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

net/publication/302915829

A Comparative Study of Various Image Filtering Techniques for Removing

Various Noisy Pixels in Aerial Image

Article · March 2016

DOI: 10.14257/ijsip.2016.9.3.10

CITATIONS READS

49 5,882

2 authors:

Sheikh Tania Raghad Rowaida

Federation University Australia Southeast University (Bangladesh)
5 PUBLICATIONS 53 CITATIONS 2 PUBLICATIONS 49 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Sheikh Tania on 20 January 2018.

The user has requested enhancement of the downloaded file.

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016), pp.113-124
http://dx.doi.org/10.14257/ijsip.2016.9.3.10

A Comparative Study of Various Image Filtering Techniques for

Removing Various Noisy Pixels in Aerial Image

Sheikh Tania and Raghad Rowaida

Bangladesh University of Engineering and Technology
[email protected], [email protected]

Abstract
In advanced image processing, aerial image processing plays an important role in
object extraction such as building extraction, road detection etc. The aerial images
captured are usually bound to suffer from Gaussian noise, salt and pepper noise, speckle
noise etc. Therefore obtaining of aerial image with high accuracy is very difficult task. A
flawless aerial image is inevitable for further object extraction process. There are a
number of filtering techniques to detach the noise for preserving the integrity of captured
aerial image. In this paper we have applied mean filter, median filter, wiener filter,
wavelet transform and curvelet transform for removal of various level of Gaussian
noise, salt and pepper noise and speckle noise added separately in an aerial image. The
performance of both the transforms and filtering methods are compared in terms of Peak
Signal to Noise Ratio (PSNR), Mean Square Error (MSE).

Keywords: aerial image, noise removal, filtering, mean, median, wavelet transform,
curvelet transform

1. Introduction
With growing application in science and engineering digital image processing is treated
as a rapidly evolving field. In the real world signals do not exist without noise, which
arises during image acquisition (digitization) and/or transmission. When images are
acquired using a camera, light levels and sensor temperature are major factors affecting
the amount of noise. During transmission, images are corrupted mainly due to
interference in the channel used for transmission. Removing noise from images is an
important problem in image processing.
In the early development of image processing, linear filters were the primary tools for
image enhancement and restoration. Their mathematical simplicity and the existence of
some desirable properties made them easy to design and implement. Moreover, linear
filters offered satisfactory performance in many applications. However, they have poor
performance in the presence of non additive noise and in situations where system
nonlinearities or Gaussian statistics are encountered [5]. In image processing applications,
linear filters tend to blur the edges and do not remove Gaussian and mixed Gaussian
impulse noise effectively. Linear noise removal methods are not so effective when
transient non-stationary wideband components are involved since their spectrum is similar
to the spectrum of noise, the basic idea that the energy of a signal will often be
concentrated in a few coefficients in the transform domain while the energy of noise is
spread among all coefficients in transform domain. Therefore, the nonlinear methods will
tend to keep a few larger coefficients representing the signal while the noise coefficients
will tend to reduce to zero. Noise removal methods based on multiresolution transforms
involves three steps: A linear forward transform, nonlinear thresholding step and a linear
inverse transform. Wavelets are successful in representing point discontinuities in one
dimension, but less successful in two dimensions. As a new multiscale representation
suited for edges and other singularity curves, the curvelet transform has emerged as a

ISSN: 2005-4254 IJSIP

Copyright ⓒ 2016 SERSC
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

powerful tool. The developing theory of curvelets predict that, in recovering images
which are smooth away from edges, curvelets obtain smaller asymptotic mean square
error of reconstruction than wavelet methods[7].
In this comparative study, various filtering algorithms are used to fully remove noise
from aerial images and to preserve the quality of them. These filtering algorithms have
various advantages and disadvantages. Among the different filters, none of them
overcome others in respect to computation cost, noise removing and quality of resultant
image. As a result, noise removal method can be improved and still is an open research
area.

2. Types of Noise
Noise represents unwanted information which collapses image quality. In the image
noise removal process, information about the type of noise present in the original image
plays a significant role. Typical images are corrupted with noise modeled with either a
Gaussian, uniform, or salt or pepper distribution. Another typical noise is a speckle noise,
which is multiplicative in nature. The behavior of each of these noises is described below.

2.2. Gaussian Noise

Gaussian noise statistical noise that has a probability density function of the normal
distribution that is also known as Gaussian distribution. In other words, the values that the
noise can take on are Gaussian-distributed. Gaussian noise is properly defined as the noise
with a Gaussian amplitude distribution. Noise is modeled as additive white Gaussian
noise (AWGN), where all the image pixels deviate from their original values following
the Gaussian curve. That is, for each image pixel with intensity value f (1 ≤ i ≤ m, 1 ≤ j ≤
ij
n for an m x n image), the corresponding pixel of the noisy image g is given by,
ij

(1)
Where, each noise value n is drawn from a zero -mean Gaussian distribution.

2.2. Salt and Pepper Noise

Salt and pepper noise is caused by sharp, sudden disturbances in the image signal and
its appearance is randomly scattered white or black (or both) pixels over the image. An
image containing salt and pepper noise will have dark pixels in bright regions and bright
pixels in dark regions. This noise type can be caused by dead pixels, analog to digital
converter errors and bit errors in transmission.

2.3. Speckle Noise

Speckle noise affects all inherent characteristics of coherent imaging, including
medical ultra sound imaging. It is caused by coherent processing of backscattered signals
from multiple distributed targets. Speckle noise is caused by signals from elementary
scatters. In medical literature, speckle noise is referred to as „texture‟ and may possibly
contain useful diagnostic information. For visual interpretation, smoothing the texture
may be less desirable. Physicians generally have a preference for the original noisy
images, more willingly, than the smoothed versions because the filter, even if they are
more sophisticated, can destroy some relevant image details. Thus it is essential to
develop noise filters which can preserve the features that are of interest to the physician.
Several different methods are used to eliminate speckle noise, based upon different
mathematical models of the phenomenon. In our work, we recommend hybrid filtering
techniques for removing speckle noise in ultrasound images. The speckle noise model has
the following form (denotes multiplication). For each image pixel with intensity value f ij

114 Copyright ⓒ 2016 SERSC

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

(1 ≤ i ≤ m, 1 ≤ j ≤ n for an m x n image), the corresponding pixel of the noisy image g ij is

given by,
(2)
Where, each noise value n is drawn from uniform distribution with mean 0 and
2
variance ϭ .

3. Some Existing Filtering Techniques

3.1. Mean Filter
Mean filtering is a simple, intuitive and easy to implement method to reduce noise in
images by reducing the amount of intensity variation between one pixel and the next. It is
often used for smoothing. The idea of mean filtering is simply to replace each pixel value
in an image with the mean or average value of its neighbors, including itself. This has the
effect of eliminating pixel values which are unrepresentative of their surroundings. Mean
filtering is usually thought of as a convolution filter. Like other convolutions it is based
around a kernel, which represents the shape and size of the neighborhood to be sampled
when calculating the mean. Often a 3×3 square kernel is used, as shown in Figure 1,
although larger kernels for example 5×5 squares can be used for more severe smoothing.
The main problem with mean filtering is that a single pixel with a very
unrepresentative value can significantly affect the mean value of all the pixels in its
neighborhood.

Figure 1. 3×3 Averaging Kernel often used in Mean Filtering

3.2. Median Filter

Median filter often does a better job than the mean filter of preserving useful detail in
the image. Like the mean filter, the median filter considers each pixel in the image in turn
and looks at its nearby neighbors to decide whether or not it is representative of its
surroundings. Instead of simply replacing the pixel value with the mean of neighboring
pixel values, it replaces it with the median of those values. The median is calculated by
first sorting all the pixel values from the surrounding neighborhood into numerical order
and then replacing the pixel being considered with the middle pixel value.
In general, the median filter allows a great deal of high spatial frequency detail to pass
while remaining very effective at removing noise on images where less than half of the
pixels in a smoothing neighborhood have been effected. As a consequence of this, median
filtering can be less effective at removing noise from images corrupted with Gaussian
noise.
One of the major problems with the median filter is that it is relatively expensive and
complex to compute. To find the median it is necessary to sort all the values in the
neighborhood into numerical order and this is relatively slow, even with fast sorting
algorithms such as quick sort.

Copyright ⓒ 2016 SERSC 115

International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

3.3. Wiener Filter

The Wiener filter is the MSE-optimal stationary linear filter for images degraded by
additive noise and blurring. Calculation of the Wiener filter requires the assumption that
the signal and noise processes are second-order stationary (in the random process sense).
Wiener filters are often applied in the frequency domain. Given a degraded image X (n,
m), one takes the Discrete Fourier Transform (DFT) to obtain X (u, v). The original image
spectrum is estimated by taking the product of X (u, v) with the Wiener filter G (u,v):
(3)
The inverse DFT is then used to obtain the image estimate from its spectrum. The
Wiener filter is defined in terms of these spectra:
H (u, v) is the Fourier transform of the point spread function (PSF).
Ps (u, v) is the Power spectrum of the signal process, obtained by taking the Fourier
transform of the signal autocorrelation.
Pn (u, v) is the Power spectrum of the noise process, obtained by taking the Fourier
transform of the noise autocorrelation.
The Wiener filter is:

(4)
Dividing through by Ps makes its behavior easier to explain:

(5)
The term Pn/Ps can be interpreted as the reciprocal of the signal-to-noise ratio. Where
the signal is very strong relative to the noise, Pn/Ps≈ 0 and the Wiener filter becomes H-
1
(u, v) -the inverse filter for the PSF. Where the signal is very weak, P n/Ps → ∞ and G (u,
v)→ 0.
For the case of additive white noise and no blurring, the Wiener filter simplifies to:

(6)
Where ϭn2 is the noise variance.

3.4. Wavelet Transform

Wavelet domain is advantageous because DWT make the signal energy concentrate in
a small number of coefficients; hence, the DWT of a noisy image consists of number of
coefficients having high Signal to Noise Ratio (SNR) while relatively large number of
coefficients is having low SNR. After removing the coefficients with low SNR, the image
is reconstructed using inverse DWT [3]. Time and frequency localization is
simultaneously provided by Wavelet transform. Moreover, wavelet methods represent
such signals much more efficiently than either the original domain or Fourier transforms
[8].
The DWT is same as hierarchical sub band system where the sub bands are
logarithmically spaced in frequency and represent octave-band decomposition. Image is
decomposed into four sub-bands and critically sampled by applying DWT as shown in
Figure 1. These sub bands are formed by separable applications of horizontal and vertical
filters. Sub-bands with label LH1, HL1 and HH1 correspond to finest scale coefficient

116 Copyright ⓒ 2016 SERSC

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

while sub-band LL1 represents coarse level coefficients [6] [3]. The LL1 sub band is
further decomposed and critically sampled to find out the next coarse level of wavelet
coefficients as shown in Figure 1. It results in two level of wavelet decomposition.

Figure 2. Image Decomposition by using DWT

Wavelet thresholding is a signal estimation technique that exploits the capabilities of
Wavelet transform for removal of noise from signal. It removes noise by killing
coefficients that are irrelevant relative to some threshold [6] .Several studies are there on
thresholding the Wavelet coefficients. The process, commonly called Wavelet Shrinkage,
consists of following main stages:

Figure 3. Block Diagram of Image Noise Removal using Wavelet Transform

• Read the noisy image as input.

• Perform DWT of noisy image and obtain Wavelet coefficients.
• Estimate noise variance from noisy image.
• Calculate threshold value using various threshold selection rules or shrinkage rules.
• Apply soft or hard thresholding function to noisy coefficients.
• Perform the inverse DWT to reconstruct the denoised image.

There are two thresholding functions frequently used as Hard threshold [4] and Soft
threshold. Hard thresholding function keeps the input if it is larger than the threshold;
otherwise, it is set to zero. Soft thresholding function takes the argument and shrinks it
toward zero by the threshold. Soft thresholding rule is chosen over hard thresholding, for
the soft thresholding method yields more visually pleasant images over hard thresholding.
A small threshold may yield a result close to the input, but the result may still be noisy.
Large threshold alternatively, produces signal with large number of zero coefficients. This
leads to a smooth signal.

3.5. Curvelet Transform

The Curvelet transform is a higher dimensional generalization of the wavelet transform
designed to represent images at different scales and different angles. Curvelet transform is
a special member of the multiscale geometric transforms. It is a transform with multiscale

Copyright ⓒ 2016 SERSC 117

International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

pyramid with many directions at each length scale. Curvelets will be superior over
wavelets in following cases: [1]
 Optimally sparse representation of objects with edges.
 Optimal image reconstruction in severely ill-posed problems.
 Optimal sparse representation of wave propagators.
The idea of the Curvelet transform is first to decompose the image into subbands
means to separate the object into a series of disjoint scales. Curvelets are initially
introduced by Candes and Donoho[9]. The Discrete Curvelet transform (DCT) takes as
input a Cartesian grid of the form f(n1, n2), 0≤n1, n2<n, and outputs a collection of
coefficients cD( j, l, k) defined by,

(7)
D
Where ø j, l, k are digital curvelet waveforms which preserve the listed properties of the
continuous curvelet. DCT can be implemented in two ways. The first method is based on
unequally spaced fast Fourier transform (USFFT) and the second is based on the
Wrapping of specially selected Fourier samples [2]. The two implementations essentially
differ by spatial grid used to translate curvelets at each scale and angle.
For the 2D image, the architecture of the DCT via Wrapping is as follows:
 Apply the 2D FFT and obtain Fourier samples.
f [n1, n2], -n/2≤ n1, n2 <n/2
 For each scale j and angle l, form the product.
Uj, l [n1, n2] f[n1, n2]
 Wrap this product around the origin and obtain.
f‟j, l [n1, n2] = W (Uj, l f) [n1, n2]
 Apply the inverse 2D FFT to each f‟j, l [n1, n2] hence collecting the discrete
coefficients cD( j, l, k).
The curvelet noise removal method consists of the following steps:
 Estimate the noise standard deviation σ in the input image.
 Calculate the Curvelet transform of the input image. We get a set of bands w j,
each band wj contains Nj coefficients and corresponds to a given resolution level.
 Calculate the noise standard deviation ϭj for each band j of the Curvelet
transform.
 For each band j do: Calculate the maximum of the band and multiply each
curvelet coefficient.
 Reconstruct the image from the modified curvelet coefficients.

4. Comparative Study
The experimental evaluation is performed on aerial image pentagon.tiff of size
1024*1024 pixels downloaded from the USC-SIPI image database. The objective quality
of the reconstructed image is measured by Root Mean Square Error (RMSE) and Peak
Signal to Noise Ratio (PSNR). RMSE is a measure of the "average" error, weighted
according to the square of the error. PSNR is defined as the ratio of signal power to noise

118 Copyright ⓒ 2016 SERSC

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

power. It basically obtains the gray value difference between resulting image and original
image.

(8)

(9)
Here f (i, j) is the original aerial image with noise, g (i, j) is the enhanced image and m
and n are the total number of pixels in the horizontal and vertical dimensions of the
image. The low value of RMSE indicates the better enhancement approach but the high
PSNR value indicates the better ones. The original noisy image and filtered image
obtained by various filtering techniques are shown in Figure 4 to 8. Visual comparison of
PSNR and RMSE of Gaussian noise, salt and pepper noise and speckle noise removal
with respect to Mean Filter, Median Filter, Wiener Filter, Wavelet Transform and
Curvelet Transform are shown in Figure 9 to 11. Table 1 to 6 show the RMSE and PSNR
values of image filtered by five different types of filters.

5. Conclusion
Noise is one of obstructions in automatic image understanding and noise reducing is
very important to improve the results of further processing. In this paper various filtering
techniques are implemented on aerial image to remove different types of noise. The
results are analyzed and evaluated. The comparative study is conducted with Mean Square
Error (MSE) and Peak Signal to Noise Ratio (PSNR). Through this work it is observed
that in case of Gaussian noise the performance of all the filters and transformation
techniques are almost same. On the other hand for salt and pepper noise the filters work
better than the transforms especially median filter performs the best. In addition wavelet
transform performs better than all the filters and transforms for speckle noise. The results
achieved through this paper are useful for various purposes to analyze the image.

Figure 4. Three Different Types of Noise and Corresponding Enhanced

Image using Mean Filter

Copyright ⓒ 2016 SERSC 119

International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

Figure 5. Three Different Types of Noise and Corresponding Enhanced

Image using Median Filter

Figure 6. Three Different Types of Noise and Corresponding Enhanced

Image using Winner Filter

120 Copyright ⓒ 2016 SERSC

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

Figure 7. Three Different Types of Noise and Corresponding Enhanced

Image using Wavelet Transform

Figure 8. Three Different Types of Noise and Corresponding Enhanced

Image using Curvelet Transform

Figure 9. Comparison of PSNR and RMSE of Gaussian Noise Removal with

Respect to MF, MedF, WF, WT and CT

Copyright ⓒ 2016 SERSC 121

International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

Figure 10. Comparison of PSNR and RMSE of Salt and Pepper Noise
Removal with Respect to MF, MedF, WF, WT and CT

Figure 11. Comparison of PSNR and RMSE of Speckle Noise Removal with
Respect to MF, MedF, WF, WT and CT

Table 1. RMSE Values for 1024*1024 Aerial Images with Different

Percentage of Gaussian Noise Denoised by MF, MedF, WF, WT and CT

Filter Type 10% 20% 30% 40% 50%

Mean 28.2451 51.7816 74.7164 94.4168 108.2353
Median 28.8517 52.6167 76.704 98.2739 112.5381
Winner 28.209 51.8465 74.8726 94.5835 108.4485
WT 28.691 52.105 75.0553 94.6587 108.4633
CT 165.7823 190.4116 213.2655 232.3373 245.2936

Table 2. RMSE values for 1024*1024 aerial images with Different Percentage
of Salt and Pepper Noise Denoised by MF, MedF, WF, WT and CT
Filter Type 10% 20% 30% 40% 50%
Mean 16.982 22.4605 26.9777 31.0613 34.8761
Median 8.8981 11.4271 18.08 28.7584 42.8411
Winner 24.9245 29.6095 33.2423 36.3916 39.3651
WT 19.9036 19.5159 22.8162 25.8868 28.9843
CT 142.8528 146.1774 150.034 153.9973 158.1127

122 Copyright ⓒ 2016 SERSC

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

Table 3. RMSE values for 1024*1024 Aerial Images with Different Percentage
of Speckle Noise Denoised by MF, MedF, WF, WT and CT
Filter Type 10% 20% 30% 40% 50%
Mean 17.4412 22.6716 26.7676 29.9276 32.2137
Median 24.6193 33.6461 40.5063 46.2263 51.0681
Winner 19.6618 26.106 30.6589 34.2094 36.3184
WT 16.1649 19.5505 22.3364 24.8429 26.6302
CT 142.2755 146.376 149.8805 152.8253 155.0543

Table 4. PSNR Values for 1024*1024 Aerial Images with Different Percentage
of Gaussian Noise Denoised by MF, MedF, WF, WT and CT

Filter Type 10% 20% 30% 40% 50%

Mean 19.1459 13.8813 10.6965 8.6638 7.4774
Median 18.9614 13.7423 10.4684 8.316 7.1388
Winner 19.157 13.8704 10.6783 8.6485 7.4603
WT 19.0099 13.8272 10.6572 8.6416 7.4591
CT 3.774 2.5709 1.5864 0.8424 0.3711

Table 5. PSNR values for 1024*1024 Aerial Images with Different Percentage
of Salt and Pepper Noise Denoised by MF, MedF, WF, WT and CT

Filter Type 10% 20% 30% 40% 50%

Mean 23.56 21.1364 19.5447 18.3204 17.3142
Median 29.1789 27.0061 23.0208 18.9895 15.5276
Winner 20.2323 18.7362 17.731 16.9448 16.2626
WT 22.1862 22.357 20.9999 19.9032 18.9215
CT 5.067 4.8672 4.641 4.4145 4.1855

Table 6: PSNR Values for 1024*1024 Aerial Images with Different Percentage
of Speckle Noise Denoised by MF, MedF, WF, WT and CT

Filter Type 10% 20% 30% 40% 50%

Mean 23.3333 21.0551 19.6126 18.6434 18.004
Median 20.3393 17.6261 16.0143 14.867 14.0018
Winner 22.2923 19.83 18.4337 17.4819 16.9623
WT 23.9933 22.3416 21.1846 20.2607 19.6573
CT 5.1022 4.8554 4.6499 4.4809 4.3551

References
[1] K. Ding, “Wavelets, curvelets and Wave Atoms for Image Denoising”, 2010 3rd International Congress
on Image & Signal Processing (2010).
[2] J. Tao and Zhao Xin, 2008, “Research and Application of Image Denoising Method Based on Curvelet
Transform” The International Archives of the Photogrammetry, Remote Sensing and Spatial Information
Sciences”, vol. 37, Part B2, Beijing, (2008).
[3] S. Arivazhagan, S. Deivalakshmi and K. Kannan, “Performance Analysis of Image Denoising System for
different levels of Wavelet decomposition”, International Journal of Imaging Science and Engineering
(IJISE), vol. 1, no. 3, (2007).
[4] Q. Pan, “Two denoising methods by wavelet transform”, IEEE Trans. Signal Processing, vol. 47, (1999),
pp. 3401–3406.
[5] I. Pitas and A. N. Venetsanopoulos, “Nonlinear Digital Filters: Principles and Applications”, NJ:
Springer Publisher, (1990).

Copyright ⓒ 2016 SERSC 123

 International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol.9, No.3 (2016)

[6] R. Kumar and B. S. Saini,“Improved Image Denoising Techniques Using Neighbouring Wavelet
Coefficients of Optimal Wavelet with Adaptive Thresholding”, International Journal of Computer
Theory and Engineering, vol. 4, no. 3, 2012.
[7] R. C. Gonzalez, R. E. Woods and S. L. Eddins, “Digital Image Processing Using MATLAB”, Pearson
Prentice Hall, (2004).
[8] I. Ram and M. Elad, “Generalized Tree-Based Wavelet Transform”, IEEE Transactions On Signal
Processing, vol. 59, no. 9, (2011).
[9] E. J. Candes and D. Donoho, “Curvelet: a surprisingly effective nonadaptive representation for object
withedges”, Proceeding of Curves and Surfaces IV, France, (1999), pp.105-121.

Authors
Sheikh Tania, She is an M. Sc. Student in Bangladesh University
of Engineering and University. She completed her B. Sc. from Dept.
of CSE, Computer Science and Engineering from Jahangirnagar
University. Now she is working as Assistant Professor in Bangladesh
University of Business and Technology.

Raghad Rowaida, She is an M. Sc. Student in Bangladesh

University of Engineering and University. She completed her B. Sc.
from Dept. of CSE, Computer Science and Engineering from
Ahsanullah University. Now she is working as lecturer in Bangladesh
University of Business and Technology.

124 Copyright ⓒ 2016 SERSC

View publication stats