors. It is sometime called as Similarity volume, missing to detect existing edges and false edge.
based segmentation. These techniques are classified with respect to pixels
This paper focuses on edge detection based image and sub pixels. Early edge deduction techniques like
segmentation. In this, edges are found between two classical operators (Sobel operator, Prewitt, Robert’s
equivalent regions in an image. The process or method to cross and Kirsch operators) all are used to compute the
detect the adequate edges in an image is called the edge first derivative to detect the edges. Laplacian Gaussian
detection. The identified pixels are ready to process in operator is used to compute the second order gradients of
edge detection algorithm. an image. The Gaussian or optimal edge detection focuses
The edge detection algorithm contains four steps on finding the sharp edges with fixed characteristics in all
namely: directions and sensitive to noise3.

2 Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology
 K. Bala Krishnan, Shiva Prakash Ranga and Nageswara Guptha

Table 1. Uses of edge detection algorithm The process of Sobel operator is shown in Figure 3.
S.No Method Uses The raw image is taken as an input. It detects two types of
1 Filtering Transform an image edges like Horizontal edges (Sy) and Vertical Edges (Sx).
2 Enhancement Measures the gradient magnitude The Magnitude (M) is calculated by adding the partial
with respect to pixels.
derivatives of Sx and Sy. The threshold is added with the
3 Detection Finds the exact edge point in the
image
magnitude to get the output image.
4 Localization Analyzes the resolution with
respect to sub pixel.

3. Edge Detection Techniques

The major classification of the edge detection technique
or algorithm is given Figure 2.

Figure 3. The process of sobel operator.

The Magnitude of the gradients (M) are calculated

with Sx, Sy (Sx indicates the Vertical and Sy indicates the
Horizontal position). The labeling of the neighborhood
pixels are represented in Figure 44.
The Sx, Sy partial derivations are given below5
Figure 2. Edge detection methods. M=
The angle of orientation of the edge is given by:
These techniques are analyzed and compared with ɵ= arctan(S𝑦 /S𝑥 )
respect to performances are given. Where ɵ is angle to find the direction.
Sx = (a2+Ca3+a4) – (a0+Ca7+a6)
3.1 Classical Operator Sy= (a0+Ca1+a2) – (a6+Ca5+a4)
The evaluation of first order gradient edge detection
techniques are well known as classical operator. Sobel
operator, Prewitt operator, Robert’s cross operator and
Kirsch edge detection are the classical edge detection
techniques. These techniques are used to find the effective
edges with respect to noisy images. It looks for minimum Figure 4. Labeling of neighborhood pixels.
and maximum value of first level gradients. The advantage
of classical operator is edge detection with respect to noise Where the threshold C=2, ɵ is the angle and a0, a1, a2,
image is high and it is very simple to use. The main drawback a3, a4, a5, a6, a7 are masks.
of these techniques is inaccurate with complex images. Sx, Sy are implemented by using the following
convolution mask which is shown in Figure 5 and 6. In
image processing, a kernel or convolution matrix or mask
3.1.1 Sobel Operator is a small matrix.
In this, set of 3*3 convolution kernels or mask are
used1. Maximum of edges are identified with respect to -1 0 1
perpendicular angle. In this, one kernel is allowed to -2 0 2
rotate with 90 degrees and another one kernel is stay on -1 0 1
this position.
Figure 5. Sx convolution mask.

Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology 3
 A Survey on Different Edge Detection Techniques for Image Segmentation

1 2 1 (M) of the edges.

The Sx, Sy partial derivations are given6
0 0 0
M=
-1 -2 -1 Sx = (a2+Ca3+a4) – (a0+Ca7+a6)
Sy= (a0+Ca1+a2) – (a6+Ca5+a4)
Figure 6. Sy convolution mask.
Where the threshold value C=1 and a0, a1, a2, a3, a4,
a5, a6, a7 are masks.

-1 0 1
-1 0 1
-1 0 1

Figure 9. Sx Convolution mask.

Figure 7. Conversion of original image into detected image.

1 1 1
0 0 0
An example of horizontal and vertical masks for edge -1 -1 -1
detection is shown in Figure 7 and 8. The raw image is
converted into masks and it is added with the Sx mask Figure 10. Sy convolution mask.
to find the deducted output. The output pixel for the
convolution mask [1,1] is represented. The deducted edge is performed and magnitude is
output pixel [1,1] = ([0,0] x -1) + ([0,1] x 0) + ([0,2] x calculated using the example which is shown in the Figure
1) + ([1,0] x -2) + ([1,1] x 0) + ([1,2] x 2) + ([2,0] x -1) + 11. The original image pixel 12 is taken and the magnitude
([2,1] x 0) + ([2,2] x 1). is calculated with Sx mask and the same procedures are
followed to calculate the Sy mask. The final value of Sy
mask is 20.

Figure 8. Horizontal and vertical mask.

3.1.2 Prewitt Operator

Prewitt operator is quite similar to the Sobel operator
with the difference of the C value is 1. The convolution
mask contains 1, -1, and 0 which is shown in Figure 9
and 10. The main advantage of this technique is to provide
a better performance on horizontal and vertical edges in Figure 11. Example of calculating the magnitude with
the images and higher responses for noisy images. The respect to Sx.
operator should have the following properties:
• Both negative and positive values should be in the all
3.1.3 Kirsch Edge Detection
convolution masks. The Kirsch operator or Kirsch compass kernel is a linear
• The final results should be zero when sum is obtained. edge that detects the maximum edge strength in a few
It is widely used technique to evaluate the magnitude predetermined directions. In this approach the single

4 Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology
 K. Bala Krishnan, Shiva Prakash Ranga and Nageswara Guptha

mask is used to rotate with 45 degrees in the eight

directions like North (N), North West (NW), West (W),
South West (SW), South (S), South East (SE), East (E),
and North East (NE) which is shown in the Figure 12. The
main advantage of this approach is the maximum effective
edges are found in each masks. The end mask is given

The Original Image

3.2 Laplacian of Gaussian

The laplacian Gaussians operator was introduced for
finding the second level gradients for 2- D spatial derivate
of an image. This operator normally deals with both input
and output as gray scale images. In this approach, two
East (E) edges North East (NE) edges
operations are performed namely smooth the images and
compute the laplacian.
Marr – Hildreth or LoG Edge Detection Algorithm7
• Smooth the raw image by Gaussian filter.
• Apply the laplacian(G) to the smoothed image.
• Smoothed image is done by the Equation (1)

S = g * I (1)
North (N) edges North West (NW) edges

Where S – smoothed image, g – Gaussian, I – image.

• Find the laplacianis done by the Equation (2)
∂ 2 s ∂ 2 s (2)
∇2S = +
∂x 2 ∂y 2
Where ∇ is second derivative gradient with respect
West (W) edges South West (SW) edges to x and y coordinates and is laplacian.
The laplacian of Gaussian is derived by equation (3),

2S= 2(g * I) = ( 2 g) * I (3)

Where 2(g * I) indicates the process for input image

and performs the Gaussian filter and ( 2 g) * I indicates
the process of takes the Gaussian and find the laplacian.
South (S) edges South East (SE) edges The first order and the second order derivation of
Gaussian is calculated by the following Equation (4) and
Figure 12. Kirsch edge detection masks.
(5) and the mask is shown in Figure 13.

Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology 5
 A Survey on Different Edge Detection Techniques for Image Segmentation

( x -m)2 (4) The advantage of the LoG edge detection method is

1 used to evaluate the second order derivate edges of images.
G(x) = e 2s2

2ps 2 The laplacian operation is used to found the images either

2
dark area or light area.
( x )2 +( y )
1 - (5)
G(x,y) = e 2s2
2ps 2 3.3 Gaussian
The Zero crossing operators performs the deduction of
σ = Standard deviation, when the = 1the one false zero crossing of an image and perform complex
dimensional equation G(x) is shown in the Figure 14. computation. Canny edge detection algorithm is widely
When the =2 the one dimensional equation G(x, y) used for optimal solutions. Robinson edge detection and
is shown in Figure 15. Canny edge detection both techniques are the used to
improve the optimized results on complex images.

Canny Edge Detection

This algorithm is focuses to separate the background
noise from complex image and to find the effective edges
for optimized solutions. It uses first derivative of an image.
It is used to measures the mean square distance, error edge
map and signals to noise ratio. The improved canny edge
detection algorithm provides the better optimal solutions
with respect to noisy images8.
Canny edge detection steps are given 9
Step1: Smooth the image and gradient filter.
Step 2: Take the gradient of the image.
Step 3: Evaluate the edges using the gradient in the x and
y coordinates.
Figure 13. 1D and 2D Mask.
Step4: Trace the edges with respect to x and y coordinates.
Step5: Non – maximum suppression.
X -3 -2 -1 0 1 2 3
Step6: Hysteresis.
G(x) -.01 -.13 -6 1 .6 .13 .011

Figure 14. One dimensional equation with respect to G(x). Edge Detection Performance
The evaluation of detecting edge performance
includes10
• Find the feasibility of false edges.
• Calculate the Possibility of missing edges.
• Mean and square distance of the edge estimate from
the true edge.
• Tolerance to distorted edges and other features such
as corners and junctions.
The first two criteria relate to edge detection, the
second two to edge localization, and the last to tolerance
to departures from the ideal edge model.
The performance is evaluated using the following
parameters such as Performance Ratio (PR), Miss Count
(MC), Peak Signal to Noise Ratio (PNSR), and Figure of
Figure 15. Second derivative detected output. Merit (FoM)11. Table 2 and Table 3 compare the effective
performance of diverse edge detection algorithms3.

6 Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology
 K. Bala Krishnan, Shiva Prakash Ranga and Nageswara Guptha

Table 2. Comparison of various edge detection techniques

S.no Operators Sensitivity Operation Advantage Disadvantage
1 Sobel Operator Maximum of edges are identified
with respect to perpendicular angle In accurate
2­ Prewitt Operator Medium Provide a better performance on Simplicity Average results with
horizontal and vertical edges in the Finding of smooth edges respire to complex
images images.
3 Kirsch Edge The maximum effective edges are
Detection found in the each masks.
4 LoG Good Considers the double edge images. Finding the exact edges. Few edges cannot be
Continuous edges can be detected Checks the pixels in wider detected.
using raw images area.
5 Robert’s cross Find the 2 – D spatial gradient
operator measurement on an image.
6 Robinson Edge The maximum effective edges are
Detection found in the each masks.
7 Canny Optimal High Used to eliminate the noise and to Finding the error rate is
Edge Detection find the effective edges high.
Better detection
Remove streaking problem. False Zero crossing.
Adaptive in nature.
Good localization

Table 3. Performance of various edge detection

techniques
Technique PR PSNR FoM MSE
Where M and N are rows and column of an image
Roberts 9.884 17.2154 0.083204 3934
is considered as an original raw image and is a
Prewitt 9.5487 17.2104 0.094365 3922
considered as a detected outputimage14.
Krisch 9.5627 17.2105 0.9536 3941
LoG 11.6908 17.2118 0.12455 3952
Canny 17.1933 17.2240 0.11672 3969 R is the maximal variation in the input image data.

The PR and FoM is measured by following equations12 4. Conclusion

In this paper the various edge detection techniques
for image segmentation are compared and their
performances are analyzed with respect to various factors.
From the survey, the canny optimal edge detection
algorithm produces better result when compared to other
techniques under sensitivity aspect. This paper will help
Where IA is the detected edges. the researchers in the area of digital image processing
II is the Ideal edges. applications.
d is the distance between actual and ideal edges.
∞ is the penalty factor for displaced edges.
To compare image compression quality MSE and 5. References
PSNR is used.
1. Beant K, Anil G. Mathematical morphological edge detec-
The PSNR and MSE are measured by the following
tion for remote sensing images. IEEE Transactions. 2011;
equations13 324–7.

Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology 7
 A Survey on Different Edge Detection Techniques for Image Segmentation

2. Ireyuwa EI. Comparision of edge detection techniques in based on improved wavelet transform. International Jour-
image processing techniques. International Journal of In- nal of Signal Processing, Image Processing and Pattern Rec-
formation Technology and Electrical Engineering. 2013; ognition. 2016; 9(4):435–42.
25–9. 9. Mohamed AZ, Reda RG, Sabah MA. Edge detection with a
3. Sharma A, Chaturvedi R, Chaturvedi U. Recent trends and preprocessing approach. Journal of Signal and Information
techniques in image segmentation using practical swarm – Processing. 2014; 5(4):123–34.
Optimization – A survey. International Journal of Scientific 10. More MD, Andurker GK. Edge detection technique: A
and Research Publications. 2015; 1–6. comparative approach. World Journal of Science and Tech-
4. Khaire PA, Thakur NV. A fuzzy set approach for edge de- nology. 2012.
tection. International Journal of Image Processing. 2012; 11. Shuai MC. A new prediction method for edge detection
403–12. based on human visual feature. IEEE 24th Chinese Control
5. Kumar S, Singh I. Comparison between edge detection and Detection Conference; 2012. p. 1465–8.
techniques. International Journal of Computer Applica- 12. Tobias OJ, Seara R. Image segmentation by histogram
tions. 2016; 15–8. thresholding using fuzzy sets. IEEE Transactions on Image
6. Lakshmi SJ, Prasanna. Image segmentation using various Processing. 2002; 11(12):1457–65.
edge detection techniques. International Journal of Emerg- 13. Pradyna J, Usha J. Sobel edge detection implementation us-
ing Technologies in Engineering Research. 2016; 129–32. ing FPGA. International Journal of Science and Research.
7. Malik S, Kumar T. Comparative analysis of edge detection 2015; 4(7):439–42.
between gray scale and color image. Communication on 14. Ramn M, Aggarwal H. Study and comparison of various
Applied Electronics. 2016; 5(2):38–43. image edge detection techniques. International Journal of
8. Liu X, Dawn Z, Xu WWE. Image edge deduction algorithm Image Processing. 2009; 3(1):1–12.

8 Vol 10 (4) | January 2017 | www.indjst.org Indian Journal of Science and Technology