Intelligent Denoising Technique for Spatial Video Denoising for real-time applications

Rasha Orban Mahmoud

Nile Institute of Commerce & Computer Technology, Mansoura, Egypt Mohamed T. Faheem, Amany Sarhan

Computers and Automatic Control Dept., Faculty of Engineering, Tanta University, Egypt

Abstract- With the wide spread of video usage in many fields of our lives, it becomes very important to develop new techniques for video denoising. Spatial video denoising using wavelet transform has been the focus of the current researches as it requires less computation and more suitable for real-time applications. Two specific techniques for spatial video denoising using wavelet transform are considered in this work: 2D Discrete Wavelet Transform (2D DWT) and 2D Dual Tree Complex Wavelet Transform (2D DTCWT). Each of these techniques has its advantages and disadvantages. The first technique gives less quality at high levels of noise but consumes less time while the second gives high quality video while consuming long. In this work, we introduce an intelligent denoising system that makes a tradeoff between the quality of the denoised video and the time required for denoising. The system first estimates the noise level in the video frame then accordingly chooses the proper of the two denoising techniques to apply on the frame. The simulation results show that the proposed system is more suitable for real-time applications where the time is critical while giving high quality videos especially at low to moderate levels of noise.

I. INTRODUCTION

The recent advancement in multimedia technology has promoted an enormous amount of research in the area of image and video processing. Included in the many image and video processing applications, such as compression, enhancement, and target recognition, is preprocessing functions for noise removal. Noise removal is one of the most common and important processing steps in many image and video systems. Because of the importance and commonality of preprocessing in most image and video systems, there has been an enormous amount of research dedicated to the subject of noise removal, and many different mathematical tools have been proposed [2].

Noise refers to unwanted stochastic variations as opposed to deterministic distortions such as shading or lack of focus. It can be added to the video signal or multiplied with the video signal. It can also be signal dependent or signal independent [6]. Based on its spectral properties, noise is further classified as white or color noise. Many types of noise effect charge-coupled device (CCD) cameras such as photon shot noise and read out noise. Photon shot noise is due to the random arrival of photons at the sensor, which is governed by Poisson distribution. Other sources of noise include output amplifier noise, camera noise and clock noise, which can be combined in a single equivalent Gaussian noise source called read out noise. Because of the high counting effect of Photon arrivals and according to the central limit theorem, the aggregate noise effect can be well approximated by Gaussian distribution. Consequently, in this

paper, an Additive White Gaussian Noise (AWGN) model is assumed. The choice is also motivated by AWGN being the most common noise model for TV broadcasting [6].

Spatial video denoising techniques use both the two dimensional Dual Tree Complex Wavelet Transform (2D DTCWT) and three dimensional Dual Tree Complex Wavelet Transform (3D DTCWT)[12], temporal video denoising techniques use temporal filtering only [16], while spatio-temporal video denoising techniques use combination of spatial and temporal denoising [16].

The need for fast and accurate video noise estimation algorithms rises from the fact that many fundamental video processing algorithms such as compression, segmentation, motion estimation and format conversion adapt their parameters and improve performance when the noise is known. The effectiveness of video processing methods can be significantly reduced in the presence of noise. When information about the noise becomes available, processing can be adapted to the amount of noise to provide stable processing methods [5].

A noise estimation technique calculates the level of white Gaussian noise, which is the most commonly assumed noise type in video processing applications, contained in a corrupted video signal. When noise variance becomes available, video denoising algorithms (e.g., 2D Discrete Wavelet Transform (2D DWT) and 2D Dual Tree Complex Wavelet Transform (2D DTCWT)) can be adapted to the amount of noise for improved performance.

Video noise can be estimated spatially or temporally. A widely used spatial noise estimation method calculates the variance (as a measure of homogeneity) over a set of image blocks and averages the smallest block variance as an estimate of the image noise variance. Spatial variance-based methods tend to overestimate the noise in less noisy images and underestimate it in highly noisy and textured images. Therefore, measures other than the variance were introduced in [5] to determine homogeneous blocks. Temporal noise estimation evaluates noise using motion information [15]. Such approach is very expensive for hardware implementations with estimation accuracy not significantly more precise than spatial methods.

This paper aims to introduce a novel intelligent denoising system for spatial video denoising. Two of the most widely used denoising techniques will be used namely: 2D DWT and 2D DTCWT. Each of the techniques has its advantages and disadvantages. DWT has the advantage of consuming minor in the denoising process; however, it has the disadvantages of producing less quality video at high levels of noise compared to the DTCWT. DTCWT has the

407978-1-4244-2116-9/08/$25.00 ©2008 IEEE

∑ =−k

mkk cc ,0. 2

( ) ( ) ( ) ( ) ( )∑ ∑ ∑ ∑++++= −−k k k k

kkkjkjkjkjkjkj xwdxwdxwdxsxf ,1,1,1,1,,,, ...φ

advantage of producing high quality video at high levels of noise while having the disadvantages of consuming minor in the denoising process compared to the DWT, which makes it unsuitable for real-time applications. Therefore, a tradeoff has to be made between the quality of the produced video and the time consumed for denoising between the two techniques to get the benefits and discarding the disadvantages of both.

The first component in the system is the noise estimator that estimates the noise contained in the frame (as we will work on spatial domain). The system will make a decision on how to handle this frame according to the noise level. This will lead to an efficient and flexible system for fast and reliable spatial video denoising.

II. VIDEO DENOISING TECHNIQUES

Denoising is still one of the most fundamental, widely studied, and largely unsolved problems in video processing. The purpose of denoising (or restoration) is to estimate the original video (or a “better” representative of it) from noisy data. Many methods for video denoising have been suggested, but the wavelet transform has been viewed by many as the preferred technique for noise removal [10]. Rather than a complete transformation into the frequency domain, as in DCT or FFT, the wavelet transform produces coefficient values which represent both time and frequency information. The hybrid spatial-frequency representation of the wavelet coefficients allows for analysis based on both spatial position and spatial frequency content. The hybrid analysis of the wavelet transform is excellent in facilitating video denoising algorithms [14]. A. Video Denoising Techniques Based on Wavelet

Transform In recent years, the multiresolution analysis, more

specifically the wavelet transform, has shown considerable success in signal denoising. Wavelet analysis is a powerful and popular tool for the analysis of nonstationary signals. The wavelet transform is a joint function of a time series of interest x (t) and an analyzing function or wavelet ψ(t). This transform isolates signal variability both in time t, and also in “scale” s, by rescaling and shifting the analyzing wavelet. The wavelet itself can be said to play the role of a lens through which a signal is observed. Therefore, it is important to understand how the wavelet transform depends upon the wavelet properties [9, 10]. There are two famous types of video denoising that use wavelet transform namely; 2D Discrete Wavelet Transform (2D DWT) and 2D Dual Tree Complex WT (2D DTCWT). B. 2-D Discrete Wavelet Transform (2D DWT)

The 2D DWT is a very modern mathematical tool. It is used in compression, denoising and watermarking applications. It is built with separable orthogonal mother wavelets, having a given regularity. The DWT gives a multiscale representation of a signal x (n). The DWT is implemented by iterating the 2-channel analysis filter bank described above. Specifically, the DWT of a signal is obtained by recursively applying the lowpass/highpass frequency decomposition to the lowpass output as illustrated in the diagram, see Fig.1. The diagram illustrates a 3-scale DWT. The DWT of the signal x is the collection of subband signals. The inverse DWT is obtained by iteratively applying the synthesis filter bank [14].

Figure 1. DWT Multi-scale representation of a signal x. DWT has the following advantages: Multi-scale signal processing technique. Number of significant output samples is very small and hence the extracted features are well characterized. Straightforward computation technique.

Although the Discrete Wavelet Transform (DWT) in its maximally decimated form (Mallat's dyadic filter tree [4]) has established an impression, its use for other signal analysis and reconstruction tasks has been hampered by two main disadvantages:

Lack of shift invariance, which means that small shifts in the input signal can cause major variations in the distribution of energy between DWT coefficients at different scales. Poor directional selectivity for diagonal features, because the wavelet filters are separable and real.

The 2D DWT is simply the application of the 1D-WT repeatedly to first horizontal data of the image, then the vertical data of the image. The discrete wavelet transform [4] is an algorithm for computing the coefficients sj,k and dj,k in the wavelet expansion of a signal.

Where j is the number of multiresolution components (or scales), and k ranges from 1 to the number of coefficients in the specified component. ф is the scaling function and the w is the wavelet function through dilation and translation as follows

( ) ( ) ( ) ( )kxwxwandkxx jj

kjj

j

kj −=−= −−

−− 2222 2,

2, φφ (2)

The scaling ф(x) function is the solution of the dilation equation

( ) ( )∑ −=k

k kxcx 22 φφ

Where the coefficients ck must satisfy the following conditions [4]: Unit vector:

Double-shift: m =1,2,…, p-1.

Approximation of order p: m= 0,1,…, p-1.Where p = (number of coefficients)/2.

While, the wavelet function w(x) can be derived from the corresponding scaling function by taking difference. For the four-coefficient scaling function, the wavelet equation is expressed as:

Where dk= (-1)2p-k-1c2p-1-k More precisely, the expansion in (1) for any arbitrary

signal f(x) may take the form ( ) ( ) ( )kxwakxaxf j

kjkk j

k −+−= ∑∑ ∑∞

−∞=

∞

−∞=

∞

=

2,0

φ (5)

Where the coefficients are given by

( ) ( )∑ −=k

k kndnw 22 φ

(4)

( )∑ =−k

kmk ck ,01

(1)

∑ =k

kc .12

408

Figure 2. Mother wavelet function (Daubechie’s 4).

( ) ( ) ( ) ( )∫ ∫∞

∞−

∞

∞−

−=−= dxkxwxfaanddxkxxfa jkjk 2. ,φ

This wavelet series expansion decomposes f(x) into an

infinite summation of wavelets at different scales. For computing the coefficients ak and aj,k in (5) when f(x) is sampled over some certain interval, the discrete wavelet transform is employed.

To use the wavelet transform for image processing we must implement a 2D version of the analysis and synthesis filter banks. Fig. 3 shows 2-Channel Perfect Reconstruction Filter Bank. C. 2-D Dual Tree Complex WT (2D DTCWT)

The dual-tree CWT comprises of two parallel wavelet filter bank trees that contain carefully designed filters of different delays that minimize the aliasing effects due to downsampling [4]. The dual-tree CDWT of a signal x(n) is implemented using two critically-sampled DWTs in parallel on the same data, as shown in Fig. 4. The transform is two times expansive because for an N-point signal it gives 2N DWT coefficients. If the filters in the upper and lower DWTs are the same, then no advantage is gained. Therefore, the filters are designed in a specific way such that the subband signals of the upper DWT can be interpreted as the real part of a complex wavelet transform and subband signals of the lower DWT can be interpreted as the imaginary part. When designed in this way the DTCDWT is nearly shift invariant, in contrast to the classic DWT.

Moreover, the dual-tree complex DWT can be used to implement 2D wavelet transforms where each wavelet is oriented, which is especially useful for image processing. (For the 2D DWT, recall that one of the three wavelets does not have a dominant orientation.) The DTCWT outperforms the critically-sampled DWT for applications like image denoising and enhancement.

One of the advantages of the DTCWT is that it can be used to implement 2D wavelet transforms that are more selective w

Figure 3. 2-Channel Perfect Reconstruction Filter Bank.

Figure 4. The Dual-Tree complex DWT of a signal x.

with respect to orientation than is the 2D DWT [14]. Let w2 represent the parent of w1 (w2 is the wavelet coefficient at the same spatial position as w1, but at the next coarser scale) [7]. Then:

y = w + n Where w = (w1, w2), y = (y1, y2) and n = (n1, n2). The noise values n1, n2 are zero-mean Gaussian with variance sigma (σ) [12], [13]. Based on the empirical histograms, the following non-Gaussian bivariate equation was used [12].

( ) ⎟⎟⎠

⎞⎜⎜⎝

⎛+−= 2

22

12

3exp.2

3 wwwp w σπσ (6)

With this equation, w1 and w2 are uncorrelated, but not independent [13]. The MAP estimator of w1 yields the following bivariate shrinkage function [1], [8].

122

21

22

22

1

1 .

3

ˆ yyy

yyw

n

+

+⎟⎟⎠

⎞⎜⎜⎝

⎛−+

=σσ

(7)

In general, the DTCWT has the following properties: • Good directional selectivity in 2-dimentions (also true for

higher dimensionality m-D); • Perfect reconstruction (PR) using short linear-phase filters; • Limited redundancy, independent of the number of scales,

2m:1 for m-D; Efficient order-N computation- only twice the simple DWT for 1-D (2m times for m-D).

III. NOISE LEVEL ESTIMATION

The effectiveness of video processing methods can be significantly reduced in the presence of noise. The level of noise can affect the performance of video denoising algorithms. Other video processing algorithms such as compression, segmentation, motion estimation and format conversion adapt their parameters and improve performance when the noise is known. Algorithms for estimating the AWGN variance are either temporal or spatial [5]. There exist few methods for purely temporal noise estimation such as [15]. These methods are challenged by the presence of object or global motion. Motion detection or motion compensation is commonly used as countermeasures. Hence, method in this area such as [15] requires more memory and is, in general, tends to be computationally expensive with estimation accuracy not significantly more precise than spatial methods [5]. For the spatial noise estimation method, the level of noise in a given digital frame is estimated from the noisy frame

data. From [2] a median value of the [ ].~0,hhλ subband is used in

H0a

H1a 2

2

H00a

H01a 2

2

H0b

H1b 2

2

H00b

H01b 2

2 x(n)

af1

af2

sf1

sf2

x(n) y(n)

c(n)

d(n)

409

the estimation process as follows. []

6745.0

).~(~ 0,hhn

Median λσ =

(8)

Where [ ].~0,hhλ are the noisy wavelet coefficients in the high-

high band of the 0th scale. Because the vast majority of useful information in the wavelet domain is confined to few and large coefficients, the median can effectively estimate the level of noise (i.e. the average level of the useless coefficients) without being adversely influenced by useful coefficients [2, 3].

It is worth to denote that the human visual system appears to have sensitivity thresholds. So, all frames with σ value less than 5dB look equally clean to human eye, but once the σ value exceeds 20 dB, the frame simply looks bad. This notation is very essential to us and will affect the decision making stage in the proposed system as will be discussed.

IV. PROPOSED VIDEO DENOISING SYSTEM

This system makes a tradeoff between time consumed in the denoising process and quality of the produced frames. To obtain high quality, we will consume more time in denoising, but in real-time applications, this is very hard to achieve [10]. Two spatial video denoising techniques will be used in the denoising process, which are 2D DWT and 2D DTCWT. The first of them gives good quality of the video at low levels of noise while consuming less time for denoising. The second achieves good quality of the video at high levels of noise but unfortunately consumes large time for denoising. Thus, in this proposed system we will make a compromise between quality and time through intelligent choice of the denoising technique according to the estimated noise level in the video frame. This technique is dedicated to and can be used in real-time applications. The noise in each frame will be estimated dynamically then the choice of the denoising technique is determined accordingly.

Previous techniques assumed that level of noise is known prior the noise removal process and did not provide the way to know it. However, Balster in [2] tried to estimate noise level, and Francois in [5] used hardware support to estimate noise. In this work, we will use the estimation method presented in [2] as it had the advantage of consuming less time in estimation and requiring less memory size making it suitable for the targeted real-time applications.

The framework of the proposed system contains three main components. A general description of the proposed video denoising system framework is presented in Fig. 5. The task of each component is as follows: Noise Estimator: This component estimates the noise level for each of the video frames. We make use of spatial image discontinuities, represented by wavelet coefficients, in order to estimate the level of standard deviation of estimated white Gaussian noise in the frame. Decision Making Component: In this component, the system chooses the suitable action towards the input frame according to the level of noise estimated. It can direct the frame to one of the two video denoising techniques in the next step (2D DWT or 2D DTCWT) or it considers the frame a clean frame. The system assumes that noise level that is less than 5db will be insignificant and unnoticeable to the human eye and will not perform the denoising thus

minimizing the total time of denoising of the video sequence. One of the two techniques will be applied if the noise level is above 5db. The first technique (2D DWT) will be chosen for low to moderate levels of noise as it will produce good denoised frames almost as those produced by the second technique (2D DTCWT) but in less time. The second technique will only be used at high levels of noises, as it gives better results. This means that, if the noise level is between 5 and 15db, then the frame contains low or moderate level of noise, and the 2D DWT denoising technique will be used. While if the noise level is over 15db, this means that the frame contains high level of noise and the 2D DTCWT denoising technique will be used. Denoising component: The system uses either the 2D DWT or the 2D DTCWT denoising technique according to the decision made at the previous component.

V. SIMULATION RESULTS

The performance of the proposed system is evaluated based on the PSNR (peak signal to noise ratio), which is the most commonly used quality metric, and on the overall time spent in the denoising process. PSNR is mathematically evaluated as follows: for a video sequence of K frames each having N×M pixels with m-bit depth, the Mean Square Error (MSE) is calculated as [11]:

∑ ∑= =

−

⎥⎥

⎦

⎤

⎢⎢

⎣

⎡−

×=

N

n

M

mjixjix

MNMSE

1 1

2

),(),(1 (9)

Where x(i,j) is the original frame and ),(_

jix is the restored frame of (i,j) location. The PSNR is calculated as:

)(

log.102

MSEmPSNR = (10)

Where m = 255. To determine the effectiveness of the proposed system, we

have to test it on both gray-scale and colored video streams to ensure that it can be used in any system. We have chosen two of the most famous videos in the field which are the grayscale ‘Akiyo‘ and colored ‘Salesman‘ video sequences which each consists of 50 frames. To have results to compare the proposed system with, we

Input Frame

s Wavelet Trans.

Noise Estimator

Denoising Using 2-D

DWT Clean Frame

In(t)

In(t)

Idwt(t) Idtcwt(t)

Decision Making

Compone

Output Frames

Denoising Using 2-D

DTCWT

12 3

5~ ≤nσYes No

15~ ≤nσ Yes No

3 2 1

nσ~Figure 5. General framework of the proposed algorithm.

In(t): input noisy frame; : standard deviation of estimated white Gaussian noise; Idwt(t): denoised frame using 2-D DWT; Idtcwt(t):

denoised frame using 2-D DTCWT.

410

applied the 2D DWT and the 2D DTCWT denoising techniques separately to the whole frames of the videos. The peak signal to noise ratio (PSNR) curves were used to measure the level of noise after denoising compared to the original image.

Fig.6-a shows the PSNR after applying 2D DWT, 2D DTCWT, and the proposed system to denoise the two video sequences. Fig.6-b shows the time consumed by DWT, DTCWT, and the proposed system and Table 1 summarizes the results obtained from the previous tests.

To further investigate the performance of the proposed system at different working environments, we considered two cases of video streams: mostly clean and mostly noised video streams. Mostly clean video are those videos that contain no or low levels of noise. While the mostly noised videos are those videos that contain high levels of noise.

To obtain mostly clean videos, we corrupted the two test videos at percentage of 90% of the whole video stream frames randomly with no or low levels noise. Fig.7-a shows the PSNR after applying DWT, DTCWT, and the proposed system to denoise the two videos. Fig.7-b shows the time consumed by DWT, DTCWT, and the proposed system and Table 2 summarizes the results obtained from the previous tests.

To obtain mostly noised videos, we corrupted the two test videos, at percentage of 90% of the whole video stream frames, randomly with high levels noise. Fig.8-a shows the PSNR after applying 2D DWT, 2D DTCWT, and the proposed system to denoise the two video sequences.

333435363738394041

Akiyo(Grayscale)

Salesman(Colored)

Video sequence

PSN

R (d

B)

DWT

DTCWT

ProposedSystem

(a) PSNR values obtained.

.

0

2

4

6

8

10

Akiyo(Grayscale)

Salesman(Colored)

Video Sequence

Tim

e C

onsu

med

(Sec

.)

DWT

DTCWT

ProposedSystem

(b) Time consumed.

Figure 7. Comparison between DWT, DTCWT and the proposed denoising system to denoise mostly clean grayscale and colored

sequence videos. Fig.8-b shows the time consumed in the DWT, the DTCWT, and the proposed system and table 3 summarizes the results obtained from the previous tests. From the figures and tables, we found many interesting remarks on the performance of the proposed systems, which can be summarized into points as follows: 1- Using the DTCWT denoising technique consumed more

time than the DWT, approximately 6 times more, for both the gray scale and colored videos. That can be explained by the high computations DTCWT takes during the denoising.

2- However, the DTCWT gives better average PSNR results than the DWT (the improvement in PSNR reaches 1dB in some cases) in both the gray scale and colored videos.

3- The performance of the Intelligent Denoising System is better than DWT in terms of the PSNR (quality of the video after denoising) as it gives better PSNR.

4- The Intelligent Denoising System consumes overall less time than using only DTCWT and more time than using only DWT, which means that it will consume less overall time to produce good overall quality. The mean value of PSNR of videos produced by the proposed system (40.2 in gray and 36.12 in colored) in the mostly clean videos is very close to those produced by DTCWT (40.6 in gray and 36.51 in colored) with a difference 0.98% and 1.07%. While the mean time consumed by the proposed system is much way less than the DTCWT (about 73.32% improvement in gray and 75.57% improved in colored).

313233343536373839

Akiyo(Grayscale)

Salesman(Colored)

Video Sequence

PSN

R (d

B)

DWT

DTCWT

ProposedSystem

(a) PSNR values obtained.

0

2

4

6

8

10

Akiyo(Grayscale)

Salesman(Colored)

Video Sequence

Tim

e C

onsu

med

(Sec

.)

DWT

DTCWT

ProposedSystem

(b) Time consumed.

Figure 6. Comparison between DWT, DTCWT and the proposed denoising system to denoise grayscale and colored sequence videos at random values

of σ.

TABLE I SUMMARY OF RESULTS OBTAINED FROM PREVIOUS TESTS FOR DIFFERENT

RANDOM VALUES OF σ.

Video Sequence

Technique Used

PSNR Time Consumed (sec)

min max mean min max mean

Akiyo (Grayscale)

DWT 32.83 51.20 37.94 0.73 1.11 0.82 DTCWT 33.51 51.76 38.61 2.97 3.15 3.08

Pro.System 34.18 48.84 38.16 0.31 3.43 1.48

Salesman (Colored)

DWT 29.84 46.78 33.99 2.25 5.04 2.35 DTCWT 30.46 47.53 34.55 8.97 9.34 9.21

Pro.System 30.45 46.19 34.35 0.71 10.18 4.64

TABLE II SUMMARY OF RESULTS OBTAINED FROM PREVIOUS TESTS

FOR MOSTLY CLEAN VIDEOS.

Video Sequence

Technique Used

PSNR Time Consumed (sec)

min max mean min max mean

Akiyo (Grayscale)

DWT 33.85 51.20 39.96 0.77 4.49 0.92 DTCWT 34.48 51.74 40.60 3.06 3.12 3.16

Pro.System 34.82 48.83 40.20 0.31 3.76 0.78

Salesman (Colored)

DWT 29.90 46.85 35.96 2.05 2.27 2.19 DTCWT 30.54 47.54 36.51 8.91 9.33 9.21

Pro.System 30.50 48.18 36.12 0.67 9.60 2.25

411

5- The mean value of PSNR of videos produced by the

proposed system (37.56 in gray and 33.63 in colored) in the mostly noised videos is very close to those produced by DTCWT (37.97 in gray and 33.89 in colored) with a difference of 1.08% and 0.87%. While the mean time consumed by the proposed system is much way less than both the DTCWT and DWT (about 75.24% improvement in gray and 48.81% improvement in colored from the DTCWT).

6- The minimum time consumed by the proposed system, which is produced when processing clean image containing noise less than 5db, is non zero as this time is consumed in noise estimation and making choice based on it.

From these points we can say that the proposed system performance is highly acceptable as it achieves high quality videos almost as good as the ones produced by DTCWT, which is considered the best spatial denoising technique available by this time, while consuming less time than it takes. Thus, it makes the proposed system most suitable for the real-time applications than both studied techniques.

VI. CONCLUSION

In this work, we were concerned with the noise removal from the real-time videos such as life broadcasts and video conferencing. In such environments, obtaining good quality videos in an acceptable time is very important, however, it is hard to achieve. The denoising techniques that produce high quality video take considerable time in the denoising process. We proposed an intelligent denoising system that makes a tradeoff between quality and time parameters of denoising.

The results produced by this system emphasis its ability to fulfill the real-time applications requirements in terms of producing high quality videos in appropriate time for these applications.

REFERENCES [1] A. Alin, and K. Ercan, "Image Denoising Using Bivariate-Stable

Distributions in the Complex Wavelet Domain", Signals and Images Laboratory, Istituto di Scienza e Tecnologie dell’Informazione “A. Faedo”, Area della Ricerca CNR di Pisa, Italy, 2004.

[2] E. J. Balster, Y. F. Zheng, and R. L. Ewing, "Combined Spatial and Temporal Domain Wavelet Shrinkage Algorithm for Video Denoising", IEEE Transactions On Circuits And Systems For Video Technology, Vol. 16, No. 2, February 2006.

[3] E. J. Balster, Y. F. Zheng, and R. L. Ewing, "Feature-Based Wavelet Shrinkage Algorithm for Image Denoising", IEEE transactions on image processing, vol. 14, no. 12, December 2005.

[4] I. Daubechies, "Ten Lectures on Wavelets", Rutgers University and AT&T Bell Labaratories, USA, 1992.

[5] L. Francois-Xavier, A. Amer, and C. Wang, “FPGA Architecture for Real-Time Video Noise Estimation", IEEE ICIP, 2006.

[6] M. Ghazal, A. Amer, and A. Ghrayeb, “A Real-Time Technique for Spatio–Temporal Video Noise Estimation”, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 17, No. 12, December 2007.

[7] R. Gomathi and S. Selvakumaran, " A Bivariate Shrinkage Function For Complex Dual Tree Dwt Based Image Denoising", Proceedings of the 6th WSEAS International Conference on Wavelet Analysis & Multirate Systems, Bucharest, Romania, October 16-18, 2006.

[8] A. Hyvarinen, P. Hoyer, and E. Oja, "Image Denoising by Sparse Code Shrinkage", Intelligent Signal Processing, IEEE Press, 2001.

[9] J. M. Lilly, and S. C. Olhede, "On the Design of Optimal Analytic Wavelets", IEEE Transactions On Signal Processing, February 2008.

[10] R. Orban, M. T. Faheem, A. Sarhan, “Comparison between Discrete Wavelet Transform and Dual-Tree Complex wavelet Transform in Video Sequences Using Wavelet-Domain”, INFOS 2008 ( The 6th International Conference on Informatics and Systems), Egypt, March 2008.

[11] C. Sang-Gyu, B. Zoran, M. Dragorad, L. Jungsik, and H. Jae-Jeong, "Image Quality Evaluation: JPEG 2000 versus Intra-only H.264/AVC High Profile", Facta Universitatis (NIˇS), Elec. vol. 20, no. 1, pp. 71-83, 2007.

[12] I.W. Selesnick, K. Y. Li, " Video Denoising Using 2D and 3D Dual-Tree Complex Wavelet Transforms ", Proceedings of SPIE , Vol. 5207, pp. 607-618, Nov 2003.

[13] L. Sendur, I.W. Selesnick, "A Bivariate Shrinkage Function For Wavelet-Based Denoising", Electrical Engineering, Polytechnic University, Metrotech Center, Brooklyn, NY 11201, 2001.

[14] L. Sendur, I.W. Selesnick, "Bivariate Shrinkage Functions for Wavelet-based Denoising Exploiting Interscale Dependency", IEEE Transactions on Signal Processing, 50(11), pp. 2744-2756, Nov 2002.

[15] B. C. Song and K. W. Chun, “Noise Power Estimation for Effective Denoising in a Video Encoder,” IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 2, pp. 357–360, March 2005.

[16] V. Zlokolica, "Advanced Nonlinear Methods for Video Denoising", Ph.D. Thesis, Faculty of Engineering, Ghent University, Germany, 2006.

TABLE III SUMMARY OF RESULTS OBTAINED FROM PREVIOUS TESTS

FOR MOSTLY NOISED VIDEOS

Video Sequence

Technique Used

PSNR Time Consumed (sec)

min max mean min max mean

Akiyo (Grayscale)

DWT 33.79 51.16 37.30 0.81 1.20 0.85 DTCWT 34.51 51.76 37.97 3.09 3.18 3.19

Pro.System 22.79 48.82 37.56 0.32 3.59 0.79

Salesman (Colored)

DWT 29.84 46.79 33.33 2.09 2.32 2.15 DTCWT 30.45 47.53 33.89 8.80 9.28 9.18

Pro.System 30.46 48.18 33.63 0.62 9.94 4.70

`

313233343536373839

Akiyo(Grayscale)

Salesman(Colored)

Video Sequence

PSNR

(dB)

DWT

DTCWT

ProposedSystem

(a) PSNR values obtained.

0

2

4

6

8

10

Akiyo(Grayscale)

Salesman(Colored)

Video Sequence

Tim

e C

onsu

med

(Sec

.)

DWT

DTCWT

ProposedSystem

(b) Time consumed.

Figure 8. Comparison between DWT, DTCWT and the proposed denoising system to denoise mostly noised grayscale and colored

sequence videos.

412