Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Noise effect analysis on edge detection in detecting digits with bilateral

filter
To cite this article: Insidini Fawwaz et al 2019 J. Phys.: Conf. Ser. 1230 012095

View the article online for updates and enhancements.

This content was downloaded from IP address 179.61.160.117 on 06/09/2019 at 13:31

MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

Noise effect analysis on edge detection in detecting digits with

bilateral filter

Insidini Fawwaz*, N P Dharshinni, Fadhillah Azmi

Faculty of Technology and Computer Science, Universitas Prima Indonesia, Indonesia

Email:*[email protected]

Abstract. Recognizing characters such as digits in natural images is a challenge in image

processing. Edge detection is the first approach in image recognition. Edge detection using
Bilateral Filter can improve image quality. However, in image processing, noise often occurs
in the image which causes objects in the image to be more difficult to detect. In this paper, we
conducted experiments to detect edges by using Bilateral filters on each edge detection
operator against digits on SVHN which added noise and analyzes the effect of noise on edge
detection. In the proposed method, edge detection of SVHN with Speckle noise does not
produce many edges compared to Salt & Pepper noise so digits in SVHN can still be seen. The
performance of the edge detection operator with Bilateral Filter can also be seen from the MSE
and PSNR values. The lowest MSE value and the highest PSNR value is on Canny when
detecting SVHN edges with Speckle noise.

1. Introduction
The Street view house number (SVHN) is an image containing digits obtained from the house number
in the Google Street View image. Recognizing multi-digits numbers from image is a difficult unsolved
computer vision problem [1]. Edge detection is the main approach in image recognition and analysis
can be based on the edge of the image containing useful information. Edge detection is used to extract
the essential features in the image. The detected edges can be wide or nonexistent due to the influence
of noise [2]. In image processing, noise often occurs in an image [3]. Noise is a disturbance
experienced by the image so as to make the image becomes less good. There is a lot of noise that can
affect image quality, including salt n pepper, gaussian and speckle noise. The effect of each noise
varies on each image.
In previous studies, we have compared edge detection using gaussian filters and edge detection
using bilateral filters where the results show edge detection using bilateral filters has increased [4].
Therefore in this research purpose to analyze the effect of noise on edge detection with bilateral filters
in reading numbers on street view house number.

2. Noise
Noise is a disturbance that can affect the appearance of the image and produce unwanted effects such
as blurred objects, invisible lines, angles, and disturbing the entire object image. The causes of noise
in the image are defects in data transmission, imperfect optics, processing techniques, non-functioning
sensors and electronic interference [5].

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

2.1 Gaussian Noise

Gaussian noise is also known as a Gaussian distribution whose Probability Density Function is the
same as a statistical disturbance. Gaussian noise generally interferes with gray values in the image of
[6]. This noise due to the characteristics of the gray level and the presence of random variables
because of the characteristics of the Density Probability Function (PDF). The effect of this gaussian
noise on the image is the appearance of colored dots whose numbers are equal to the percentage of
noise. This is given as
(1)

Where g = gray value, = standard deviation and = mean.

2.2 Salt & Pepper Noise

The image which has salt & pepper noise in the image will show dark pixels in bright areas and bright
pixels in dark areas [7]. This noise generally occurs due to sharp and unexpected changes in image
signals, noise arises and causes dead pixels, analog-to-digital converter errors, etc. [8].
The pdf of Impulse noise is given by:

(2)

If b>a, gray level b will appear as a bright spot on the image. on the contrary, the gray level a will
appear like a dark spot. If there is no probability of zero value, it will form fine grains such as salt (salt
& pepper).

2.3 Speckle Noise

The nature of noise is mutipative, meaning that the greater the intensity of the image or the brighter
the image, the clearer the noise. Speckle Noise is a noise model that gives a black mark on the part in
the form of a noise-affected point. This noise appears during imperfect image capture due to weather,
image capture devices and so on [6].
The equation is given by:

v(x,y) (3)

where v(x,y) is corrupted image, is multiplicative component, and is intensity from

speckle noise.

3. Bilateral Filter
Bilateral filters are widely used for denoising [9] image filters. Bilateral filters can smooth the image
while maintaining the edge that has two kernel filters: spatial kernel and range kernel [10]. The
weights can be calculated from the Gaussian function [11]. Spatial kernel ( ) is the realization of
spatial proximity measurement is by the equation

(4)

2
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

While range kernel ( ) means the weighting of pixels in accordance with the size of the difference in
intensity of the pixel with the intensity at the pixel which is the center of analysis on the image.
Calculation of range kernel in each pixel is shown in the equation.

(5)

Bilateral Filtering ( ) computation as follows:

(6)

Where I is a grey level image, N(p) is a neighborhood of p, Ip is the intensity value at pixel location p,
Wp is the normalization factor which ensures pixel weights sum to 1

(7)

4. Edge Detection
Edge detection is used to extract the essential features [12]. Edge is the essential characteristic of an
image [13]. The edges of an image contain crucial information from the image A significant step in
image processing is edge detection. Image edges can represent the objects contained in the image,
shape, size, and sometimes also information about the texture.

4.1 Roberts Operator

Roberts operator is a first-order operator, which performs 2-D spatial gradient measurements to find
edges [2]. Roberts operators have fast computing rates and are simple operators. The image to be
detected by the edge is combined with the mask of the operator and the gradient magnitude to be
calculated directly. The Roberts operator uses the following 2x2 convolution kernels::

4.2. Canny Operator

Canny is one of the modern edge detection algorithms. This algorithm is focuses to separate the
background noise from complex image and to find the effective edges for optimized solutions. In
1986 John Canny proposed three criteria to be the basis of filter mining to optimize edge detection in
noise-sensitive images. The first step in implementing Canny is to smooth the image [2]. Then find the
image gradient in both vertical and horizontal directions [12]. Canny uses a Gaussian filter before
applying the mask. Gaussian filters can reduce the noise as much as possible [3]. Canny edge detectors
are not sensitive to noise causing Canny can detect weak edges [2].

4.3. Frei Chen Operator

Frei Chen mask contains the base vector. This is implemented in the 3x3 image area represented by
the number of nine Frei Chen masks [14].

3
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

5. Results and Discussion

The dataset used in this research are Street View House Number (SVHN) from Google Street View
image. In this research we used svhn training for edge detection.

Figure 1. Sample SVHN dataset images

Table 1. Edge detection results on SVHN with bilateral filter

SVHN Robert Canny Frei Chen

MSE : 21780.2 MSE : 21250.2 MSE : 21771.5

PSNR : 4.75018 PSNR : 4.85716 PSNR : 4.75192

MSE : 15283.4 MSE : 14581.7 MSE : 15274.6

PSNR : 6.28861 PSNR : 6.49272 PSNR : 6.29111

MSE : 9655.43 MSE : 9278.44 MSE : 9647.29

PSNR : 8.28309 PSNR : 8.45605 PSNR : 8.28675

4
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

In table we only show 3 images from the 50 datasets that we have used. The experiment was carried
out using bilateral filters on each edge detection operator. The result shows that the edge detection of
Robert can't produce an edge at all. This also happened with the other edge detection of Robert. Canny
produces pretty good edge detection. The digit in the picture can be seen clearly. As for Frei Chen,
there were several missing edges which resulted in the digit not being clearly visible.

Table 2. Edge detection results with Bilateral Filter on SVHN by adding Gaussian Noise

SVHN with GAUSSIAN ROBERT CANNY FREI CHEN

NOISE

MSE : 21633.5 MSE : 18444.3 MSE : 21620.9

PSNR : 4.77953 PSNR : 5.47219 PSNR : 4.78208

MSE : 15781.7 MSE : 12879.5 MSE : 15777.1

PSNR : 6.14925 PSNR : 7.0318 PSNR : 6.15053

MSE : 10210.1 MSE : 7972.1 MSE : 10205

PSNR : 8.04052 PSNR : 9.11508 PSNR : 8.04268

From the results of Robert's edge detection that we have done, the edges of the image do not appear at
all, but in some other edge detection results there are some visible edges even though they do not form
digits such as the first image above. Canny produces very many edges. The Gaussian noise in SVHN
is also detected as an edge so that the digits of the SVHN cannot be seen. The edge of the digit on Frei
Chen's edge detection is quite clearly seen in the first image, but in the next image the edge of the digit
is not very clearly visible.

Table 3. Edge detection results with Bilateral Filter on SVHN by adding Salt & Pepper Noise

SVHN with SALT & ROBERT CANNY FREI CHEN

PEPPER

MSE : 21279.2 MSE : 18714.3 MSE : 21251.5

PSNR : 4.85126 PSNR : 5.40906 PSNR : 4.85692

5
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

MSE : 15623.3 MSE : 13501.6 MSE : 15623.2

PSNR : 6.19309 PSNR : 6.82695 PSNR : 6.1931

MSE : 9940.13 MSE : 8384.41 MSE : 9931.06

PSNR : 8.15688 PSNR : 8.89608 PSNR : 8.16085

On the edge detection of Robert and Frei Chen shows Salt & Pepper noise is also considered an edge.
The digits in SVHN appear vague in the first image, while in the next image the digits are no longer
visible. Whereas Canny shows digits that is dimly covered with the edge of the Salt & Pepper noise on
the SVHN.

Table 4. Edge detection results with Bilateral Filter on SVHN by adding Speckle Noise

SVHN with SPECKLE ROBERT CANNY FREI CHEN

NOISE

MSE : 21382.2 MSE : 17049.5 MSE : 21377.1

PSNR : 4.83029 PSNR : 5.81367 PSNR : 4.83131

MSE : 15590 MSE : 12570.1 MSE : 15557.5

PSNR : 6.20233 PSNR : 7.13742 PSNR : 6.24114

MSE : 10047.9 MSE : 8415.62 MSE : 10040.7

PSNR : 8.11006 PSNR : 8.87994 PSNR : 8.11316

In the first picture, digits are still visible even though the Canny edge detection is not very clear. In the
second picture, the digits on the Canny and Frei Chen edge detection can still be seen, but the digits on
Robert's edge detection looks vague. And in the third picture, the digits on edge detection can no
longer be seen at all. Only Speckle noise is detected by Canny. On Frei Chen's edge detection, the
digits can still be detected even if it looks vague.
Figure 2 shows the performance comparison of the edge detection operators for detecting edges in
the original SVHN and SVHN which has been given noise where MSE in Robert and Frei Chen is not
much different. While the lowest MSE is owned by Canny in detecting SVHN edges with Speckle
noise.

6
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

Figure 2. Comparative performance of all operators in detecting edges on SVHN based on MSE

Figure 3. Comparative performance of all operators in detecting edges on SVHN based on PSNR

The highest PSNR is owned by Canny in detecting SVHN with Speckle Noise, while Robert and Frei
Chen operators have the highest PSNR when detecting the original SVHN edge.

6. Conclusion and Future Work

In this research, adding Gaussian noise to SVHN causes the digits to be difficult to detect by the edge
detection operators. Only a few SVHNs can be detected by the operator Frei Chen. Salt & Pepper
noise in the SVHN produces many edges so that the digits appear vague and invisible. Compared to
Salt & Pepper noise, Speckle noise doesn't produce too much edge. The digits in SVHN are also still
visible when edge detection is performed. The performance of the edge detection operator with
Bilateral Filter can also be seen from the MSE and PSNR values, where the lowest MSE value and the
highest PSNR is on Canny when detecting SVHN edges with Speckle noise. The performance of the
proposed algorithm can be improved by using the filter parameters or by using several other filters.
Also this can be modified by adding some other type of noise. The modification of the proposed
method will inspire future research.

7
MECNIT 2018 IOP Publishing
IOP Conf. Series: Journal of Physics: Conf. Series 1230 (2019) 012095 doi:10.1088/1742-6596/1230/1/012095

References
[1] Y Netzer et al 2011 Reading digits in natural images with unsupervised feature learning NIPS
Workshop on Deep Learning and Unsupervised Feature Learning.
[2] Bin L and Yeganeh M S 2012 Comparison for image edge detection algorithms 2 pp 01-04.
[3] Singh S and Singh R 2015 Comparison of various edge detection techniques.
[4] Fawwaz I et al 2018 The edge detection enhancement on satellite image using bilateral filter
308.
[5] Irum I et al 2015 A nonlinear hybrid filter for salt & pepper noise removal from color image 13
pp 79-85.
[6] Boyat A K and Joshi B K 2015 Noise models in digital image processing, International Journal 6.
[7] Sharma S and Kumar E K 2016 A comparison of salt and pepper noise removal filters.
[8] Kumar N and Nachamai M 2017 Noise removal and filtering techniques used in medical
images.
[9] Ancal et al 2018 An efficient image denoising scheme for higher noise levels using spatial
domain filters 11 p 625-634.
[10] X. Leng et al 2016 Hybrid bilateral filtering algorithm based on edge detection 10 pp 809-816.
[11] J. A and Sekhar C 2013 Bilateral edge detector.
[12] Thepade S. D and Shinde Y D 2015 Improvisation of content based image retrieval using color
edge detection with various gradient filters and slope magnitude method.
[13] Srisuk S 2014 Bilateral filtering as a tool for image smoothing with edge preserving properties.
[14] Mishra A, Mahapatra S, and Banerjee S 2017 Modified frei-chen operator based infrared and
visible sensor image fusion for real-time applications 17 p 4639.
[15] SVHN dataset downloaded [online], http://ufldl.stanford.edu/housenumbers/