DOI 10.4010/2016.1647 ISSN 2321 3361 © 2016 IJESC `

Research Article Volume 6 Issue No. 6

A Study of Image Noising and Filtering Technique Surabhi1, Neha Pawar2 Research Scholar1, Assistant Professor2 SDDIET Department of Computer Sc. Barwala Haryana, India

Abstract: In the world of image processing field, removal of high density desire is always a famous area for research. In the sources of noise in images arise during image acquisition, digitization or transmission. Now study for removing fixed impulse noise from color images. Impulse noise is produced by errors in the data transmission generate in noisy sensors or message channels, or by errors through the data capture from digital cameras. Several nonlinear filters have been proposed for renovation of images impure by salt and pepper noise. Among these standard has been recognized as trustworthy method to remove the salt and pepper noise without harmful the edge details.

Keywords: Image Processing, Impulse Noise, Bilateral Filter, Image Quality Metric

1. INTRODUCTION Color images are very often ruined by impulsive noise, which Computer image processing methods mainly take two is introduced into the image by faulty pixels in the camera categories. First, the space domain processing; that is in the sensors, broadcast errors in noisy channels, poor lighting image gap of the image processing. The other is the image conditions and aging of the storage material [1]. The spatial domain. It should be use rate domain through the suppression of the disturbances introduce by the impulsive orthogonal. In the world of image processing field, removal noise is indispensable for the success of additional stages of of high density desire noise is always a famous area for the image processing pipeline. Gray scale image cover no research. The sources of noise in images arise during image colour information. They represent the brightness of the acquisition, digitization or transmission. Now study for image. This image contains 8 bits/pixel, which means it, can removing fixed impulse noise from color images. Impulse have up to 256 (0-255) different brightness levels. A ‘0’ noise is produced by errors in the data transmission generate represents black and ‘255’ denotes white. In among values in noisy sensors or message channels, or by errors through from 1 to 254 represent the different gray levels. As they the data capture from digital cameras. Several nonlinear contain the intensity information, they are also denoted to as filters have been proposed for renovation of images impure intensity images. Colour images are considered as three band by salt and pepper noise. Among these standard median filter monochrome images, where each band is of a different has been recognized as reliable method to remove the salt colour. Each band offers the brightness information of the and pepper noise without harmful the edge details. corresponding spectral band. Classic colour imagery are red, green and blue images and are also referred to as RGB Transformation in various frequency domain. Next, do images. This is a 24 bits/pixel image. reversal processing further and then it can be stop processing for image. It is also based on the actual characteristics of the ii. Image Filters image, noise and spectral division of the demographic A variety of adaptive filtering techniques has been proposed characteristics of the law. Scientists derived many de-noising for enhancing images degraded by noise [8:34]. By adjusting approaches. One of the most sensitive ways of noise energy their parameters, depending on local characteristics of the is generally concentrated in high-frequency and spectral input image, these filters preserve sharp continued changes images situated in a incomplete range of this characteristic. (edges) in the image while reducing noise. All of the filters And then low-pass filtering approach is used to de-noising or are useful for additive suppression, but are the image handing out. This is the first class of generally ineffective in eliminating impulsive noise that image processing methods. Another way is processing in the appears as very large spikes of short duration. Various filters incidence domain (such as: Fourier transform, are discussed here available for the impulse noise transform). Image denoising is often used in the field of elimination. By adjusting their parameters, dependent on photography or publishing where an image was somehow local characteristics of the input image, these filters tainted but needs to be improved before it can be printed. For reservation sharp sustained changes (edges) in the image this type of request we need to know something about the while reducing noise. poverty process in order to develop a model for it. When we have a model for the degradation process, the opposite 2.1.1 Bilateral Filter process can be applied to the image to restore it back to the original form. This type of image restoration is frequently Bilateral Filter is a nonlinear, non-iterative technique, which used in space exploration to help eliminate arte facts uses both range filtering and domain filtering instead of only generated by mechanical in a spaceship or to domain filtering used by the conventional techniques. compensate for in the optical system of a telescope. Bilateral filter is capable of smoothing the image and Image denoising finds applications in fields such as preserving edges as well. The bilateral filter can also be astronomy defined as a weighted average of nearby pixels, in a custom very similar to Gaussian convolution. The difference is that

International Journal of Engineering Science and Computing, June 2016 6892 http://ijesc.org/

the bilateral strain takes into account the change in value with iii. Noise in Images the neighbors to protect edges while smoothing. During the capture, transmission, processing or acquisition of an image it can have many kind of variations in its original The key idea of the bilateral filter is that for a pixel to affect form, this variation is frequently random and has no another pixel, it should not only inhabit a nearby location but particular pattern. In many cases, it reduces image quality. also have a similar value. Bilateral filter overcomes the This chance variation in image is called noise. Generally limitations of Gaussian low pass filtering and anisotropic noise gives an image an undesirable appearance, the most diffusion techniques of image smoothing. As bilateral significant factor that noise can cover and reduce is the filtering is the combination of both range and domain visibility of certain features within the image. The noise filtering. Hence in this both operations like good filtering is present in image can be either in additive form or in achieved at the boundary of images because of domain multiplicative form. These both forms can be represented as filtering and crisp edges can be preserved with the help of below: range filtering. When applied to the color images bilateral filter operates on the three bands of color (RGB) at same time Additive noise equation- w(x,y) = s(x,y) + n(x, y) Eq. (1.1) and it averages the perceptually similar colors together, hence Multiplicative noise equation- w(x, y) = s(x, y) n(x, y) no artifacts occur at the edges, which occurs when three Eq.(1.2) bands are filtered separately [5]. In the above equations s(x, y) represents the original signal, n(x, y) is the noise introduced in signal, w(x, y) is the image corrupted by noise and (x, y) is the pixel location. There are different sources of noise in a digital image, depending upon sources noise can be: Dark current noise, , Amplifier noise and Quantization noise, usually following types of are most common in image processing:

 Impulse noise (Salt and Pepper Noise)  Speckle Noise

3.1.1. Gaussian Noise Gaussian noise is a noise which has Gaussian distribution, which has a bell shaped probability distribution function. This noise is evenly distributed over the signal. This means that in the loud image each pixel has a value which is the sum of the true pixel value and a random Gaussian spread noise Fig2.1 Edge preserving by using bilateral filter during value. The probability distribution function of Gaussian noise denoising. [5] is

F(g) = Eq. (1.3) 2.1.2 Median Filter

The low pass filtering methods like averaging filter do blur of In the above equation g represents the gray level, m is the image, hence it also the edges of image with the mean or average, and σ is the standard deviation of the noise. noise. Low pass filter methods works well for Gaussian The following graph shows the distribution: sound but they cannot remove salt and pepper noise from the image, Hence for the removal of salt and pepper noise from image a non-linear filter called Median Filter (MF) is used. It swaps the gray level of all the pixels in the neighborhood.

Median filtering is very effective in eliminating sharp

‘spikes’ from an image. When sound is introduced in an image low pass filtering just shapes the noise where the

Median filtering succeed sin removing most of the noise [7].

If necessary, several Passes with a median filter may be needed. The median filtering actually military pixels with Fig.3.1Probability thickness Function for Gaussian noise. very distinct gray levels to have a gray level that is more similar to its neighbors. 3.2.2 Impulse (Salt-and-Pepper) Noise Model 2.1.3 Trimmed Median Filter The PDF of (bipolar) impulse noise is given by Trimmed Median Filter for noise detecting and removing, the pa for z  a whole of the image is scan and processed pixel by pixel. If  the current pixel has the utmost value (255) or minimum p(z)  pb for z  b (1.4)  value (0) then the pixel is careful as noisy pixel. To current 0 otherwise the value of a noisy pixel, a two-dimensional window  concentrated in that pixel is created. If b  a , gray level b will appear as a light dot in image. Conversely, level a will appear like a dark dot. If either P a

International Journal of Engineering Science and Computing, June 2016 6893 http://ijesc.org/

4.2.2 MSE or Pb is zero, the impulse noise is called unipolar. Impulse noise is found in situations where quick transients, such as It stands for the mean squared disparity between the original faulty [10] switching take place during imaging. Four image and distorted image. The mathematical definition for impulse noise models are reported in recent papers. MSE is:

3.2.3 Speckle Noise MN 2 Speckle noise is a multiplicative noise. This type of noise MSE1/ M  N ( aij  b ij ) occurs in almost all coherent imaging organisms such as ij11 laser, and SAR (Synthetic Aperture Radar) images. (1.7) The source of this noise is the random meddlesome between the coherent returns. Fully developed speckle noise has the 4.2.3 IMAGE ENHANCEMENT FACTOR: characteristic of multiplicative noise. Speckle noise monitors IEF is mathematically defined as: a gamma distribution and is given as:

F(g) = Eq. (1.5)

In the above equation variance is and g is the gray level. (1.8) The gamma distribution for the speckle noise is shown as following in Fig. 1.3. Where ἠ is noisy image, Y is Original image and Ỷ is denoised image. These metrics were used to check the quality of the image. It checks which technique is giving better output. The output of SMF, TMF and Bilateral Filter is tested using these parameters.

v. Related Work

Zhang and Allebach in 2008, which is an extension of the Bilateral Filter given by Tomasi and Manduchi [5].ABF can sharpen the image by increasing slope of the edges without Fig3.2Gamma Distribution for Speckle noise [10]. producing overshoot and undershoot. The Bilateral Filter can do only smoothing of image and preserve edges but it cannot iv. Image Metrics increase the slope of edges. For ABF two modifications are There are basically two approaches for image Quality done in the traditional Bilateral Filter. (1)An offset Ϛ is measurement:- introduced in the range filter in ABF and (2) Offset Ϛ and the width of range filter σr both are locally adaptive. By using  Subjective measurement: A number of viewers are these adaptive offset and width the histogram is transformed selected, tested for their visual capabilities, shown a by range filter so that the slope of edge can enhanced. The series of test scenes and queried to score the quality value of Ϛ and σr is estimated by training procedure, during of the scenes. It is the only “correct” method of this training procedure the MSE between the original and the quantifying chart image quality. However, restored image is minimized so that correct value of subjective estimation is usually too inconvenient, parameters can be estimated. The results of ABF are better time-consuming and expensive. than bilateral filter and Optimal Unsharp Masking (OUM)  Objective measurement: These are the automatic algorithm. ABF gives better image denoising and sharpening algorithms for quality assessment that could analyze then Bilateral Filter. As compared to OUM, ABF images images and report their quality without human have sharper edges and there is no overshoot and undershoot participation. Such methods could eliminate the as is present in OUM. ABF can either smooth or sharp an need for expensive subjective studies. image depending upon the value of σd (domain filter) and σr Objective image excellence metrics can be classified (range filter) or transformation of histogram. [2] according to the obtainability of an original (distortion-free) Xie, Ann Heng and Shah 2008 introduced a new edge image, with which the unclear image is to be compared. preserving filter called Saliency Bilateral Filter SBF given by them is a little extension to the bilateral filter. Like bilateral 4.1 Evaluation of Image Quality Metric filter SBF smooth away the noise and small scale structures 4.1.1. PSNR while retaining important features from the image. But the difference between the BF and SBF is that in Bilateral filter Objective image quality appraisal methods were mainly kernels used for the filtering are constant, which is not the based on simple mathematical actions such as the Euclidian case in SBF. In SBF different kernels are used for different distance between the pixels of the original image taken as the image regions. In this smooth regions of image are averaged position and its distorted type. The Peak Signal to Noise with broad filtering kernel and strong boundaries are Ratio is one of the most broadly used metrics until now due preserved with sharp kernel. The parameter σr suggest the to its analytical and computational easiness. This makes the value of range filtering kernel, larger the value of σr, broad PSNR sensible for the optimization of image code filtering will be the range filtering kernel and little the effect of range and quality enhancement systems [3]. component of filter will be for domain component. The SBF 2 PSNR= 10log10255 /MSE (1.6) method of edge preserving outperforms the bilateral filter and Adaptive Bilateral filter in many applications. [6]

International Journal of Engineering Science and Computing, June 2016 6894 http://ijesc.org/

Ming Zhang and Bahadir introduced 2008. In this framework range of functions. It uses gradient descent algorithm to they give two contributions about bilateral filtering. First, minimize error. The back propagation algorithm depends on they give an empirical study about the optimal parameter the minimizing error of the network by using the derivatives selection of bilateral filter and secondly they give a of the error function. Results show that the proposed method multiresolusion bilateral filter, which is a combination of gives better image enhancement than ABF. [11] bilateral filter and wavelet thresholding. In the optimal Anna Gabiger Rose and other give a new method for the parameter theory they show that the value of σr is more selection of bilateral filter parameters by using noise adaptive critical than the value of σd in bilateral filter and the optimal parameter tuning. As the domain filter acts as a low pass value of σr is linearly proportional to the standard divergence filter and the range filter acts as a non-linear component and of noise σn. Various connections between σr, σd and σn are plays an important role in edge preserving. These both filters shown by authors. In the second contribution about BF they are adjusted by using parameters which are tuned using noise show that low frequency noise component or coarse-grain adaptive tuning in this framework. In this work the range noise in images can be eliminated by the combination of filter is adjusted to noise level by making adaptive on the bilateral filter with wavelet thresholding. This is done by basis of standard deviation of noise. Hence this work is integrating bilateral filtering and wavelet thresholding. In this adjusted to sound level by making adaptive on the basis of technique an image is decomposed into low and high standard deviation of noise. Hence this work is introduced to frequency components and bilateral filtering is applied on the overcome the drawback of other bilateral filtering techniques calculation subbands and wavelet thresholding on the detail [5], [2], [6], [3] which uses pair of parameter values which subbands. Hence noise components can be identified and have to be tested for every image to choose the optimal removed effectively. This approach is better than bilateral values. In this work provision for both the geometric as well filter for the removal of low frequency noise components in as photometric parameters selection is given. [1] image. [8] Anand and sahambi 2009, take into account the noise vi. Techniques for image reduction occurred in Magnetic Resonance Imaging (MRI) technique, Noise cannot be removed without the loss of some which used bilateral filter in the undecimated wavelet information in the form of image detail. However noise- domain. This technique is efficient for denoising of MRI and reduction algorithms have been developed to reduce noise effectively preserves the relevant edge feature and removes without image information too much. To get pure signal by the noisy coefficients. The reconstructed MRI data has high detecting and removing noise, a wide variety of filtering Peak signal to noise ratio (PSNR) as compared to the algorithms have developed. These include both Temporal classical wavelet domain denoising approaches. In this Filters and spatial filters. For example, the median image method to remove the complex white Gaussian noise present filter give good results for impulse filtering while the mean in reconstructed MRI data at first the data is decomposed into image is best for Gaussian noise filtering. 3-level UDWT (Undecimated Wavelet Transform) coefficients using haar . Then the approximation is basically classified into two types: coefficient obtained in third level is passed through bilateral 1) Linear Technique filter. The resultant is a denoised form of the approximation 2) Non Linear Technique coefficient. The detailed coefficients at each level are denoised by using soft thresholding. Hence good denoised 6.1 Linear Techniques image can be produced by using the combination of UDWT In Linear techniques, without classifying pixel into noisy and and Bilateral Filtering. [4] non noisy pixels, a noise reduction formula is applied for all Kenny and Ashidi 2011 proposed method for the sharpness pixels of image linearly. As this algorithm is applied for both enhancement of image corrupted by blurring. To make an noisy and non noisy pixels thus the main drawback of this image blurring free or for the sharpness enhancement of algorithm is it damages the non noisy pixels. image many methods used. Unsharp Masking (UM) is one 6.2 Non- Linear Techniques method which operates by adding a fraction of the high pass It’s a two step process filtered version of the input image to the original data. The 1) Noise detection use of high pass operator in UM makes it very sensitive to 2) Noise Replacement noise, hence it also enhances the noise while sharpness There are many algorithms but with low noise condition such enhancement of image. UM also produces overshoot and algorithms works well but there performance is poor in case undershoot in the image, which results in ringing or halo of high noise conditions. To improve the range of noise artifacts in area of sharp edges due to smoothing of edges. reduction, non linear Techniques, MMF, CWMF (Centre Histogram Equalization (HE) is another technique used for Weighted Media Filter), AMF (adaptive Median Filter) sharpness enhancement but it works only for noise free Algorithm are proposed. images. Hence the need for the proposed algorithm occurs.

[7] VII. Conclusion T. Ravi 2012 which is an improvement of ABF for sharpness Impulse noise is one of the most important factors in enhancement. They use a combination of ABF and a back degrading of image quality. In this paper, a novel technique propagation neural network in this algorithm. Back is presented for detecting and removing of impulse noise and propagation further improves the image sharpening ability of Filtering method while the significant information of image, ABF. As given by [2] ABF is advantageous over UM for such as edges and texture, are remind untouched. Noise is an image sharpening it can be further improved by adding neural irrelevant data that obscures authenticity of original data. The network back propagation algorithm to it. Back propagation impulse noise corruption is common in digital image. algorithm is multilayered network, which can compute wide

International Journal of Engineering Science and Computing, June 2016 6895 http://ijesc.org/

REFERENCES Er.Neha Pawar is from Ambala City. She completed her B.Tech in 2008 and M.Tech in [1] A. G Rose, M. Kube, P. Schmitt, R. Weigel and R. Rose 2011. Er. Neha Pawar has teaching experience (2011), “Image Denoising Using two-sided Filter With of five years. Noise-Adaptive Parameter Tuning,” Fraunhofer Institute for Integrated Circuits, D-91058 Erlangen, Germany.

[2] B. Zhang and J. P. Allebach (2008), “Adaptive bilateral filter for sharpness improvement and noise removal,” IEEE Trans. on Image Process.vol.17, no.5, May2008.

[3] C. H. Lin, J. S. Tsai, and C. T. Chiu (2010), “Switching two-sided filter with a Texture/Noise Detector for Universal Noise taking away” IEEE Trans. On.Image Processing, vol. 19, pp.2307-2320, Sept. 2010.

[4] C. Shyam Anand and J. S. Sahambi (2009), “MRI Denoising Using mutual Filter in Redundant Wavelet Domain,” IIT Guwahati, India, 2009.

[5] C. Tomasi and R. Manduchi (1998), “joint filtering for gray and color images,” in Proc. ICCV pp.839–846 Bombay, 1998.

[6] Jun Xie, Pheng Ann Heng, and Mubarak Shah (2008), “Image Diffusion Using Saliency Bilateral Filter,” IEEE Trans. on information technology, vol.12, no.6 Nov 2008.

[7] Kenny Kal Vin Toh and Nor Ashidi Mat Isa (2011), “Locally adaptive bilateral cluster for image deblurring and sharpness enhancement,” IEEE Trans. on Cons. Elec. Vol. 57, No. 3, Aug 2011.

[8] Ming Zhang and Bahadir K. Gunturk (2008), “Multiresolution Bilateral Filtering for Image Denoising,” IEEE Trans. on Image Process. Vol.17, No.12, Dec 2008.

[9] Pawan Patidar, Manoj Gupta, Sumit Srivastava and Ashok Kumar Nagawat (2010) ,“Image De-noising by Various Filters for Different Noise,” International Journal of Computer Applications , vol. 9, no. 4, Nov 2010.

[10] Rafael C. Gonzales and Richard E. Woods, “,” 3rd edition, 2011.

[11] T. Ravi, C. H. Mounika, C. H. Rajesh Babu, T. Prasanth (2012), “Quality improvement of image using adaptive bilateral filter and neural networks.” International journal of modern Eng. Research, vol. 2, issue 2, Mar- Apr. 2012.

Author’s Profiles Er. Surabhi is from Panchkula. Born on 26 December 1990. She completed B.Tech (Computer Science) from Swami Devi Dyal Institute of Engineering, Barwala, India in the year 2012. She is pursuing M.Tech (Computer Science) from SDDIET, Barwala, India.

International Journal of Engineering Science and Computing, June 2016 6896 http://ijesc.org/