Real-Time Non-linear Noise Reduction Algorithm for Video

Chinatsu Mori and Seiichi Gohshi

Kogakuin University, 1-24-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo, 163-8677, Japan

Keywords:

Video Noise Reducer, 4KTV, 8KTV, Real Time, Non-linear Signal Processing, Image Quality.

Abstract:

Noise is an essential issue for images and videos. Recently, a range of high-sensitivity imaging devices

have become available. Cameras are often used under poor lighting conditions for security purposes or night

time news gathering. Videos shot under poor lighting conditions are afﬂicted by signiﬁcant noise which

degrades the image quality. The process of noise removal from videos is called noise reduction (NR). Although

many NR methods are proposed, they are complex and are proposed as computer simulations. In practical

applications, NR processing of videos occurs in real-time. The practical real-time methods are limited and the

complex NR methods cannot cope with real-time processing. Video has three dimensions: horizontal, vertical

and temporal. Since the temporal relation is stronger than that of horizontal and vertical, the conventional

real-time NR methods use the temporal relation to reduce noise. This approach is known as the inter-frame

relation, and the noise reducer comprises a temporal recursive ﬁlter. Temporal recursive ﬁlters are widely

used in digital TV sets to reduce the noise affecting images. Although the temporal recursive ﬁlter is a simple

algorithm, moving objects leave trails when it reduces the high-level noise. In this paper, a novel NR algorithm

that does not suffer from this trail issue and shows better performance than NR using temporal recursive ﬁlters

is proposed.

1 INTRODUCTION

Imaging technology advanced in the 21st century and

HDTV (1920 × 1080) resolution cameras have be-

come a reasonably priced commodity. Recently, high-

sensitivity imaging devices have also become widely

available and video cameras can work under poor

lighting conditions. This high-sensitivity imaging

technology makes 4K/8K ultra-high-resolution video

systems possible. The size of one 4K imaging pixel

is 1/4 that of an HDTV pixel and the size of one 8K

pixel is 1/16 that of an HDTV pixel. The light en-

ergy collected by one pixel is proportional to the size

of the imaging cell; therefore, the light energy col-

lected by one 4K or 8K pixel is 1/4 or 1/16 that of

an HDTV pixel. Since imaging cells generate a volt-

age that is proportional to the collected light energy,

4K/8K imaging cells generate a lower voltage than

those of HDTV imaging cells. The light intensity is

often insufﬁcient when 4K/8K imaging, which causes

noise to appear in videos, degrading the image qual-

ity.

Aside from 4K/8K videos, noise is also a crucial

issue in security cameras. Crimes are often commit-

ted after sunset. In the night time, the lighting condi-

tions are worse and the recorded videos usually con-

tain a lot of noise. When using recorded videos to

investigate a crime, noise is often a problem when try-

ing to identify the person of interest. Low noise and

high resolution, such as 4K/8K, are important factors

for high-quality videos. There are many signal pro-

cessing methods for reducing the noise and improv-

ing the resolution of recorded videos to achieve high-

quality videos.

Noise reduction (NR) is a signal processing

method for reducing noise in recorded videos, and

super-resolution (SR) is a signal processing method

for improving the video resolution. Unfortunately,

these two technologies are trade-offs. Noise occurs

as small dots that have high-frequency elements. The

high resolution is also created by high-frequency ele-

ments. If we try to reduce the noise in a video, the

high-frequency elements are reduced and the video

becomes blurry; this is the ﬁrst issue with NR. The

second issue is real-time signal processing, which is

essential for all video systems. Although there are

many NR approaches, most of them are proposed for

still images. There are no real-time requirements for

still-image NR. The frame rates of video systems are

50/60 (analogue TV/HD/4K) or 120 Hz (8K). This

means that the NR processing for a frame has to be

ﬁnished within 25/16.7 ms for current practical video

Mori, C. and Gohshi, S.

Real-Time Non-linear Noise Reduction Algorithm for Video.

DOI: 10.5220/0006839803210327

In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications (ICETE 2018) - Volume 1: DCNET, ICE-B, OPTICS, SIGMAP and WINSYS, pages 321-327

ISBN: 978-989-758-319-3

321

systems (analogue TV/HD/4K). Due to these time

constraints, it is impossible to adopt as complex NR

algorithms for videos as used for still images. In this

paper, a novel real-time NR algorithm for videos is

proposed. It exploits video characteristics that are dif-

ferent from those of still images and reduces noise

without blurring, unlike conventional NR algorithms

for videos.

2 PREVIOUS WORKS

Still images have horizontal and vertical (spatial)

axes, but videos havespatial and temporal axes. Many

two- (2D) and three-dimensional (3D) NR systems

have been proposed by researchers.

In (Malfait and Roose, 1997), (Kazubek, 2003),

(Piurica et al., 2004), (Nai-Xiang et al., 2006), (Se-

lesnick and Li, 2003), and (Pizurica et al., 2003), the

authors proposed the use of spatial (two-dimensional)

and spatiotemporal (three-dimensional) ﬁlters to re-

move video noise. However, spatial ﬁlters only con-

sider spatial information; therefore, these ﬁlters can

cause spatial blurring at high noise levels. Using

a combination of temporal and spatial information

can reduce this blurring effect. This approach can

also be used to improve the ﬁltering performance at

low noise levels. In (Malfait and Roose, 1997), a

wavelet domain spatial ﬁlter in which the coefﬁcients

are manipulated using a Markov random ﬁeld im-

age model has been proposed. A Wiener ﬁlter was

utilized in the wavelet domain to remove the image

noise in (Kazubek, 2003). Noise reduction using

the wavelet transform was proposed in previous stud-

ies (Piurica et al., 2004)(Nai-Xiang et al., 2006)(Se-

lesnick and Li, 2003)(Pizurica et al., 2003)(Gupta

et al., 2004)(Mahmoud and Faheem, 2008)(Jovanov

et al., 2009)(Luisier et al., 2010), which provides a

high performanceand results in images of a high qual-

ity. However, currently, these approaches are only

feasible at the computer simulation level, and they

do not work in real time. The wavelet transform is

a complex algorithm; therefore, it is difﬁcult to ap-

ply it to NR, and it is not cost-effective. Currently,

there are no practical real-time NR systems employ-

ing the wavelet transform method. The authors in (Pi-

urica et al., 2004) proposed a fuzzy logic-based image

noise ﬁlter that considers directional deviations.

In addition, a recursive estimator structure has

been proposed to differentiate a clean image from a

ﬁlm-grain noisy image where the noise is considered

to be related to the exposure time in the form of a

non-Gaussian and multiplicativestructure (Nai-Xiang

et al., 2006). In addition, a pixel-based spatiotemporal

adaptive ﬁlter that calculates new pixel values adap-

tively using the weighted mean of pixels over mo-

tion compensated frames has been proposed in (Se-

lesnick and Li, 2003). An edge preserving spatiotem-

poral video noise ﬁlter that combines 2D Wiener and

Kalman ﬁlters has been presented in (Pizurica et al.,

2003). The authors of (Gupta et al., 2004) proposed

a nonlinear video noise ﬁlter that calculates new pixel

values using a 3D window. In this method, the pixels

are arranged with respect to the related pixel values

in the form of a 3D window according to their differ-

ence and the average of the pixels in the window af-

ter weighting them with respect to their sorting order,

which gives good results in the case of no or slow lo-

cal motion, but it deforms image regions in the case of

abrupt local motion. For local motion, the 3D ﬁltering

performance of this method is low. To improvethe 3D

ﬁltering performance of the method proposed in (Nai-

Xiang et al., 2006), video de-noising uses 2D and 3D

dual-tree complex wavelet transforms. The authors of

(9) proposed 2D wavelet-based ﬁltering and temporal

mean ﬁltering that uses pixel-based motion detection.

The authors in (Mahmoud and Faheem, 2008) pro-

posed 2D wavelet-based ﬁltering and temporal mean

ﬁltering that uses pixel-based motion detection. The

authors in reference (Jovanov et al., 2009) proposed

a wavelet transform-based video ﬁltering technique

that uses spatial and temporal redundancy. A content

adaptive video de-noising ﬁlter was also proposed re-

cently (Luisier et al., 2010).This method ﬁlters both

impulsive and non-impulsive noises, but the ﬁltering

performance is low in cases with Gaussian noise with

high variance. In this work, a new pixel-based spa-

tiotemporal video noise ﬁlter that incorporates motion

changes and spatial standard deviations into the de-

noising algorithm is proposed.

Bilateral ﬁltering has also been proposed for NR

(Yang et al., 2009). Although it is a simple algorithm,

in principle, it could cause spatial blurring in station-

ary areas. Our eyes are sensitive to blurring in station-

ary areas than in moving areas. Stopping the video

signal, we perceive a large blur in the moving areas.

However, when playing the same video again, you

cannot ﬁnd the same blur. The reason for this is that

our dynamic eyesihgt is inferior to the static eyesight.

Since NRs employing recursivetemporal ﬁlters do not

cause spatial blurring but cause blurring in moving ar-

eas, they give the perception of a better image qual-

ity. Many other proposals have been made to reduce

noise in images and videos (Dabov et al., 2007)(Le-

brun et al., 2013)(Portilla et al., 2003)(Kaur et al.,

2002)(Rudin and S.Osher, 1992)(Elad and Aharon,

2006). However, none of them are sufﬁciently fast

for their use with real-time videos.

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

322

Videos have a strong correlation along the tempo-

ral axis compared to the horizontal and vertical axes.

This characteristic has been used to reduce noise in

videos. Conventional real-time NR algorithms use

temporal correlation to reduce noise (Kondo et al.,

1994)(Brailean et al., 1995)(Yagi et al., 2004). Frame

memory is required to exploit temporal correlation.

This is called inter-frame signal processing. Although

the memory cost has been reduced, the overall cost is

still high if we use it for many frames. Traditionally, a

recursive temporal ﬁlter with one frame memory was

used in this conﬁguration. Most of digital TV sets are

equipped with the recursive temporal type noise re-

ducers. However, there is an issue in the motion areas

which have blur trails becuse the recusive ﬁlters have

inﬁnite responses. This issue is discussed in the next

section.

Figure 1: A conventional real-time video noise reducer.

Figure 2: A noisy video frame.

Figure 3: The processed result of Figure 2 using the con-

ventional method (Figure 1).

3 ISSUES WITH THE

CONVENTIONAL METHOD

A block diagram of a real-time noise reducer is shown

in Figure 1. The parameter, α, is set based on a range

0:1 (low level noise) to 0:3 (high level noise). It re-

duces the pixel value changes using a temporal re-

cursive low pass ﬁlter at every pixel. Currently, only

this type of noise reducer is practical since it is cost-

effective. It can work in real time and is commonly

used in TV systems (TI, 2011). AAs mentioned ear-

lier in this section, the stationary areas have the same

pixel values. However, the pixels in the moving areas

change their values in every frame. Although con-

ventional noise reducers successfully reduce noise in

stationary areas by averaging the values of each pixel,

they create a motion trail blur behind the moving ob-

jects.

Figure 2 shows a frame from a noisy video. In this

video, the camera is panning from left to right. Fig-

ure 3 shows the processed result using NR shown in

Figure 1. Although the noise is reduced in Figure 3,

there is a trail from left to right in accordance with

the camera panning direction. If the noise is high and

visible, the recursive ﬁlter in the noise reducer has to

work more heavily, i.e. with a larger recursive coefﬁ-

cient (α in Figure 1). The larger recursive coefﬁcient

reduces the noise. However, as shown in Figure 3, it

also causes blur in the moving areas. This type of NR

involves a trade-off between the strength of NR and

the extent of blurry trails.

A video signal can be written as f(x, y, t). Here,

x is the horizontal axis, y is the vertical axis and t is

the temporal axis. We assume noise as n(t), and the

video with noise can be expressed as follows:

(x, y, t) = f(x, y, t) + n(t) (1)

The noise reduction process of the conventional

method shown in Figure 1 can be expressed as fol-

lows:

Fn(x, y, t) = (1− α) f

(x, y, t − 1) + αf

(x, y, t) (2)

The spatial position (x,y) is the same in all frames

and only the temporal parameter t changes. There-

fore, Equation 2 can be written as follows:

Fn(t) = (1 − α) f

(t − 1) + αf

(t) (3)

Equation 3 is a recursive ﬁlter that has an inﬁnite im-

pulse response (IIR). Theoretically, IIR leads to in-

ﬁnite trails in movement areas. In the real video,

the trails continue until the output of the IIR ﬁlter

becomes smaller than the least signiﬁcant bit (LSB)

level. A temporal ﬁnite impulse ﬁlter (FIR) does not

cause the long trails associated with IIR. However,

a couple of frames of temporal relation cannot re-

duce noise to the practically required level. If we

increase the number of frames in memory, blur/trail

occurs. The spatial processing (intra-frame) NR does

not cause trails or blur. It does not work well because

Real-Time Non-linear Noise Reduction Algorithm for Video

323

the spatial correlation is not strong compared with

the temporal relation in images/videos. The spatial

NR causes a spatial blur instead of the temporal blur

that is caused by temporal recursive NR. The conven-

tional NR is a kind of low pass ﬁlter (LPF). Noise

in videos looks like it comprises high-frequency el-

ements. However, noise comprises a wide range of

frequencies, including low-frequency elements and

DC. NR works as an LPF against noise which elimi-

nates the high-frequency elements while retaining the

low-frequency elements. Although the peak level of

noise decreases, the noise changes its shape and be-

comes low-level widespread spots. Since human eyes

are sensitive to the low-frequency elements, the fre-

quency shifted low-level noise becomes more visible.

This means that conventional NR changes the noise

shape and makes it more visible.

Figure 4: Proposed NR signal prosessing.

4 PROPOSED METHOD

Figure 4 shows an image comparison of the conven-

tional and the proposed NR. In Figure 4, the hori-

zontal axis is the horizontal/vertical line of the video

and the vertical axis is the level of the video. Fig-

ure 4(a) is the input of the NR ﬁlter, which is a video

with noise. Figure 4(b) is the conventional NR pro-

cessed result of Figure 4(a). As discussed in the pre-

vious section, the levels of noise are reduced but be-

come widespread, as shown in Figure 4(b). In Fig-

Figure 5: Block diagram of proposed NR signal prosessing.

Figure 6: Processed result of Fig. 2 by the proposed

method.

ure 4(b), the levels of noise are lower than those in

Figure 4(a) after the application of the LPF. How-

ever, the noise spreads over wider areas than that in

Figure 4(a). Noise becomes more visible with LPF

especially for high-noise videos that are shot under

poor lighting conditions. When these kinds of videos

are processed by the conventional NR equipped with

LPF, the low-level widespread noise appears every-

where. The conventional noise reducer changes the

noise frequency from high to low, which makes the

noise more visible. If the noise is converted to the

high-frequency areas , it becomes less visible. Fig-

ure 4(c) shows the processed result obtained by us-

ing the proposed method. In Figure 4(c), the levels

of noise become lower but the noise does not spread.

The ends of the noise become sharp edges that con-

tain high-frequency elements. Therefore, the noise is

successfully converted to the high-frequency areas.

We propose a novel nonlinear FIR for NR. Here,

we assume f(x, y, t− 1), f(x, y, t), and f(x, y, t +1) are

three sequential frames. The target frame for process-

ing is f(x, y, t). f(x, y, t − 1) and f (x, y, t + 1) are the

reference frames. We also assume the noise in the

video is Gaussian noise with deviation because it is

the most common noise for images created under poor

lighting conditions. Noise is the undesired signal. If

f(x, y, t) contains noise, the level of f(x, y, t) is higher

or lower compared with the true value. However, al-

though the video contains noise, f(x, y, t) may be the

true value. The proposed method changes the value of

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

324

f(x, y, t) according to the following three cases, which

occur depending on their probability.

• f(x, y, t − 1) ≤ f(x, y, t) ≤ f(x, y, t + 1) or

f(x, y, t + 1) ≤ f(x, y, t) ≤ f(x, y, t − 1) =⇒ the

output of the NR is f(x, y, t)

• f(x, y, t) is the highest=⇒ the output of the NR is

f(x, y, t) − δ

• f(x, y, t) is the lowest=⇒ the output of the NR is

f(x, y, t) + δ

Condition 1.: if f(x,y,t) is in the middle, f(x,y,t) does

not contain noise and no signal processing is neces-

sary for f(x,y,t). The output of the NR is f(x,y,t). Con-

dition 2.: if f(x,y,t) is the highest of the three sig-

nals, f(x,y,t)- is the output of the NR. Condition 3.:

if f(x,y,t) is the lowest of the three signals, f(x,y,t)+ is

the output of the NR. A block diagram of the proposed

signal processing is shown in Figure 5. The proposed

NR comprises two frame memories, one comparer,

one adder, one subtracter, and one selector. The com-

parer has three inputs. It compares f(x,y,t) with the

other two signals, f(x,y,t-1) and f(x,y,t+1). The output

of the comparer is three bits, which represent three

conditions: f(x,y,t) is the highest, f(x,y,t) is in the

middle, and f(x,y,t) is the lowest of the three values

. These three bits are introduced to the selector. This

approach is sufﬁciently simple to embody as a real-

time noise reducer.

In Figure 5, the top left is the video input of the

NR ﬁlter and the bottom right is the output of NR

ﬁlter. f(x,y,t-1), f(x,y,t), and f(x,y,t+1) are obtained

with the two frame memories. By comparing f(x,y,t)

with the other two values, the order of f(x,y,t) is ob-

tained. If the value of f(x,y,t) is in the middle (case 1),

f(x,y,t) is the output of the NR. If f(x,y,t) is the high-

est, f(x,y,t)- is the output of NR (case 2). If f(x,y,t) is

the lowest, f(x,y,t)+ is the output of NR. f(x,y,t)- and

f(x,y,t)+ are created by the adder and the subtracter.

The three paths, f(x,y,t), f(x,y,t)-, and f(x,y,t)+, are the

inputs of the selector, and one of them is selected as

the output of the comparer. The block diagram shown

in Figure 5 indicates practical hardware that could im-

plement the proposed algorithm. It is a simple and

compact design for the development of real-time NR

hardware.

5 EXPERIMENT

5.1 Simulation Results

Computer simulations were conducted to compare the

peak signal-to-noise ratios (PSNRs) of the proposed

and conventional NR methods. Figure 7 shows stills

from ﬁve video sequences. In Figures 7(a) and (e), the

train and marching people are moving and the cam-

era is panning slowly. In Figure 1(b), the camera was

moved using a circular dolly, whereas in Figure 7(d),

it was dollied in and then zoomed back. The woman

stood at the same place in both sequences and did not

move signiﬁcantly. Figure7(c) shows a music concert

with ﬂashing lights and confetti.

We preparedtest video sequences by adding Gaus-

sian noise ( = 7) to Figures 7(a)(e). We then compared

the PSNRs of the proposed NR method with those of

conventional NR using computer simulations. Ow-

ing to space limitations, only the results for Figures

7(a) and (b) are presented in Figures 8 and 9. Herein,

the horizontal axis shows the frame number, and the

vertical axis shows the PSNR. The blue lines show

the PSNRs for the videos with added noise compared

with the original videos. These stay constant because

a constant level of noise ( = 7) was added. The yellow

green and purple lines show the results of process-

ing the videos with conventional NR using parame-

ters = 0.2 and 0.5, respectively, whereas the brown

lines show the results of processing the videos using

the proposed method.

Although the conventional NR method reduced

the noise in the videos, its PSNRs are lower than those

for the noisy test videos. This means that it reduced

noise and degraded the resolution. In contrast, the

proposed method (brown lines) always yields PSNRs

higher than those of the noisy test videos, as shown in

both Figures 8 and 9. These results indicate that the

proposed NR method outperforms conventional NR.

5.2 Low Luminace Video

Figure 6 shows the processed result of Figure 2 using

the proposed method three times sequentially. Com-

paring Figure 3 with Figure 6, the image quality of

Figure 6 is better than that of Figure 3. Blur in Figure

6 is less than that in Figure 3 and noise is greatly re-

duced. Note that Figures 2, 3, and 6 are just computer

simulation results.

We apply the proposed method to an actual video.

Figure 10 shows a video frame shot under 3.5 lx il-

lumination by a high-sensitivity video camera. Al-

though 3.5 lx illumination is not sufﬁcient for imag-

ing, noise is not visible. In the video, the doll is rotat-

ing and the hair ornament is curving due to centrifugal

force. Figure 11 shows a video frame shot under 0.4

lx illumination taken by the same video camera. Even

though a high-sensitivity camera is used, noise is visi-

ble everywhere. Figure 12 shows the processed result

of Fig. 10 by the proposed method. Comparing Fig-

Real-Time Non-linear Noise Reduction Algorithm for Video

325

Figure 7: Video sequences.

Figure 8: Simulation results for the sequnce in Figure 7(a).

Figure 9: Simulation results for the sequnce in Figure 7(b).

ure 12 with Figure 11, noise is reduced and there is no

motion blur, which is apparent in Figure 3. In partic-

ular, the moving thin hair ornament that is curved due

to the motion is not blurry.

It should be noted that the proposed method does

not cause any blur in moving areas, unlike the conven-

tional NR. As shown in Figure 6 , the proposed NR

algorithm is simple, cost-effective, and can process

videos in real time. However, the noise levels differ

Figure 10: Image shot under 3.5 lx illumination.

Figure 11: Image shot under 0.4 lx illumination.

Figure 12: Processed result of Fig. 7 by the proposed

method.

depending on the video. It is necessary to precisely

detect the noise level to make the NR work in real

time. Future work will focus on developing a method

to detect the noise level automatically. Combining the

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

326

proposed NR and an automatic noise level detector

can reduce video noise effectively without human in-

tervention.

6 CONCLUSION

A novel NR algorithm that can process videos in real

time is was proposed. It does not sufferhave from

the artifactsartefacts that the afﬂict conventional NR

algorithmshas, such as trails behind moving objects

trails. The Computer simulations of a video with

added noise and shot in a dark room are were pre-

sented with the noise added video and the shot in a

dark room. Although the proposed NR is composed

with comprises a simple algorithm, it can remove

video noise effectively. Therefore, iIt is should not

be difﬁcult to develop a real- time hardware based on

the proposed ideasmethod. The Future work should

focus on how to detect the noise level automatically.

REFERENCES

Brailean, J. C., Kleihorst, R. P., Efstratiadis, S., Katsagge-

los, A. K., and Lagenfdijk, R. L. (1995). Noise re-

duction ﬁlter for dynamic image sequeces : Review.

Proceedings of the IEEE, 83:1272–1292.

Dabov, K., Foi, A., Katkovnik, V., and Egiazarian, K.

(2007). Image denoising by sparse 3-d transform-

domain collaborative ﬁltering. IEEE Transactions on

Image Processing, 16(8):2080–2095.

Elad, M. and Aharon, M. (2006). Image denoising via

sparse and redundant representations over learned dic-

tionaries. IEEE Transactions on Image Processing,

15(12):3736–3745.

Gupta, N., Swamy, M. N., and Plotkin, E. (2004). Low-

complexity video noise reduction in wavelet domain.

IEEE 6th Workshop on Multimedia Signal Processing,

pages 239–242.

Jovanov, L., Pizurica, A., Schulta, S., Schelkon, P.,

Munteanu, A., Kerre, E., and Philips, W. (2009).

Combined wavelet domain and motion compensated

video denoiding based on video codec motion estima-

tion method. IEEE Transactions on Circuits and Sys-

tems for Video Technology, 19(3).

Kaur, L., Gupta, S., and Chauhan, R. (2002). Image denois-

ing using wavelet thresholdingh, indian conference on

computer vision. Graphics and Image Processing.

Kazubek, M. (2003). Wavelet domain image denoising by

thresholding and wiener ﬁltering. IEEE Signal Pro-

cessing, 10(11):324–326.

Kondo, T., Fujimori, Y., Horishi, T., and Nishikata, T.

(1994). Patent: Noise reduction in image signals: Pct,

ep0640908 a1, ep19940306328.

Lebrun, M., Buades, A., and Morel, J.-M. (2013). A nonlo-

cal bayesian image denoising algorithm. SIAM Jour-

nal on Imaging Science, 6(3):1665–1688.

Luisier, F., Blue, T., and Unser, M. (2010). Surelet for

orthonormal wavelet domain video denoising. IEEE

Transactions on Circuits and Systems for Video Tech-

nology, 20(6).

Mahmoud, R. O. and Faheem, M. T. (2008). gcomparison

between dwt and dual tree complex wavelet transform

in video sequences using wavelet domain. INFOS.

Malfait, M. and Roose, D. (1997). Wavelet based image de-

noising using a markov random ﬁeld a priori model.

IEEE Transaction on Image Processing, 6(4):549–

565.

Nai-Xiang, L., Vitali, Z., and Yap, T. P. (2006). Video

denoising using vector estimation of wavelet coefﬁ-

cients. ISPACS.

Piurica, A., Zlokolica, V., and Philips, W. (2004). Noise re-

duction in video sequences using wavelet-domain and

temporal ﬁltering. Wavelet Applications in Industrial

Processing, Edited by Truchetet, Frederic. Proceed-

ings of the SPIE, 5266:48–49.

Pizurica, A., Zlokolica, V., and Philips, W. (2003). Com-

bined wavelet domain and temporal video denois-

ing. IEEE Conference on Advanced Video and Signal

Based Surveillance (AVSS2003), pages 334–341.

Portilla, J., Strela, V., Wainwright, M. J., and Simoncelli,

E. P. (2003). Image denoising using scale mixtures of

gaussians in the wavelet domain. IEEE Transactions

on Image Processing, 12(11):1338–1351.

Rudin, L. and S.Osher (1992). Nonlinear total variation

based noise removal algorithms. Physica D, 60(1-

4):259–268.

Selesnick, I. W. and Li, K. Y. (2003). Video denoising using

2d and 3d dual-tree complex wavelet transforms. In

Wavelet Applications in Signal and Image Processing

(SPIE 5207), 5207:607–618.

TI (2011). TVP5160 3D Noise Reduction Calibration Pro-

cedure Application Report: SLEA110 May, Texas In-

strument Manual.

Yagi, S., Inoue, S., Hayashi, M., Okui, S., and Gohshi, S.

(2004). Practical Video Signal Processing, pp. 143-

145, ISBN4-274-94637-1,. Ohmusha (in Japanese).

Yang, Q. X., Tan, K. I., and Ahuja, N. (2009). Real time

o(1) bilateral ﬁltering. Computer Vision and Pattern

Recognition (CVPR).

Real-Time Non-linear Noise Reduction Algorithm for Video

327