1

Dictionary based Approach to Edge Detection

Nitish Chandra*, Kedar Khare
Department of Physics, Indian Institute of Technology, Delhi.

Abstract—Edge detection is a very essential part of image All the above methods use generic operators and do not
processing, as quality and accuracy of detection determines the consider the image characteristics. In this paper, we demon-
success of further processing. We have developed a new self strate an edge detection method based on image characteristics
learning technique for edge detection using dictionary comprised
of eigenfilters constructed using features of the input image. using a dictionary based algorithm. This method is different
The dictionary based method eliminates the need of pre or from traditional methods as it incorporates features of a given
post processing of the image and accounts for noise, blurriness, image, rather than using a set of pre-defined functions.
arXiv:1509.00714v1 [cs.CV] 2 Sep 2015

class of image and variation of illumination during the detection In section II, we review popular existing algorithms for edge
process itself. Since, this method depends on the characteristics detection to which the results are compared in later sections.
of the image, the new technique can detect edges more accurately
and capture greater detail than existing algorithms such as In section III, we present our algorithm of dictionary based
Sobel, Prewitt Laplacian of Gaussian, Canny method etc which edge detection. In section IV, we have compared the results
use generic filters and operators. We have demonstrated its with the popular techniques for different classes of images.
application on various classes of images such as text, face, We have also presented the application of the new method in
barcodes, traffic and cell images. An application of this technique barcode reading, counting of red blood cells and the results are
to cell counting in a microscopic image is also presented.
compared with existing algorithms. We conclude and discuss
Index Terms—Edge detection, dictionary, self-learning, eigen- further applications in section V.
filters

II. R EVIEW OF E DGE D ETECTION T ECHNIQUES

I. I NTRODUCTION
Some of the most common techniques for edge detection

I MAGE processing techniques to enhance the quality of

images or to extract information from an image have
become ubiquitous, in imaging systems such as cameras,
employ gradient operators i.e filtering the image with one or
more masks. These methods do not have any provisions for
image characteristics and noise content. Methods such as Sobel
microscopes, etc. Edge detection is an essential component and Prewitt have gradient operator in orthogonal directions.
of image processing as edges play an important role in For continuous function these are expressed as,
visual perception. Edges in images can be defined as the  ∂f ∂f 
points where sharp changes of colour and/or intensity occur. ∇f (x, y) = [Gx Gy ]T = (1)
Current edge detection techniques utilize differential operators ∂x ∂y
or combination of filtering operations or results from various In practice for digital image processing calculation of partial
techniques are combined together. There are many standard derivatives, convolution is performed using a small kernel or
types of edge detection techniques, like Sobel, Prewitt and mask, as images are comprised of discrete pixels. For Sobel
Laplacian of Gaussian that detect any discontinuities in color method following two 3 x 3 kernels are used, one detects
or intensity from one pixel to another in an image. vertical edges and other horizontal edges.
In general, there are two categories of edge detection tech-    
−1 −2 −1 −1 0 −1
niques. First, gradient methods in which maximum and min-
Gx =  0 0 0  ∗Γ Gy = −2 0 2  ∗Γ (2)
   
imum in first derivative is considered and other is Laplacian
method which searches for zero crossing of second derivative. 1 2 1 −1 0 1
In practice, edges get blurred and noisy and the level of
contamination depends on illumination, focusing mechanism, where, Γ is the image in which the edge has to detected.
and noise level in electrical components of imaging system. Prewitt edge detection uses a similar gradient method but with
Second derivative methods are more sensitive to noise than slightly different kernels. The magnitude and direction of the
first derivative. Thus, there are three fundamental steps that edge from Sobel anf Prewitt methods, can be obtained by,
are performed during edge detection, image smoothing for
q
G = G2x + G2y Θ = arctan(Gy , Gx ) (3)
noise removal, detection of edge points, and edge localization.
In Sobel and Prewitt methods these steps are performed An advanced technique used method which accounts for
separately and the results depend on the accuracy of all the noise content and edge characteristics is Laplacian of Gaus-
steps. However, in Laplacian of Gaussian smoothing step is sian (LoG), in which regions of rapid intensity changes are
incorporated and Canny method consists of all the steps. highlighted by the Laplacian of the image. In this method, the
image is first smoothed, generally by using a Gaussian kernel
*Nitish Chandra, e-mail: [email protected].
to reduce the sensitivity to noise. The Laplacian of the image
with pixel intensity I(x, y) is given by,
 2

III. T HE D ICTIONARY BASED E DGE D ETECTION

∂2I ∂2I
L(x, y) = 2
+ 2 (4) In recent years there has been a growing interest to study
∂x ∂y
sparse representation of images using an overcomplete dictio-
A smoothing filter has to be applied before the second nary D ∈ Rn×K that contains K sample columns of a basis,
derivative to reduce the effect of noise, LoG can be represented {dj }Kj=1 . The linear combination of these bases can be used
by a hybrid filter which incorporates both low pass filter to decompose a signal y ∈ Rn , or an image. The representation
and second derivative, Gaussian centered at 0 and standard can be exact y = Dx or approximate, k y − Dx kp ≤ 
deviation of σ [1], depending of amount of noise, where ’p’ can be 1,2 or ∞,
but generally p = 2 is used. The accuracy of approximation
1  x2 + y 2  − x2 +y2 2 depends upon the quality of dictionary. Methods such as basis
LoG(x, y) = − 1− e 2σ (5)
πσ 4 2σ 2 pursuit and matching pursuit are used to construct dictionaries
with reasonable approximation. Since, we are interested in the
Canny edge detector is more complex algorithm which features of the input image and one of the best way to extract
is superior to other edge detectors. The key objectives that key features is principal component analysis.
underpins Canny method are, low, error rate, localized edge
points, and single edge point response [2]. In essence, Canny
edge detector finds the optimal solution of the formulations of
the aforementioned objectives. A two dimensional Gaussian
function is used to reduce the noise, followed by gradient
operation to calculate the strength and direction of edge at
every point. Let, f (x, y) be the input image and G(x, y) =
x2 +y 2
exp− 2σ2 denote the Gaussian function, the smoothed image
is given by

fs (x, y) = f (x, y) ? G(x, y) (6)

where, ? denotes convolution. Further using equations 1 (a)

and 3, strength and direction of edges are calculated. A
gradient method generally produces wide edges, nonmaxima
suppression is used to thin out the edges.
Over the years, these techniques have been improved upon (b)
but they work better only in specific cases and still are generic
and do not use the characteristics of the image. An improved Fig. 1: Input Image: (a) Illustration of image Γ indicating
Sobel method which uses soft-threshold wavelet de-noising small segments. (not to scale). (b) Illustration of 3D dictionary
combined with traditional Sobel technique but it works most formed using small segments (si ) of the image.
efficiently on images with white Gaussian noise [3]. Another
improved edge detection method uses a two level edge fusion Consider a test image, Γ of size NxN pixels (fig.1a, test
model. On first level, typical edge detectors are used, where image has N = 512) in which color discontinuities, intensity
every pixel an edge score is calculated for every pixel [4]. Then variation are the key features for edge detection. In order to
results are combined at the second level using the Hadamard make the process more specific to an image, the filters have
product with two additional edge estimations, based on edge to be formed by analyzing the features of the image itself.
spatial characteristics. The entire image is divided into small sections or sub-images
All the above techniques use gradient operator along with (s1 , s2 , · · · sn ) of size n x n, (in this case n is 4, fig.6c) to
some smoothing filters and combination algorithms for edge form the dictionary.
detection. The major problem of these methods is that they The choice of size of the segment cannot be arbitrary. The
are generic, so edge detected depends on noise, blurriness and size of the segments is related to the kernels derived, which
illumination of the image. However, a method that and elim- are used for detecting the edges. Optimal size of the segments
inates the process of pre-processing the image and depends is 3 x 3 to 5 x 5, if the size of segments is too small, kernels
on the features of the image being analysed is more likely formed will not be sufficient to capture all the important image
to capture details in the image accurately. A self-learning characteristics. If the size is too large like 8 x 8, we observe
algorithm which to adapts the filters and operators to the that the image extracted captures very less detail because key
input image, irrespective of the class of image can provide the features of the image, is also interpreted as noise and hence
solution. In this paper, we have demonstrated an edge detection is eliminated.
method based on image characteristics, using a dictionary The dictionary is composed of s1 , s2 , · · · sn , which are
constructed from the image itself. This method is different evidently images. All the elements of the dictionary are
from traditional methods as it incorporates features of a input averaged, given by Ψs , and subtracted from each element,
image, rather than on a pre-defined operators. Φi to reduce the noise and increase the sharpness, due to low
 3

16
1X
Ψe = ei ∆i = ei − Ψe (10)
n i=1

4x4

Eigenvalue
16x1
Problem for
AAT 4x4
} Eigenfilter

Concatenated Matrix (A)

(n2 x N2/n2) 16x1
4x4
Eigenvector

(a) (b) (c)

Fig. 2: Matrix formed by elements of dictionary. Fig. 3: (a) 3d array of images (ei ) after convolution with
Γ and thresholding. Intermediate edge images obtained after
convolution with eigenfilter. (a) Edge image obtained by
pass and high pass filtering nature of averaging and subtraction eigenfilter 14. (b) Edge image obtained by eigenfilter 15.
operations, respectively.
The edge detection by individual eigenfilters is along one
n direction, as shown in fig. 3. Edges detected by each of the
1X
Ψs = si Φi = si − Ψs (7) eigenfilters for a binary object is shown in the appendix.
n i=1
To improve the quality of the edges, the edge images have
To find the key variations in each sub-image, principal to be combined. From edge image obtained, it is apparent
components have to analyzed. But, the calculation of principal that neighboring filters have similar characteristics. Thus, to
components for each sub-image involves solving of a large enhance the edges, each pair of consecutive elements in the β,
eigenvalue problem which may not be efficient to calculate. maximum for each pixel in consecutive images is calculated,
To get round this problem, we devised a method in which the maximum of ei and ei−1 , as each consecutive eigenfilter is of
n x n elements in the dictionary are transformed into column similar characteristics (such as derivative filters) but in almost
vectors of size n2 x 1 and concatenated in matrix A (fig. perpendicular direction (see fig. 12c and 12e, in appendix).
2) of size n2 x N 2 /n2 . For the calculation in this paper, N
= 512 and n = 4, hence the size of A is 16 x 16,384. The αi = max(ei , ei−1 ) (11)
technique to construct the dictionary based on PCA, which is
where, αi are stored in a 3D array β 0 . The second last
similar to the calculation of eigenfaces for recognition[8]. The
element of the new 3D array of 15 elements β 0 gives the
eigenvectors of covariance matrix AAT of size n2 x n2 (16
desired edge because the last eigenfilter just reduces the
x 16) are calculated. This reduces the number of eigenvalue
intensity of the image (as shown in appendix) and is not
problems for each sub-image to just eigenvalue problem for
useful for edge detection. The number of eigenfilter can vary
one matrix.
depending on the size of sub-images taken and eigenvectors
considered for convolution.
AAT →
−
vi = λi →
−
vi (8)
IV. R ESULTS AND A NALYSIS
ei = Γ ∗ ui (9) In this section, we have shown the application of the
dictionary method in edge detection on various classes of
Here, λi are the eigenvalues and →−vi are the eigenvectors, ∗ image and provide solution to the problems with traditional
represents convolution of original image (Γ) with ui , which methods. First, we have applied standard Sobel and Laplacian
are the rearranged eigenvectors to give ei , intermediate edge of Gaussian edge detection techniques on a Chinese character,
image stored in a 3D array (fig. 2). These n2 eigenvectors of which comprises of straight lines and curves and compared
size n2 x 1 (16 x 1) which are rearranged to form a dictionary the results with dictionary based method.
of 16 kernels (ui ) of size n x n, i.e 4 x 4 (fig.??). We have The edge found by the dictionary based method is much
termed the derived kernels as eigenfilters. They have to be more pronounced than the standard methods for simple im-
convoluted with image to detect edges along one direction, ages. To get the same performance in terms of edge continuity
like a filter. The name eigenfilter originates from the work from the traditional techniques, the edge image has to be fur-
of Vaidyanathan and Nguyen [9]. After convolution of each ther subjected to morphological image processing techniques,
eigenfilters with Γ, filtered images are stored stored in a 3D such as dilation. The difference in the edges detected can be
array (fig.3a). The average Ψe of this 3D array is subtracted observed from the following cases.
from individual elements of the array to give another 3D array For complex images such as the Lena image shown in fig.
β. 6a the output from dictionary based method is well defined,
capture more features of the image and produces a continuous
 4

(a) (b) (c)

(d) (e) (f)

Fig. 5: Edge detection in traffic image. (a) Original Image. (b) Grayscale image. (c) Edge detected by Sobel method. (d)
Laplacian of Gaussian (e) Canny method (f) Dictionary based method.

images such as text, face or cell images than the existing

techniques, specially when the image is not captured at the
highest resolution.
In practical situations, image is contaminated with noise,
varying illumination shadows, glare etc. and these affect
(a) (b) (c) (d) the quality of edges detected. Here, we have demonstrated
application of various edge detection techniques on an images
Fig. 4: Comparision with different methods of edge detection.
of roads full of cars (see fig. 5a). The image is first converted
The edges detected by dictionary based method are more
to grayscale image (fig. 5b) and the same image is used for all
detailed. (a) Original image of a Chinese character. (b) Edge
the methods without any pre-processing. Edges images from
detected by Sobel method. (c) Edge detected by LoG method.
Sobel. Laplacian of Gaussian, Canny and Dictionary based
(d) Edge detected by dictionary based method.
method has been shown in fig. 5c, 5d, 5e and 5f, respectively.
It is quite evident from the results that dictionary based method
procudes the best results with pre-processing. It also takes care
of the varying illumination, noise, glare from headlights of the
car and shadow which are shown as edges in output from LoG
and Canny as they are more sensitive than Sobel method. This
shows the advantage of having an algorithm that adapts to the
input image and uses its feature for edge detection rather than
a generic operator.
(a) (b) (c)
Accurate and fast edge detection techniques have many ap-
Fig. 6: Comparision between Sobel and Dictionary method plications in image processing. In the following sub-sections,
of edge detection. (a) Original image of Lena. (b) Edge by we will demonstrate the application and advantages of dictio-
Prewitt method. (c) Dictionary based method. nary based method in barcode reading and cell counting.

A. Barcode Reading
edge image as compared to Prewitt method. Also, for images
which are not completely focused, the new algorithm performs Barcodes have been used to store information for a long
much better. The calculation of eigenfilter by dividing into time. Primitive 1D barcodes are now being replaced by 2D
small segments increases the probability of capturing small barcodes which can contain much more data. But for reading
peaks. The subtraction of Ψe , reduces the remaining noise, these codes special hardware is required. Also, techniques to
improves edges as well as suppresses the background. The read 1D barcodes or stacked 2D barcodes such as, second
new dictionary based gives better results for every class of derivative, peak location and EM algorithm require high
 5

resolution of scanned image [10]. In recent years, new smart- cells usually requires special apparatus which are expensive.
phone based reading barcode is gaining popularity. But still, For using techniques of snapshots of a sample to count and
for camera based interpretation of barcode is still limited, segment image, edge detection technique becomes extremely
as binarizing the image to compute the bit pattern is not important.
accurate in cases where image is blurred or distorted. To read
the a 2D barcode segmentation of image, size detection and
coordinates of four corners is required [10]. With accurate
edge detection, the process of segmentation, size detection
and corner locations can be enhanced. The edge image can
also be used to calculate the rotation or tilt in the image more
precisely.
As shown in fig. 7a, we have a 2D barcode named Aztec
Code. When edges are detected using Sobel and dictionary
based methods the results are not very different for image with
highest resolution. However, when the same image is blurred,
as would happen in most practical cases, the advantage of self (a) (b)
learning algorithm becomes evident. With the new algorithm
which can detect edges from low resolution images, camera
based interpretation of barcodes can be done more accurately
using smart phones or web-cameras.

(c) (d)

(a) (b) (c) Fig. 9: Application on red blood cells. (a) Original RBC
sample. (b) Edge detected by Sobel method. (c) Edge from
Fig. 7: Edge detection in highest resolution image of Aztec
Dictionary based method. (d) Estimation of number of RBC.
Code. (a) High quality Aztec Code image. (b) Edge detected
by Sobel method. (c) Dictionary based method. With the use of the method described in this paper, it is
possible to identify edges more accurately and in much more
detail. In this paper, we have used an image of RBC cells taken
from microscope with objective of 10x magnification(fig.9a).
The final image has dark background which has to be
removed to make the edges more clearer and display in
familiar form. For removing the background, a threshold on the
intensity values of the pixels is applied to remove background.
It is fairly simple, threshold varies slightly for different images
(a) (b) (c) and can be estimated from the histogram of the image. After
Fig. 8: Edge detection in blurred image of Aztec Code, blurred applying the threshold, advantage of proposed method over
by averaging filter of size 12. (a) Blurred Aztec Code image. the existing techniques become evident.
(b) Edge detected by Sobel method. (c) Dictionary based While for Sobel method the edge detected are fainter and
method. discontinuous. The dictionary based method is able to detect
very faint boundaries and also of cells which are in the process
Another possible application can be in scanning of text of division. Sobel method is unable to detect cells that are
images, where the edges are grainy or fonts are small to joined partially or completely and in some cases it has not
be captured clearly. Edge detection technique combined with picked up the cells at all.
average image capturing hardware can be used to produce high In fig. 10, a magnified view of two sections of image and
quality image. The output edge image from Prewitt method is corresponding edges detected by Sobel and dictionary methods
not continuous as compared to output from dictionary method has been presented. In first, the cells are joined together to
in the case where boundary of letters are not continuous and form a long string and in second, there is a cluster of cells of
this could lead to errors in cases where small font size is used. different sizes and present at varying spacing. These are two
types of sections in which estimating the boundary of cells
B. Counting of Red Blood Cells and subsequently estimating the number of cells can be most
In Biomedical imaging, edge detection algorithms can be difficult. As seen from the figure, dictionary based method
used to count cells, detection of tumors etc. Counting of produces a better edge image than Sobel method. To estimate
 6

the number of cells, we have taken the advantage of circular At the beginning of the paper, we have highlighted major
shape of the red blood cells and fitted the edge image with issues that exists with the traditional methods. In general, they
circles using circular Hough transform based algorithm which reqires pre-processing of the image and it depends on the type
is also available in Matlab [11] [12]. This approach is used of noise, class of image etc. A method that takes the inputs
because of its performance in presence of noise and varying image and used its features for calculation edges eliminates
illumination. The radius of circles fitted has slight variation this step. Furthermore, dictionary based method demonstrated
to account to cells in different stages. The range of radius in this paper is not affected by noise, blurriness, class of image
of circle depends on the magnification of original image and or varying illumination, shown through various examples and
typical cells size which is obtained from edge image. Another provides a smooth solution to the problem of edge detection.
point to consider is that not all cells are circular and cells We have demonstrated its application in variety of areas
which are in the process of division look like two joined or and compared the results to that of existing algorithms. It can
half circles. Inter cellular space can also have similar circular detect accurate edges from blurry image, enabling accurate im-
features that can be picked in number estimation. From fig. age segmentation thus making barcode reading using cameras
10d, it can be seen that fitting depends on the quality of edges more reliable. It also improves the quality of text in cases
and hence the accuracy of count. where letters are grainy with small fonts. Dictionary based
method can also be used to count the number of red blood
cells in a sample without the need for specialized equipments.

R EFERENCES
[1] Rafael C. Gonzalez, Richard E. Wood, Digital Image Processing, Chapter
10, Second Edition, Prentice Hall, Upper Saddle River, New Jersey 07458.
[2] John Canny, A Computational Approach to Edge Detection, IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, Vol. PAMI-8, No.
6, November 1986.
[3] Wenshuo Gao, Lei Yang, Xiaoguang Zhang, and Huizhong Liu, An
Improved Sobel Edge Detection, IEEE, 978-1-4244-5540-9/10, 2010.
[4] Xibin Jia, Haiyong Huang, Yanfeng Sun, Jianming Yuan, and David
M. W. Powers, A novel edge detection approach using a fusion model,
Springer Multimed Tools Appl, DOI 10.1007/s11042-014-2359-6, 2014.
Fig. 10: (a) Magnified view of RBC sample; (b) Edge detected [5] I. G.T. Shrivakshan, A Comparison of various Edge Detection Techniques
by Sobel method; (c) Edge detected by Dictionary based used in Image Processing, International Journal of Computer Science, Vol.
9, Issue 5, No. 1, September 2012.
method; (d) Estimation of number of cell by approximating [6] Michal Aharon, Michael Elad, and Alfred Bruckstein, K-SVD: An Algo-
cells to a circle. Hence, better detection of edges can lead to rithm for Designing Overcomplete Dictionaries for Sparse Representa-
better approximation and more accuracy of counting. tion, IEEE Transactions On Signal Processing, Vol. 54, No. 11, November
2006.
[7] Hongying Zhua, Ikbal Sencana, Justin Wonga, Stoyan Dimitrova, Derek
Thus, a few precautions that should be taken in fitting circles Tsenga, Keita Nagashimaa, and Aydogan Ozcan, Cost-effective and Rapid
Blood Analysis on a Cell-phone, NIH Public Access, Lab Chip. 13(7):
to count number of cells. In this paper, we have used range 12821288, April 7 2013 . doi:10.1039/c3lc41408f.
of radius from 3 to 8 pixels for both methods. From fig. 10, [8] Matthew Turk, Alex Pentland, Eigenfaces for Recognition, Journal of
estimation by Sobel method is not accurate as it misses many Cognitive Neuroscience, Volume 3, Number 1, 1991.
[9] P. P. Vaidyanathan and T. Q. Nguyen, Eigenfilters: A new approach to
cells, counts inter cellular spaces and size of the estimated least-squares FIR filter design and applications including Nyquist filters,
circle is not close to the cell size. Excluding the cells that IEEE Trans. Circuits System, Vol. 34, pp. 1123, Jan. 1987.
are not completely contained within the image, total number [10] Hao Wang, and Yanming Zou, 2D Bar Codes Reading: Solutions for
Camera Phones, World Academy of Science, Engineering and Technol-
of cells are 226, and the cells number of cells detected by ogy International Journal of Computer, Control, Quantum and Information
dictionary based method is 225, with an error of 0.44%. Engineering, Vol:1, No:6, 2007.
For applications such as counting of cells, the dictionary [11] T.J Atherton, D.J. Kerbyson. Size invariant circle detection, Image and
Vision Computing, Volume 17, Number 11,pp. 795-803, 1999.
based method has advantage over other methods as it detects [12] H.K Yuen, .J. Princen, J. Illingworth, and J. Kittler, Comparative
more accurate and continuous edges and allows fitting circles study of Hough transform methods for circle finding, Image and Vision
which not only counts the RBCs, but also estimates the average Computing, Volume 8, Number 1, pp. 7177, 1990.
and variation of radii of the cells.

V. C ONCLUSION
In this paper, we have proposed a new edge detection
algorithm using dictionary comprised of the key features of the
image. Since the eigenfilters are derived from the input image,
it becomes image specific and sensitive to its features. This
enables the algorithm to adapt to different classes of images
and produce better results. Its shown in the paper, results from
the new method is much more accurate and detailed than
existing algorithms.
 7

A PPENDIX A
E IGENFILTERS AND C ORRESPONDING E DGE I MAGES

(a) Eigenfilter (b) Edge (c) Eigenfilter (d) Edge

#1 #1 #2 #2

(e) Eigenfilter (f) Edge (g) Eigenfilter (h) Edge

#3 #3 #4 #4

(i) Eigenfilter (j) Edge (k) Eigenfilter (l) Edge

#5 #5 #6 #6

(a) Eigenfilter (b) Edge (c) Eigenfilter (d) Edge

#13 #13 #14 #14
(m) Eigenfil- (n) Edge (o) Eigenfilter (p) Edge
ter #7 #7 #8 #8

(e) Eigenfilter (f) Edge (g) Eigenfilter (h) Edge

#15 #15 #16 #16
(q) Eigenfilter (r) Edge (s) Eigenfilter (t) Edge
#9 #9 #10 #10 Fig. 12: Eigenfilter displayed as an image with respective edge
images (before combining for final edge image).

(u) Eigenfilter (v) Edge (w) Eigenfil- (x) Edge

#11 #11 ter #12 #12
Fig. 11: Eigenfilter displayed as an image with respective edge
images (before combining for final edge image).