Received February 17, 2019, accepted February 25, 2019, date of publication March 13, 2019, date of current

version March 26, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2902579

An Efficient Edge Detection Approach to Provide

Better Edge Connectivity for Image Analysis
MAMTA MITTAL1 , (Member, IEEE), AMIT VERMA2 , (Life Member, IEEE), IQBALDEEP KAUR2 ,
BHAVNEET KAUR3 , MEENAKSHI SHARMA3 , LALIT MOHAN GOYAL4 ,
SUDIPTA ROY 5 , AND TAI-HOON KIM6
1 Department of Computer Science and Engineering, G.B. Pant Government Engineering College, New Delhi 110020, India
2 Department of Computer Science and Engineering, Chandigarh Group of Colleges, Mohali 140307, India
3 University
Institute of Computing, Chandigarh University, Mohali 140413, India
4 Department of Computer Engineering, J.C. Bose University of Science and Technology, YMCA, Faridabad 121006, India
5 RCIL, Washington University in Saint Louis, Saint Louis, MO 63110, USA
6 Department of Convergence Security, Sungshin Women’s University, Seoul 136-742, South Korea

Corresponding authors: Sudipta Roy (sudiptaroy01@yahoo.com) and Tai-Hoon Kim (taihoonn@daum.net)

ABSTRACT An edge detection is important for its reliability and security which delivers a better understand-
ing of object recognition in the applications of computer vision, such as pedestrian detection, face detection,
and video surveillance. This paper introduced two fundamental limitations encountered in edge detection:
edge connectivity and edge thickness, those have been used by various developments in the state-of-the-
art. An optimal selection of the threshold for effectual edge detection has constantly been a key challenge
in computer vision. Therefore, a robust edge detection algorithm using multiple threshold approaches
(B-Edge) is proposed to cover both the limitations. The majorly used canny edge operator focuses on two
thresholds selections and still witnesses a few gaps for optimal results. To handle the loopholes of the
canny edge operator, our method selects the simulated triple thresholds that target to the prime issues of the
edge detection: image contrast, effective edge pixels selection, errors handling, and similarity to the ground
truth. The qualitative and quantitative experimental evaluations demonstrate that our edge detection method
outperforms competing algorithms for mentioned issues. The proposed approach endeavors an improvement
for both grayscale and colored images.

INDEX TERMS Edge, edge connectivity, edge detection, edge width uniformity, threshold.

I. INTRODUCTION its evaluations. In 1981, Kitchen and Rosenfeld [3] made

EDGES are directly coupled with shape variations in the an attempt for detecting the edges based on its local good
distribution of pixel intensities. They carry significant sub- form by means of a quality evaluation method. The acquired
stantiation about an object [1]. Thus, edge detection persis- outcomes were similar to consequences obtained from known
tently considered as an elementary operation achieved at low- ideal positions when were compared with the variation in
level image processing and various computer vision applica- thresholds, noise and blurriness. Non-requirement of the prior
tions. The two primary approaches to edge detection from an information about true locations, made the proposed method
image are: the thresholding/enhancement technique and the applied in multiple applications. Lack in width uniformity
edge fitting technique. The former, depicts the discontinuity was the key limitation of the proposed approach [3]. The work
in image attributes those are enhanced by processing using over uniformity, accuracy and sensitivity assessment of edged
operators. The latter, involves in the fitting of / ideal edge images was made by Zhu [4]. The improvement in which was
pixels. Fig. 1 [2] illustrates the edge enhancement/ thresh- made in [5], by working over the partial derivate fails using
olding edge detection method. Numerous researchers have local information around edge terminals for effective edge
committed efforts in designing effective edge operators and linking.
In 1998, Russo [6], adopted the fuzzy reasoning for
The associate editor coordinating the review of this manuscript and
successful detection of edges without being misled by the
approving it for publication was Mohammad Shorif Uddin. noises. Fast and effective outcomes were observed from the
2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
33240 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

assimilating nucleus (SUSAN) principal and BFA. Concep-

tually, a parametric fuzzy intensifier operator (FINT) was
developed to improved weak edge information that produces
another fuzzy set. BFA was utilized to optimize the involved
parameters in fuzzy membership function and FINT. The
experiments concluded with acceptable outcomes. Günen
and Atasever [18] proposed a backtracking search clustering
based edge detection algorithm for noisy images. It was
concluded that obtained outcomes performed better than
FIGURE 1. Edge enhancement/thresholding edge detection system.
traditional methods even though statistical outcomes vary
according to the modified ground truth images. Ma et al. [19]
proposed a novel de-noising principle based edge detection
simulations. By the time, an improvement was made in [7] method on Synthetic Aperture Radar (SAR) images. Out-
and [8] by using soft-threshold wavelet technique for the comes from experiments were obtained to be successful in
removal of noise and practice of Sobel edge detection opera- developing edge visual effect of SAR images and together
tor for detection from the image. got down the false edges.
Khamy et al. [9] took a step towards the improve- Zhang et al. [20] focus to improve the Sobel Edge
ment of clear edges with an attempt in an improvement of algorithm by enhancing the gradient template using the
Sobel approach. Lack in satisfactory outcomes led a further FPGA technology. Better outcomes were observed from
enhancement in state-of-the-art using ACO approach [10]. the improved algorithm. Cao et al. [21] work over parallel
In the year 2006, Dollar et al. [11] proposed a supervised designed and implementation for an Otsu- optimized canny
learning-based algorithm for edge and boundary detection. operator using a Map Reduce parallel programming model.
In this adaptive learning stage, the algorithm used an extended Ostu method aims to optimize the canny operator’s dual
version of the Probabilistic Boosting Tree classification threshold and performance whereas MapReduce parallel pro-
method for discriminative model learning that selects and gramming operates to resolve speed and cost issues. The
amalgamate a huge count of features across various scales. outcomes acquired better performance in comparison to the
In the year 2011, Pablo et al. [12] target to minimize existing methods.
the image segmentation issues to that of contour detection. Based on the literature survey, it is observed that the
He adopted spectral clustering-based globalization contour research on edge detection have run parallel and are intermit-
detection collaborated to multiple detections of local cues. tent from the last 40 years. Incremental progress is reported
Secondly, he developed a segmentation algorithm with an aim in successive works for years. Along with this, works are
to transform contour detected outcomes into a hierarchical reported from various universities across continents. From
region tree using generic machinery. In the year 2013, Lim the literature review it has been identified that the existing
et al. [13] described a new mid-level feature called sketch methods lack in accurate detection of edges, fails in edge
tokens that is supervised information in the form of hand connectivity and acquiring of optimal qualitative and quan-
drawn contours in images. He clustered the human generated titative analysis. Therefore, in this paper, an effective edge
contours patches to generate sketch token classes and random detection algorithm is proposed which successfully comply
forest classifier for object detection from images. with the following conditions: 1) The algorithm should have
Agarwal [14] compared the edge detection of bacterial better edge connectivity. 2) The algorithm should provide
foraging algorithm (BFA) and canny method. It was con- with improved edge width uniformity. 3) The algorithm must
cluded that BFA obtained better outcomes than canny as produce acceptable entropy value. 4) The results should have
bacterial foraging uses swarm intelligence whereas, canny maximum similarity and minimum error to the respective
uses gaussian filter. Later, Yang et al. [15] aim at the issue ground truth.
of driving relationship between local phase vector and local
attenuation, if a higher dimensional signal is not intrinsically
one-dimensional signal. To overcome the discussed issue II. PROBLEM CONSTRAIN
modified differential phase congruency method was devel- Edge detection techniques are classified into two domains:
oped, which proved a higher dimensional space frequency Spatial Domain and Frequency Domain. Basically, spatial
in it is equal to the minus of the scale derivate of the local domain executes direct operations over the image pixels,
attenuation. In the year 2017, Xuan and Hong [16] focus such as operator-based methods. Operator based approaches
to resolve the limitations of a traditional canny method by are categorized into first order and second order methods.
introducing differential operation on amplitude gradient his- Technically, first order method estimates the first derivation
togram. The experiments show that the proposed method of image gradient and second order method approximate sec-
was robust to noise and could detach the targets from its ond derivation of the image gradient. Sobel, Prewitt and
background. Later, Verma and Parihar [17], developed a fuzzy Roberts edge detectors are categorized as first order methods
system for edge detection by the smallest unvalued segment and Laplacian & Canny as second order methods, respec-

VOLUME 7, 2019 33241

 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

TABLE 1. Limitations of spatial domain methods.

tively. Limitations of Spatial Domain Methods are discussed TABLE 2. Classification of physical measures of edge quality and their
quality indices.
in Table 1.
Unlike Spatial, an image can be converted into a Frequency
Domain using Fourier Transform method. Later to which
various operations were performed. Thus, fine details were
extracted from low frequencies and corresponding image
contour were obtained from high frequencies. Frequency
domain carries a limitation of huge search space.
The foremost objective was the rapid and accurate edges
detection of object from an image. In the initial phase,
an attempt for effective edge detection was made by Gonzalez
and Woods [22]. Prewitt operator was a discrete differentiable
operator that computes the gradient of an image intensity
function. It was limited to eight possible orientations, from
which the majority were not accurate. Though it was easy to
implement yet it was highly sensitive to noise.
Later, first order differential-based edge detection
approach was developed, named Sobel edge detection algo-
rithm. The gradient was a ration of a functional variation and
an image can be considered as a sampling point group of
continuous functions of a gray image. Therefore, substantial
variations in the value of a gray image can be verified
using discrete approach gradient functions. The major issue
with the Roberts operator was its sensitivity to noise while
evaluation of direction difference. Thus, Sobel proposed an
omnidirectional differential operator, having a partial deriva-
tive in m and n direction. It was a3 × 3 convolution kernelled
row edge operator [9] of point E(c, d), namely rate change. For digital images the mentioned (2) is further
Ac = {E(c+1, d − 1)+(2 × E (c+1, d))+E(c+1, d +1)
 reduced to (3).
− {E (c − 1, d − 1)+(2 × E (c − 1, d))



G(c, d) = |Ac | + |Ad | (3)


+ E(c − 1, d +1)

Ad = {E(c − 1, d +1)+(2 × E (c, d +1))+E(c+1, d +1)  Sobel was a weighting algorithm which detects the edge
− {E (c − 1, d − 1)+(2 × E (c, d − 1))

by weighting all the adjacent point pixels in each direction



+ E(c+1, d − 1)

(Left, Right, Up and Down) and identifies edge according
(1) to phenomena that it can reach to extreme value in the edge
points.
The above formula applies weightage average difference of
Thus, Sobel obtains smooth effect over noise but the mag-
image intensity in neighborhood of E(c, d). Here, the gradi-
nitude of the edges degrades drastically which leads to the
ent directs at the highest variation of the function E(c, d) and
production of false edges and extreme low pixel localization
its size can be computed from (2)
q due to the influence of local average concept. Therefore,
G(c, d) = A2c + A2d (2) Sobel was not preferred edge detection approach for high
accuracy.
Equation (2) presents that the gradient amplitude E(c, d) By 1986, a widely known Canny operator was proposed
was intensified unit distance in the direction of the highest by Canny [23]. The canny operator uses a Gaussian filter to

33242 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

FIGURE 2. Edge detection can be easily performed on normal images but there is always an issue in the processing with low contrast images.
Therefore, simulations are performed on such images with an aim to identify an intensity, for generating the final image fit for edge detection. It has
been identified that intensity below 20/8.33 causes double edges and automatic noise generation. Whereas, intensity above 20/8.33 causes highly
smooth image with no edge information.

smooth the image and for noise removal. Technically, canny use non-maxima suppression method
{8 × (φ) − 1} for enhancing the signal w. r. t. noise ratio, as the suppression
d= (4) method results in one-pixel wide ranges as the outcome.
2
The canny edge detector has few of loopholes such as: 1) It
Once the gradient magnitude has been computed non-
lacks in time consumption. 2) The use of the huge count
maximum suppression was applied, where the algorithm
of parameters in the canny edge operator algorithm leads to
eliminated the non-edge pixels.
"
# infinite tweaking for obtaining a little better result. 3) There
1 − E2 is still a need of connecting the resulting edges for extrac-
Guass Kernal = √ .exp (5) tion of complete edges those are obvious from the human
2π × φ 2 × (φ)2
perspective. 4) The Gaussian smoothing causes a lack in the
Basically, smoothing of an image was performed using localization of the detected edges (it directly depends on the
Gaussian filter and reduction in noise was made specific of size of Gaussian kernel). 5) Gaussian smoothing causes a
particular standard deviation σ . Computation of the gradi- blurriness in the corners, junctions and edges. 6) The corner
ent magnitude for partial derivatives using finite-difference pixels lead in open-ended edges and missing junctions.
approximations over which non-Maxima suppression was In later year, Rong et al. [24] proposed an improved Canny
applied. Edge Detection Algorithm (ICA). Firstly, a gravitational field
If I [u, v] was an image. Gaussian Smoothing Filter was intensity concept had been introduced in the algorithm to
presented as Gs [u, v, δ], where δ depicts the Gaussian spread replace the image gradient that was earlier used in tradi-
and manage the smoothness degree. tional canny edge operator. Secondly, developed two adaptive
Sm [u, v] = Gs [u, v, δ] · I [u, v] (6) threshold selection methods for an image with less and rich
edge information. In ICA the templates of the image gradient
Sm [u, v] was smoothed array used to compute α and β partial with a2 × 2 neighboring area operator are improved with a
derivatives A [u, v] and B [u, v] as: replacement to 3 × 3 neighboring area operator.
A[u, v] ≈ (Sm [u, v + 1] − Sm [u, v] Traditional Canny Edge  operator
 had 2 ×  2 neighboring

−1 1 1 1
+ Sm [u + 1, v + 1] − Sm [u + 1, v]) /2 area operator as: Gx =
−1 1
and Gy =
−1 −1
. And
B[u, v] ≈ (Sm [u, v] − Sm [u + 1, v] the ICA has improved to 3 × 3 neighboring area operator as:
+ Sm [u, v + 1] − Sm [u + 1, v + 1]) /2  √ √ 
2 2
− 4 0
By averaging finite differences, the computation of α and 4 
Gx =  −1 0 −1
 
β partial derivatives were made. Gradient magnitude and  √ √ 
2 2

orientation were figured as: − 0
q 4 4
Mg [u, v] = A[u, v]2 + B[u, v]2 (7) and
 √ √ 
Finally, hysteresis thresholding was applied along the 2 2
 4 1
edges. Basically, hysteresis used two thresholds: upper 4 
Gy =  √ 0 0 0 .
 
threshold and lower threshold. Here, the pixel gradient above √ 
2 2
 
the upper threshold was marked as an edge and the pixel − −1 −
gradient below the lower threshold was discarded. If the pixel 4 4
gradient lies between the two thresholds, it would be consid- The edges in image only occupy a little part, therefore an
ered as candidate edge points. The candidate edge points in inversely proportional relation had been observed between
connection with edge edges would be considered as an edge. gradient magnitude of the edge pixels ad edge pixels count.

VOLUME 7, 2019 33243

 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

FIGURE 3. Performing edge detection on low contrast image has always been a challenging chore in computer vision. Therefore,
the proposed algorithm step forward to minimize the low performance rate on such images. Firstly, to verify the suitability of the
inputted image for edge detection gray thresh value is calculated by the proposed algorithm. It has been witnessed that computed
gray thresh value sometimes falls in a situation when image turns unfit for result generation because of a lack in image contrast.
Therefore, conduct of simulative experiment is made to identify the minimum threshold value below which image turns unfit. From
the experiment it has been analyzed that threshold value below 0.2 fails in acquiring of edges from an image (As presented in
IM (1)+ and IM (2)+ ). Whereas on increasing the contrast of an image to an extent (As presented in IM (1)++ , IM (1)+++ , IM (2)++
and IM (2)+++ ), proposed algorithm results to a threshold value 0.2 and above, respectively. It has successfully led to a generation of
edges from the corresponding images (the slightly generated edge pixels are encircled). Lastly, with an aim to increase the contrast of
an image, imadjust() is called. Finally, the complete process is executed again on the adjusted image. (Detailed discussion in made in
Step-3 of Section-3).

FIGURE 4. The proposed method handled the other limitation of edge detection i.e. Edge Connectivity with an implementation of triple
intensity thresholds. Simulations are performed on an image with an aim to achieve a complete coverage to the grayscale intensity range
such that no intensity lead to generation of compromising edge detection outcomes. Multiple combinations were selected for simulations
(as presented in above images) but the combination of 64, 85 and 170 not only provided the complete coverage to the grayscale intensity
range but also lead to a generation of more connected edges with an improved noise handling in the image.

Hence, selection of threshold had a great relationship with also it enhances the ability of noise suppression and edge
the mean of gradient magnitude and standard deviation. For information. But its computing speed was relatively slow and
images with less edge information the gradient magnitude when analyzed qualitatively, a further improvement is still
of majority of the pixels was located in a small range in required.
images, which ends up to a relatively small mean of gra- Thus, to overcome the drawbacks of the above studied
dient magnitude and the standard deviation. Thus double- approaches, a new efficient edge detection algorithm is pro-
threshold selection method for such images had been made posed. The major edge quality measures which should always
under consideration. Whereas, in images with a large field be considered for effective outcomes in the state-of-the-art are
of view, the contrast of each part of the entire image was discussed in Table 2.
inconsistent, and the image gradient’s standard deviation was
large relatively. Therefore, for this kind of images, a selection III. PROPOSED ALGORITHM
to double-threshold for each pixel was made under consider- This section covers a detailed discussion of the proposed
ation. Finally, it had been identified even the ICA not only algorithm (B-Edge). The pictorial flow of the data in the
keeps the advantages of the traditional Canny algorithm, but algorithm is presented in Fig. 5. The algorithm is as follows:

33244 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

FIGURE 5. Data flow diagram of the proposed algorithm.

FIGURE 6. Process over the pixels for edge formation in all the four directions: (a) the current pixel
is considered to be an edge pixel, if the value of right pixel is larger than the selected threshold.
(b) If the value of left pixel is larger than the selected threshold, the current pixel is considered to
be an edge pixel. (c) If the value of below pixel is larger than the selected threshold, the current
pixel is considered to be an edge pixel. (d) The current pixel is considered to be an edge pixel, only
if the value of above pixel is larger than the selected threshold.

Step-1: The image Iimg , is read from the database. Here, m and n depict the pixel dimensions and ara is an
Step-2 (Graythresh and Threshold Computation Phase): array of an input image. Graythresh provides an average of
The graythresh of each pixel is computed by (8), an image, which is different for every inputted image.
Step-3 (Intensity Adjustment Phase): The corresponding
threshold value (φ) of an image is then computed by (9),
P 
n Pm
α=1 β=1 ara ψ · 20
ψ= (8) Phi = (9)
m·n 8.33
VOLUME 7, 2019 33245
 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

FIGURE 7. Illustrates (a) original Image. (b) Prewitt’s Outcomes. (c) Sobel’s Outcomes. (d) Canny’s Outcomes. (e) Improved Canny’s Outcomes.
(f) Our Outcomes. (g) Ground Truth.

With an aim to increase the contrast of an image Iimg , the com- to an implementation of blurriness with a low magnitude
puted 9 is multiplied with a constant 20/8.33. The constant which results in double edges and automatic noise generation
20/8.33 is made under consideration because as per the fur- which minimized the generation of acceptable outcomes (See
ther simulative experiment it has been witnessed that a value Fig. 2). Here, the computed value of 8 sometimes falls in
greater than 20/8.33, results to a higher 8 value. This leads to a situation when image Iimg turns unfit to obtain a result
an implementation of the additional step of blurriness with a because of the lack in image contrast. Therefore, a conduct
high magnitude which results in high smoothness in an image of simulative experiment is made to identify the minimum
that causes a reduction in edge information. Whereas, a lesser 8 value below which Iimg turns unfit. From the simulative
value than 20/8.33, results to a lower 8 value. This leads experiments it has been identified that the intensity of an

33246 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

FIGURE 8. Illustrates (a) original Image. (b) Prewitt’s Outcomes. (c) Sobel’s Outcomes. (d) Canny’s Outcomes. (e) Improved
Canny’s Outcomes. (f) Our Outcomes. (g) Ground Truth.

image below 0.2 results in a black outcome due to low Step-4 (Grayscale Conversion Phase): A grayscale image
contrast and unsatisfactory presence of the color difference in technically variates in an intensity range of 0-255. From
the image. Thus, our target is to increase the image contrast the literature review it has been identified that an optimal
(See Fig. 3). ‘‘imadjust()’’ function is called in the proposed selection of the threshold for effectual edge detection has
method with an aim to automatically increase the image con- constantly been a key challenge in computer vision. It has
trast. Thus, with an implementation of the blurriness effect in been noted that a non-consideration to complete intensity
an algorithm our method handles the first limitation of edge range has always resulted to insufficient improvement in
detection i.e. ‘‘Edge Thickness’’ satisfactorily. edge detection. Even double threshold method of Canny edge

VOLUME 7, 2019 33247

 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

FIGURE 9. The results in comparison of previously proposed methods through entropy on various human body parts.

operator detects the edge rely on gradient magnitude-based (See Fig. 4). To ensure acceptable outcomes the image Iimg
information, which requires an improvement in edge connec- is firstly converted to grayscale by (10).
tivity and image information. Therefore, a multiple threshold
(ara (α, β, :))
P
approach is used in our proposed method. ara (α, β) = (10)
Here, triple intensity threshold values set as α, β, γ have 3
been defined that aims a complete coverage to the grayscale where, α ∈ {1, 2, · · · , n} and β ∈ {1, 2, · · · , n}. Here, inten-
intensity range such that no intensity leads to a generation sities 64, 85 and 170 are made under consideration because
of compromising edge detection outcome. In this method, it is witnessed that the intensity 64 fulfills the requirement of
the threshold selection is based on automatic simulated low intensities for appropriate edge detection and its property
results conducted in color processing. From the simula- also matches to properties of intensities greater than 220.
tions 64, 85 and 170 are selected as best possible intensity Whereas, intensity 85 majorly carries a grayscale shade that
values for contributing an improvement in edge detection falls between the range 70-120 and its property also matches

33248 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

TABLE 3. Parametric evaluation of selected images through entropy difference.

to properties of intensities closer to 135. In the case of inten- identified from above operations than the scanning over the
sity 170, if the intensity value slightly changes, then it leads pixels in the horizontal direction is applied and detected
to the creation of double edges. Thus, selecting these triple edge pixels of the object will be updated in the empty
intensities delivers a complete coverage to intensity range, array O (as described in Fig. 6). Finally, the array is called
therefore leads to effective edge detection with improved and the resultant edged image is revealed. It is performed
edge connectivity. This algorithm can obtain threshold auto- as,
matically, which has higher practical value in the practical if (ara (α − k, β−k) > Selected Thresholds(8))
engineering application. Update O (α, β) by 1
Step-5 (Best Solution Selection): An appropriate range Where, α ∈ {1, 2, . . . , n},β ∈ {1, 2, . . . ., n} and k ∈
check has been examined. The individual threshold that falls {−1, 0, 1}
under ideal range of shortlisted intensities is considered as Step-6: Repeat the step 5 with respect to vertical direction
one of the three finalized thresholds for the edge detec- now.
tion than the latest obtained. Unlike to the above scenario,
if the individual threshold would not fall under ideal range, IV. SIMULATION RESULTS WITH PERFORMANCE
then the general threshold is continuing to take its respec- ANALYSIS
tive position for the threshold. Initialize a zero array O to Using the Berkeley dataset (BSDS500) the proposed edge
any user defined size [α, β]. Once the three thresholds are detection algorithm has been evaluated both qualitatively

VOLUME 7, 2019 33249

 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

TABLE 4. Parametric evaluation of selected images through MSE.

and quantitatively. BSDS500 is an extension of BSDS300, The outcomes of the proposed algorithm (B-Edge) are
encompasses original 300 images for validation/ train- displayed in Fig. 7-8 (f). Prewit’s and Sobel’s results are
ing 170 and 200 images for testing [12], [13]. This marginally similar to the ground truth as less continuity in the
benchmark operates by performing comparisons between detected edges are identified but many more enhancements
machine-generated contours and ground truth data are shown have been observed from the Canny’s results, which are
in Fig. 7-9, that allows segmentation assessment in a similar much similar to the ground truth images. On observation,
framework by concerning region boundaries as contours. over-segmentation has been identified in the results of Sobel.
Since Sobel effectively highlights noise initiated in real world
A. QUALITATIVE ANALYSIS pictures, the detected edges are analyzed to be thicker in
Qualitative assessment of obtained outcomes together with comparisons to the other (See Fig. 7-9 (c)). Unlike these,
the outcomes of four major benchmarked edge detection the outcomes from Canny’s method have more connected
algorithms [9], [22]–[24] are shown in Fig. 7-9. Fig. 7- edges. Even then, in some cases double edges are encountered
8 (a) present images selected for operations. Fig. 7-8 (b- by Canny (Fig. Fig. 7-8 (d)). A minute enhancement in the
e) correspondingly, displays the results of Prewitt [22], Canny method is observed with the implementation of ICA
Sobel [9] Canny [23] and Improved Canny (ICA) [16]. (See Fig. Fig. 7-8 (e)).

33250 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

TABLE 5. Parametric evaluation of selected images through PSNR.

Although the B-Edge did not satisfactorily detect the com- improved outcomes in comparison to the existing methods by
plete background edges in the Row-1 of the Fig. 7 as pre- efficiently handling noise, double edges, edge connectivity
sented in the corresponding ground truth but it has success- and entropy. In image Brain-3, B-Edge effectively identified
fully identified the object’s edges. Even, a discontinuity can the tumor using edges, whereas from the outcomes of Prewitt
be observed from the ICA’s outcome for the same image. and Sobel a huge edge disconnection is observed which made
Maximum identical edges to the corresponding ground truths these methods inappropriate for analysis of medical images.
have been identified by the B-Edge in Row- 6 of Fig. 7, Unlike to Prewitt and Sobel, ICA turns unable to handle the
and Row- 4 & 5 of Fig. 8. High edge connectivity has been noise which leads to high entropy values. In case of Brain-1
witnessed in Row- 2 & 5 of Fig. 7 and Row- 2, 3 & 4 of and Brain-2 double edges and low information content are
Fig. 8. In case of noise handling, acceptable results have been witnessed from Prewitt and Sobel. Where ICA obtained the
acquired and observed by the proposed method in Row- 1, 2, edges from background region of an image which leads to
4 & 6 of Fig. 7 and Row- 1, 3, 4 & 5 of Fig. 8. In contrast, high entropy value. In image ankle, knee and elbow high edge
high noise occurrence can be noticed from the outcomes of discontinuity can be observed from the outcomes of Prewitt
the existing algorithms. & Sobel and high noise presence can be witnessed from the
Despite all, the proposed algorithm is efficiently managing outcomes of ICA for the respective images.
over-segmentation issues, edge continuity, uniformity of the
detected edges and has preserved more local edge informa- B. QUANTITATIVE ANALYSIS
tion. Like the above images, the proposed algorithm B-Edge A complete quantitative assessment through Entropy [25],
is applied over the MR images of various human body parts MSE (Mean Squares Error), PSNR (Peak Signal to Noise
such as elbow, ankle, knee and brain. The results are presented Ratio) [26] & finally the SSIM (Structural Similarity
in Fig. 9. On the execution of the B-Edge on human body Index) [27] is discussed in this section. The outcomes of the
parts it is witnessed that the proposed algorithm obtains the quantitative analysis are presented in Fig. 7-9.

VOLUME 7, 2019 33251

 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

TABLE 6. Parametric evaluation of selected images through SSIM.

C. ENTROPY
The content of an information that is present in an
image is measured through entropy function introduced
by E. Shannon in 1948 [25] which is defined as
follows:
XL
H (I ) = − pi log pi (11)
i=0

In (11), I represent an image and pi is the rate of recur-

rence of pixels with intensity i. Shannon’s entropy states
that, lesser the value of H (I ) designates information with
low content, however on the other hand too high value of
H (I ) indicates the presence of noise and double edges. There-
fore, H (I ) should always lie in the middle of both extremes
for best results. The graphical analysis of obtained results
in comparison to previously proposed methods through
FIGURE 10. The graphical analysis of the results obtained from human
entropy difference on human body parts are presented body parts images in comparison of previously proposed methods
in Fig. 10. through entropy.
The successful discovery of the meaningful edges is
detected through the proposed approach. The attained entropy
difference values are less than canny but higher than Prewitt’s, extremes. Even though noise has been encountered in some
Sobel’s and relies nearby canny in some scenarios, which cases, the acquired entropy difference values of the proposed
state that the entropy difference values fall in between the two approach are lesser than those of Canny and ICA. Therefore,

33252 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

encompasses a procedure that analyzes the further pieces of

evidence from the image and connects the short contours
into long ones. The obtained results are made compared
with traditional algorithms. On observation, it is concluded
that: 1) Developed method can successfully detect robust &
thin edges with less noise proportion. 2) It provides with
better edge continuity. 3) It provides better entropy value.
There are several directions to improve B-Edge: 1) The pro-
posal algorithm does not effectively work over blur images.
2) The time computation needs to be improved. 3) More
optimal outcomes can be attained with the implementation
with the deep learning approach. Finally, we would encour-
age that the future work to practice over more challenging
databases. Better object detection and recognition can be
acquired if the algorithm proposed is used in real time. This
proposed edge detection method can be used in various device
FIGURE 11. The graphical analysis of the obtained results in comparison and application such as robots, automations, and medical
of previously proposed methods through similarity. equipment for safety, advancement and betterment of the
society.

it results in edges that restrain lesser noise proportions and REFERENCES

more information content. Collectively, it also delivers the [1] S. Loncaric, ‘‘A survey of shape analysis techniques,’’ Pattern Recognit.,
shade differences that is present in respective images. The vol. 31, no. 8, pp. 983–1001, 1998.
obtained values of entropy difference for selected images are [2] I. E. Abdou and W. K. Pratt, ‘‘Quantitative design and evaluation of
enhancement/thresholding edge detectors,’’ Process. IEEE, vol. 67, no. 5,
presented in Table III. Finally, it makes B-Edge algorithm pp. 753–763, May 1979.
a reliable method for MR images analysis of different body [3] L. Kitchen and A. Rosenfeld, ‘‘Edge evaluation using local edge coher-
parts. ence,’’ IEEE Trans. Syst., Man, Cybern., vol. 11, no. 9, pp. 597–605,
Sep. 1981.
[4] Q. Zhu, ‘‘Efficient evaluations of edge connectivity and width uniformity,’’
D. MSE (MEAN SQUARE ERROR) AND PSNR Image Vis. Comput., vol. 14, no. 1, pp. 21–34, 1996.
The mean square error of the respective outcome is computed [5] O. Ghita and P. F. Whelan, ‘‘Computational approach for edge linking,’’
J. Electron. Imaging, vol. 11, no. 4, p. 479, 2002.
from (12).
[6] F. Russo, ‘‘Edge detection in noisy images using fuzzy reasoning,’’ IEEE
1 Xm−1 Xn−1 Trans. Instrum. Meas., vol. 47, no. 5, pp. 1102–1105, Oct. 1998.
MSE = ||O (s, t) − P (s, t)||2 (12) [7] W. Gao, X. Zhang, L. Yang, and H. Liu, ‘‘An improved Sobel edge
mn 0 0
detection,’’ in Proc. 3rd Int. Conf. Comput. Sci. Inf. Technol. (ICCSIT),
Here, ‘O’ depicts the matrix data of source image, ‘P’ depicts Jul. 2010, pp. 67–71.
[8] S. Gupta and S. G. Mazumdar, ‘‘Sobel edge detection algorithm,’’ Int.
the matrix data of obtained image, the total rows and columns J. Comput. Sci. Manag. Res., vol. 2, no. 2, pp. 1578–1583, 2013.
count of the image pixels is presented as ‘m’ and ‘n’, ‘s’ [9] S. E. El-Khamy, M. Lotfy, and N. El-Yamany, ‘‘A modified fuzzy Sobel
depicts the index value of rows and ‘t’ depicts the index edge detector,’’ in Proc. IEEE 17th Nat. Radio Sci. Conf., Feb. 2000,
pp. 1–9.
value of columns. The low MSE represents satisfactory edged
[10] D.-S. Lu and C.-C. Chen, ‘‘Edge detection improvement by ant colony
results while a high MSE represents the non-acceptable out- optimization,’’ Pattern Recognit. Lett., vol. 29, no. 4, pp. 416–425, 2008.
comes [26]. [11] P. Dollar, Z. Tu, and S. Belongie, ‘‘Supervised learning of edges and
The obtained values of MSE for images are presented object boundaries,’’ in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern
Recognit., Jun. 2006, pp. 1–8.
in Table 4. It has been observed that the MSE values of [12] P. Arbeláez, M. Maire, C. Fowlkes, and J. Malik, ‘‘Contour detection and
Prewitt and Sobel are relatively higher for several images hierarchical image segmentation,’’ IEEE Trans. Pattern Anal. Mach. Intell.,
because of over segmentation as discussed earlier. Pre- vol. 33, no. 5, pp. 898–916, May 2011.
[13] J. J. Lim, C. L. Zitnick, and P. Dollár, ‘‘Sketch tokens: A learned mid-
witt and Sobel’s results overtake the results of Canny for level representation for contour and object detection,’’ in Proc. IEEE Conf.
image IM-1, IM-2, IM-26 and IM-27 in Table 4. In con- Comput. Vis. Pattern Recognit., Jun. 2013, pp. 3158–3165.
trast, B-Edge achieves the minimum MSE values. A study [14] A. Agarwal and K. Goel, ‘‘Comparative analysis of digital image for edge
reveals the lack of effectiveness in predicting the human detection by using bacterial foraging & canny edge detector,’’ in Proc. 2nd
Int. Conf. Comput. Intell. Commun. Technol., Feb. 2016, pp. 125–129.
response to image quality through PSNR method. There- [15] Y. Yang, K. I. Kou, and C. Zou, ‘‘Edge detection methods based on
fore, SSIM was introduced. PSNR has been computed modified differential phase congruency of monogenic signal,’’ Multidi-
from (13). mensional Syst. Signal Process., vol. 29, no. 1, pp. 339–359, 2016.
[16] L. Xuan and Z. Hong, ‘‘An improved CANNY edge detection algorithm,’’
V. CONCLUSION in Proc. 8th IEEE Int. Conf. Softw. Eng. Service Sci. (ICSESS), Nov. 2017,
To overwhelm the issue of traditional approaches, our pp. 275–278.
[17] O. P. Verma and A. S. Parihar, ‘‘An optimal fuzzy system for edge detection
proposed method is capable to obtain desirable results from in color images using bacterial foraging algorithm,’’ IEEE Trans. Fuzzy
both normal and medical images. The proposed system Syst., vol. 25, no. 1, pp. 114–127, Feb. 2017.

VOLUME 7, 2019 33253

 M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

[18] M. A. Günen, Ü. H. Atasever, and E. Beşdok, ‘‘A novel edge detection AMIT VERMA from Kurukshetra University,
approach based on backtracking search optimization algorithm (BSA) Kurukshetra, India, and the Ph.D. degree in image
clustering,’’ in Proc. 8th Int. Conf. Inf. Technol. (ICIT), May 2017, processing. He is currently a young and innova-
pp. 116–122. tive Academician and a Researcher. He is also
[19] X. Ma, S. Liu, S. Hu, P. Geng, M. Liu, and J. Zhao, ‘‘SAR image the Head and a Professor with the Department of
edge detection via sparse representation,’’ Soft Comput., vol. 22, no. 8, Computer Science Engineering, Chandigarh Engi-
pp. 2507–2515, 2018. neering College. He has over 14 years of experi-
[20] K. Zhang, Y. Zhang, P. Wang, Y. Tian, and J. Yang, ‘‘An improved Sobel
ence in teaching with two years of experience in
edge algorithm and FPGA implementation,’’ in Proc. 8th Int. Congr. Inf.
industry. He has got the exposure of working with
Commun. Technol., 2018, pp. 243–248.
[21] J. Cao, L. Chen, M. Wang, and Y. Tian, ‘‘Implementing a parallel a telecommunication industry at Canada. He has
image edge detection algorithm based on the Otsu-canny operator on guided around 60 students for their M.Tech. Thesis. He has been the Chair of
the Hadoop platform,’’ Comput. Intell. Neurosci., vol. 2018, May 2018, various national and international conferences. He has always been actively
Art. no. 3598284. involved in research and development. He has to his credit a number of
[22] R. Gonzalez and R. Woods, Digital image processing. 2002. research papers with a number of research publications in coveted and
[23] J. Canny, ‘‘A computational approach to edge detection,’’ IEEE referred journals, including Springer and the IEEE. He is a Reviewer of
Trans. Pattern Anal. Mach. Intell., vol. PAMI-8, no. 6, pp. 679–698, various research journals. He is a Life Member of ISTE and ACM. His
Nov. 1986. research interests include image processing, simulation and modeling, and
[24] W. Rong, Z. Li, W. Zhang, and L. Sun, ‘‘An improved CANNY edge face detection and recognition.
detection algorithm,’’ in Proc. IEEE Int. Conf. Mechatron. Automat., 2014,
vol. 2, no. 2, pp. 577–582.
[25] O. P. Verma, N. Agrawal, and S. Sharma, ‘‘An optimal edge detection using IQBALDEEP KAUR is currently an Associate Pro-
modified artificial bee colony algorithm,’’ Proc. Nat. Acad. Sci., India fessor with the Department of Computer Science
Sect. A, Phys. Sci., vol. 86, no. 2, pp. 157–168, 2016. and Engineering, Chandigarh Engineering Col-
[26] P. Ganesan and V. Rajini, ‘‘Assessment of satellite image segmentation in lege, Mohali, India. Her research interests include
RGB and HSV color space using image quality measures,’’ in Proc. Int. machine learning, pattern recognition, classifica-
Conf. Adv. Elect. Eng., Jan. 2014, pp. 1–5. tion, data mining, and knowledge discovery.
[27] Z. Wang, A. C. Bovik, H. R. Sheikh, S. Member, E. P. Simoncelli, and
S. Member, ‘‘Image quality assessment: From error visibility to structural
similarity,’’ IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612,
Apr. 2004.
[28] R. Laishram, W. K. Kumar, A. Gupta, and K. V. Prakash, ‘‘A novel
MRI brain edge detection using PSOFCM segmentation and canny algo-
rithm,’’ in Proc. Int. Conf. Electron. Syst., Signal Process. Comput. Tech- BHAVNEET KAUR received the B.Sc. degree
nol. (ICESC), Jan. 2014, pp. 398–401. (Hons.) in computer science from Delhi Univer-
[29] A. K. Ray, N. B. Bahadure, and S. Routray, ‘‘SVM-LWT enabled fuzzy sity and the M.C.A. degree from Sikkim Mani-
clustering-based image analysis for brain tumor detection,’’ Int. J. Pure pal University, in 2014. She is currently pursuing
Appl. Math., vol. 117, no. 19, pp. 13–22, 2017. the Ph.D. degree in computer applications with
[30] D. Bhawna, M. Neetu, and M. Megha, ‘‘Comparative analysis of the University Institute of Computing, Chandigarh
edge detection techniques for medical images of different body parts,’’ University. Her research interests include digital
in Proc. 4th Int. Conf. Recent Develop. Sci., Eng. Technol., 2017, image processing, computer vision, and computer
pp. 164–176. graphics.
[31] Z. Unlu, Y. S. Tokmak, and A. Sahillioglu, ‘‘Suddenly developed hem-
orrhagic olecranon bursitis releated to traumatic asymptomatic hetero-
topic ossification,’’ Phys. Med. Rehabil. Res., vol. 2, no. 3, pp. 1–2,
2017. MEENAKSHI SHARMA received the M.C.A.
degree (Hons.) and the M.Tech. and Ph.D. degrees
in computer science and engineering from Kuruk-
shetra University, in 2012. She is currently an
Associate Professor and a Research Coordina-
tor with the University Institute of Comput-
ing, Chandigarh University. Her research interests
include data compression, digital image process-
ing, and data warehousing.

MAMTA MITTAL received the degree in computer LALIT MOHAN GOYAL received the B.Tech.
science and engineering from Kurukshetra Univer- (Hons.) in computer engineering from Kurukshetra
sity, Kurukshetra, in 2001, and the masters’ degree University, Kurukshetra, the M.Tech. degree
(Hons.) in computer science and engineering from (Hons.) in information technology from Guru
YMCA, Faridabad, and the Ph.D. degree in com- Gobind Singh Indraprastha University, New Delhi,
puter science and engineering from Thapar Uni- and the Ph.D. degree in computer engineering
versity Patiala. She was teaching from last 16 years from Jamia Millia Islamia, New Delhi. He has
with emphasis on data mining, machine learning, 16 years of teaching experience in the area of
soft computing, and data structure. She is currently theory of computation, parallel and random algo-
with the G.B. Pant Government Engineering Col- rithms, distributed data mining, and cloud com-
lege, New Delhi (under the Government of NCT Delhi), and also supervises puting. He is currently with the Department of Computer Engineering, J.C.
Ph.D. candidates of Guru Gobind Singh Indraprastha University, Dwarka, Bose University of Science and Technology, YMCA, Faridabad, India. He is
New Delhi. She has published and communicated a number of research currently working on the project sponsored by the Indian Council of Medical
papers in SCI, SCIE, ESCI, and Scopus Indexed journals. She has attended Research, Delhi. He has published research papers in SCI indexed and
many workshops, FDPs, and seminars. She is an active member of CSI. Scopus indexed journals and conferences. He is a Reviewer of many reputed
journals and conferences.

33254 VOLUME 7, 2019

M. Mittal et al.: Efficient Edge Detection Approach to Provide Better Edge Connectivity for Image Analysis

SUDIPTA ROY received the Ph.D. degree in TAI-HOON KIM received the B.E. and M.E.
computer science and engineering from the degrees from Sungkyunkwan University, South
Department of Computer Science and Engineer- Korea, and the Ph.D. degrees from the Univer-
ing, University of Calcutta. He is currently with sity of Bristol, U.K., and from the University
the Radiological Chemistry and Imaging Labora- of Tasmania, Australia. He is currently with the
tory, Washington University in Saint Louis, USA. Department of Convergence Security, Sungshin
He has more than five years of experience in Women’s University, South Korea. He has pub-
teaching and research. He is an author of more lished 300 research papers in international journals
than 30 publications in refereed national and inter- and conferences. His main research interests
national journals and conferences, including the include security engineering for IT products,
IEEE, Springer, Elsevier, and many others. He is an author of one book and IT systems, development processes, and operational environments. He is an
many book chapters. He holds an U.S. patent in medical image processing Editor of the Elsevier, Springer, and many more journals.
and filed an Indian patent in smart agricultural system. His research inter-
ests include biomedical image analysis, image processing, steganography,
artificial intelligence, big data analysis, machine learning, and big data
technologies. He is an International Advisory Committee Member and a
Program Committee Member of many conferences. He has served as a
Reviewer for many international journals, including the IEEE, Springer,
Elsevier, IET, and many others, and international conferences. He is an
Associate Editor of the IEEE ACCESS, IEEE journal, and an Editorial Member
of the International Journal of Computer Vision and Image Processing and
IGI Global Journal.

VOLUME 7, 2019 33255