See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/252015677

Fast vectorization and upscaling images with natural objects using canny edge
detection

Article · April 2011

DOI: 10.1109/ICECTECH.2011.5941823

CITATIONS READS

5 335

1 author:

Aniket Bera
Purdue University
110 PUBLICATIONS 2,458 CITATIONS

SEE PROFILE

All content following this page was uploaded by Aniket Bera on 21 November 2020.

The user has requested enhancement of the downloaded file.

 Fast Vectorization and Upscaling Images with Natural Objects
using Canny Edge Detection

Aniket Bera
Jaypee Institute of Information Technology, Noida
[email protected]

Abstract— In this paper a new, faster approach which is is limited only by the resolution of the display, and the file
different from all the other conventional image vectorization size of vector data generating the image stays the same.
techniques. Using canny edge detection we are able to find the Printing the image to paper will usually give a sharper,
sharp edges in the image and the assigning shades to each higher resolution output than printing it to the screen but can
identifiable segment using random colour extraction from the use exactly the same vector data file.
original image. Finally mapping the colour blobs with the
SVG Schema and generating a scalable vector image. This III. IMAGE PRE-PROCESSING
technique is efficient for natural and well as non-It can be
directly used in security cameras for live image enhancement. 1. Noise Reduction
If several images of same object differing only in noise
Keywords- Vectorization, Enhancement, Canny Edge realization are available then temporal averaging reduces the
Detection, Image Upscaling noise [1].
When temporal averaging is not possible, methods
developed from linear filtering or smoothing is sometimes
I. INTRODUCTION
used. Early techniques involved general statistical ideas or
One of the most prominent application fields is in frequency domain concept [1],[2].
security and object detection and analysis. This area is Chin-Chen Chang [3] describes methods for a Fast Noise
characterized by the extraction of information from image Reduction Method Based on Human Visual System.
data for the purpose of analyzing the object and its features. The motivation to propose this noise detection scheme is to
Many times the object is a very small part of the image detect impulse noise faster, and, since the final judge of
and it needs to be upscaled/vectorized to be properly studied. the image quality is the human eye, they think that some
Most important application of upscaling is in
kinds of impulse noise that can be accepted.
preservation of old images and in security.
Most of the security cameras have a very low resolution.
Using good interpolation techniques we can extract
important information for the data even which at a low
resolution.
II. ABOUT VECTOR IMAGES
Computer displays are made up from grids of small
rectangular cells called pixels. The picture is built up from
these cells. The smaller and closer the cells are together, the
better the quality of the image, but the bigger the file needed
to store the data.
If the number of pixels is kept constant, the size of each
pixel will grow and the image becomes grainy (pixelated)
when magnified, as the resolution of the eye enables it to
pick out individual pixels.
Vector graphics files store the lines, shapes and colors
that make up an image as mathematical formulae.
A vector graphics program uses these mathematical
formulae to construct the screen image, building the best
quality image possible, given the screen resolution. The
mathematical formulae determine where the dots that make
up the image should be placed for the best results when
displaying the image. Since these formulae can produce an
image scalable to any size and detail, the quality of the image Fig. 1 Noise Cleaning Method

___________________________________
978-1-4244 -8679-3/11/$26.00 ©2011 IEEE

164

Authorized licensed use limited to: University of Maryland College Park. Downloaded on November 21,2020 at 17:57:32 UTC from IEEE Xplore. Restrictions apply.
 2. Grey Scale Conversion The Canny algorithm basically finds edges where the
grayscale intensity of the image changes the most.
Take an RGB color image as input
These areas are found by determining gradients of the
and produce a grayscale image [4] as image. Gradients at each pixel in the smoothed image are
output . determined by applying what is known as the Sobel-operator.
Avoiding gamma correction issues, we assume linear
color and grayscale values. The operator consists of a pair of 3×3 convolution
The process is controlled by three parameters: the degree kernels as shown. One kernel is simply the other rotated by
of image enhancement, Ȝ ; the typical size of relevant image 90°.
features in pixels, ı ; and the proportion of image pixels
assumed to be outliers, Ș.

Table 1. Table for Values of Gx and Gy

These kernels are designed to respond maximally to

(a) (b)
edges running vertically and horizontally relative to the pixel
Fig. 2. (a) Original Image, (b) Enhanced Grayscale Image
grid, one kernel for each of the two perpendicular
orientations. The kernels can be applied separately to the
input image, to produce separate measurements of the
IV. CANNY EDGE DETECTION gradient component in each orientation (call
these Gx and Gy).
These can then be combined together to find the absolute
1. Gaussian Filter magnitude of the gradient at each point and the orientation of
. Every Image is first smoothened by applying a Gaussian that gradient. The gradient magnitude is given by:
Filter. The kernel of a Gaussian filter with a standard
deviation of = 1.4
Typically, an approximate magnitude is computed using:
which is much faster to compute.

The angle of orientation of the edge (relative to the pixel

grid) giving rise to the spatial gradient is given by:

3. Non-Maximum Suppression

For Non-Maximum Suppression we have used the

algorithm proposed by Alexander Neubeck and Luc Van
Goo [5].
The results were quite good compared to other
algorithms.

(a) (b)

Fig. 3. (a) Grayscale Image (b) After Gaussian Filter

2. Gradient Detection - Sobel

165

Authorized licensed use limited to: University of Maryland College Park. Downloaded on November 21,2020 at 17:57:32 UTC from IEEE Xplore. Restrictions apply.
 V. COLOUR IDENTIFICATION
Modularized random distributed colour capturing using
fractal decomposition [7-11] in the original image was done.
The images changed the L, M and S cone excitation at
each spatial location, and the contrasts along three post-
receptoral axes [L + M, L í M or S í (L + M)].

Individual scenes varied substantially in their mean

chromaticity and luminance, in the principal color-luminance
axes of their distributions, and in the range of contrasts in
Fig. 4. Non-Maximum Suppression Output their distributions.
Chromatic contrasts were biased along a relatively narrow
The purpose of this step is to convert the smoothened range of bluish to yellowish-green angles, lying roughly
edges in the image of the gradient magnitudes to edges with between the S - (L + M) axis and a unique blue-yellow axis
sharp features.
Basically this is done by preserving all local maxima in
the gradient image, and deleting everything else.
4. Double thresholding
Even after non-maximum suppression, the image consists
of rough lines many of which are not actually edges.
Therefore we need to segregate the lines which are strong
(more possibility of a true edge) and weak (incorrectly
detected as edge).
These are generally detected due to rough and uneven
surfaces. The simplest way to discern between these would
be to use a threshold, so that only edges stronger that a Fig. 6. Colour Distribution Blobs
certain value would be preserved.
The Canny edge detection algorithm uses double
thresholding. Edge pixels stronger than the high threshold VI. CONTOUR TRACING AND MAPPING
are marked as strong; edge pixels weaker than the low
threshold are suppressed and edge pixels between the two With our colour palettes ready, we contour traced colour
thresholds are marked as weak. with border thresholding.
After the contour tracing we used the standard SVG Code
5. Connected Component Tracing to create a circular Gradient.

For image processing, a blob is defined as a region of

connected pixels [6,12]. Blob analysis is the identification
and study of these regions in an image. The algorithms
discern pixels by their value and place them in one of two
categories: the foreground (typically pixels with a non-zero
value) or the background (pixels with a zero value).

Fig. 7. Colour Contour

(a) (b)

Fig. 5. (a) Accepted Edges (b) Rejected Edges

166

Authorized licensed use limited to: University of Maryland College Park. Downloaded on November 21,2020 at 17:57:32 UTC from IEEE Xplore. Restrictions apply.
 Fig. 8. SVG Smooth Gradient Convert

a. SVG Coding
<radialGradient id="pink" cx="50%" cy="50%" r="50%" fx="50%" fy="50%">
<stop offset="0%" style="stop-color:rgb(255,192,203);stop-opacity:0"/> Fig 12: Image Resolution (Pixels) vs Processing Time (Rounded
<stop offset="100%" style="stop-color:rgb(178,34,34);stop-opacity:1"/> Seconds)
</radialGradient>
REFERENCES
V. FINAL RENDERING
[1] I. T. Young, J. J. Gerbrands, and L. J. v. Vliet, Fundamentals of
Image Processing, 1995, Delft: PH Publications. pp. 91-98
[2] R. C. Gonzalez, and R. E. Woods, Digital Image Processing, 1st
Edition, Addison-Wesley, 1992, pp.
[3] 253-254, 268-270, 298-30Chin-Chen Chang, Ju-Yuan Hsiao, and
Chih-Ping Hsie, “A Fast Noise Reduction Method Based on Human
Visual System,” in ICICS-PCM 2003, Singapore.
[4] Mark Grundland, Neil A. Dodgso, “Decolorize: fast, contrast
enhancing, color to grayscale conversion,” unpublished.
[5] Alexander Neubeck and Luc Van Goo, “Ef¿cient Non-Maximum
Suppression, The 18th International Conference on Pattern
Recognition (ICPR'06)
Fig. 9. Original Image [6] Sarah Sookman, Matrox Imaging, “Blob Analysis and Edge Detection
In the Real World”
[7] Chapeau-Blondeau, F., Chauveau, J., Rousseau, D., and Richard, P.
Fractal structure in the color distribution of natural images. Chaos,
Solitons and Fractals 42, 1 (2009), 472 – 482
[8] Omer, I., and Werman, M. Color lines: Image specic color represen-
tation. Computer Vision and Pattern Recognition, IEEE Computer
Society Conference on 2 (2004), 946 – 953
[9] Mandelbrot, B. The fractal geometry of nature. Freeman, 1983
[10] Special issue on digital encoding of graphics, Proc. IEEE 68,
755-929 (1980).
[11] W. K. Pratt, Digital Image Processing, John Wiley & Sons,
Inc., New York, 1978.
[12] T . Pavlidis, Structural Pattern Recognition, Springer-Verlag,
Fig. 10. Normal Linear Upscaled Image
Berlin, 1977

Fig. 11. Vectorized Image

VI. RESULT

167

Authorized
View publicationlicensed
stats use limited to: University of Maryland College Park. Downloaded on November 21,2020 at 17:57:32 UTC from IEEE Xplore. Restrictions apply.