>Paper # 13< 1

Abstract—Reliably transporting captured video content over Wireless Sensor Networks (WSNs) poses several challenges due to the unique energy constraints presented by individual sensor nodes. The nature and amount of video content is another challenge that further complicates WSN data reliability problem. Motivated by the aforementioned factors, in this work, we propose an architecture that promises to enhance received video quality along-with sensor network lifetime maximization using a three prong strategy: splitting source coded video into prioritized streams, use path diversity to route video packets and partially and progressively decode the data packets as they traverse the multi-hop network. Further intrigued by quality enhancement for received video, we apply our proposed framework to an actual video sequence and compare the benefits that can be achieved using multipath multi-stream distributed reliability (MMDR) framework that we propose. We show that with little bit of processing within the network, significant enhancements in quality of video can be achieved for same level of energy spent to deliver video over the end-to-end path. We propose selective budgeting of network energy resources for processing within the network. We use low density parity check (LDPC) codes for channel coding the video sensor data. We show that the proposed architecture outperforms prevalent end-to-end channel coding schemes by considerable margins. Further, based on the results achieved, we discuss the associated rate-distortion tradeoffs in design of video sensor networks.

Index Terms—Wireless sensor networks, distributed error recovery, path diversity, reliable video

I. INTRODUCTION & MOTIVATION

IDE scale availability of low-cost hardware such as CMOS cameras and microphones has fostered the

development of sensor networks to carry multimedia content [13]. For instance, Crossbow’s Imote2 multimedia board [14] adds multimedia capabilities to the Imote2 platform for Wireless Sensor Networks (WSNs) allowing for capturing images, video as well as audio for playback. This opens door for a variety of applications for WSNs along with posing certain challenges in terms of reliably routing the multimedia content from sensor nodes back to the base station. An example of such an application can be in surveillance and

Manuscript received January 7th, 2009. This work was supported in part by NSF Award CNS-0721550, NSF Award CCF 0728996, NSF Award CCF-0515253, and NSF Award CNS-0430436.

Saad B. Qaisar is PhD Candidate with the Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48824 USA (phone: 517-355-3769; fax: 517-353-1980; e-mail: qaisarsa@ egr.msu.edu).

Hayder Radha is Professor and Associate Chair (Research) with the Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48824 USA (e-mail: radha@egr.msu.edu).

security related scenarios where images or video data captured by a remotely deployed camera mounted on a sensor node might need to be reliably sent back to the central base station. An inherent problem in applications involving video data is the amount of data generated along with ensuring minimal expenditure of energy in routing it back to the base station. Using a single path to the destination would fast deplete resources on that path, resulting in early network partitions and lesser than expected network lifetime.

In [16], Puri et al. discuss the challenges faced by any video transmission scheme for wireless sensor networks. They conclude that a broadband network of wireless video sensors is subjected to three principal constraints: limited processing capabilities, limited power/energy budget and information loss endemic to the harsh-loss prone wireless communication environment. A related issue is the high transmission costs for energy ravenous sensor nodes. According to some studies, there is a 1:2000 ratio between transmission cost and computational cost [21] in WSNs, thus making a strong case in favor of in-network processing. Any data reliability framework should keep these limitations under consideration. These limitations and the information losses call for robust coding algorithms, protocols and architectures that build error robustness within the WNSs.

One way to add reliability to sensor data is through end-to-end channel coding where WSN data is channel coded to sustain itself over entire end-to-end multi-hop path. End-to-end channel coding proves too costly for energy constrained WSNs [3]. Given the sensor transmission costs, retransmission based schemes may not be feasible [4]. Thus, motivated by both error prone nature of links as well as sensor energy limitations, we provide an architecture that can be used in order to reliably transport video content in WSNs. The proposed framework ensures near optimal use of network resources as well as network lifetime maximization by avoiding partitions in the network due to over usage of few paths. Particularly, we propose a strategy to add reliability in WSNs using a three prong approach that builds on: a) multi-stream coding of video b) multipath transport and c) distributed progressive error recovery through partial recovery of sensor data at intermediate nodes (discussed later).To the best of our knowledge, no work exists in this regard with these three components working together in WSNs. We propose a solution that is best suited to multi-stream coding of video as prevalent in most modern video coding architectures [16]. The proposed architecture is distributed in nature such that no complete knowledge is assumed regarding end-to-end network conditions. Using LDPC codes, we compare the performance of the proposed framework with end-to-end channel coding

Multipath Multi-stream Distributed Reliable Video Delivery in Wireless Sensor Networks

Saad B. Qaisar and Hayder Radha, Senior Member, IEEE

W

978-1-4244-2734-5/09/$25.00 ©2009 IEEE 207

>Paper # 13< 2

and show that it significantly outperforms it in terms of energy efficiency for level of reliability thus achieved.

Rest of the paper is organized as follows. In section II, we provide background to various techniques used in this paper. We present the proposed architecture for reliable video transmission in WSNs in section III. Section IV gives the simulation setup for a sample video application. We present results and related discussion for the proposed framework in section V. We conclude in section VI.

II. BACKGROUND

In this section, we provide a quick overview to few of the concepts used in the proposed framework.

A. Video Coding Techniques In order to understand the WSN video reliability problem,

we take a quick glance on the nature of coding techniques available for coding video content. Possible candidates for WSN can be adapted form of either of layered coding (LC), multiple description coding (MDC) or distributed video coding (DVC).

A typical layered coder includes one base layer and one or more successive enhancements layers that can be used together to achieve a desired level of video resolution at the destination. The base layer has higher priority than the enhancement layers as the loss of base layer makes the information received from the enhancement layers useless [17].

Multiple description coding fragments a single media stream into M independent sub streams ( 2M ) referred to as descriptions. These descriptions are then sent over multiple channels and combined at the receiver depending on the desired video resolution [7].

Predictive video coding (PVC) as used in H.26x or MPEG-x employs two modes in order to encode video: Intra-coding (I) mode exploiting the spatial correlation in the frame that contains the current block by using a block transform such as the discrete cosine transform (DCT), Inter-coding or motion compensated predictive (P) mode that exploits both spatial and temporal correlation in the video sequence [16]. Thus, in PVC, the encoded video frames can be partitioned into an I-stream and a P-stream.

A more recent work on low complexity encoding of sensor video data is characterized by distributed video coding (DVC) which incorporates concepts from source coding with side information, creating an intra-coded I-frame along with a side information counterpart Wyner-Ziv (WZ) frame [16]. The DVC framework, using its PRISM architecture, promises to provide robustness, light source-encoder architecture and flexibility in distributing the computational burden of motion estimation between transmitter and receiver. Without going into further details, we can notice here as well that data can be partitioned into multiple streams consisting both intra-coded as well as side information frames.

Figure 1. A Wireless Sensor Network snapshot with multiple paths to destination node

In summary, video coding techniques currently in vogue create multiple streams of data that may differ in their significance. Any video sensor reliability framework should consider these facts in the design of the technique.

B. Multipath Transport (MPT) MPT has been studied at length for both wired and wireless

data [9]. Multipath routing focuses on finding maximally disjoint paths from source to destination node. Various routing algorithms used in sensor networks such as directed diffusion, dynamic source routing (DSR), ad-hoc on demand distance vector (AODV) return multiple paths to the destination node [9]. Figure 1. shows a WSN with multiple paths leading from source node to destination. Routing diversity—where packets are purposely sent on different routes to insure against the failure of a single route—can increase error robustness in WSNs [2]. Thus, sensor energy limitations and multi-stream nature of video content generated makes an appealing case in favor of MPT for avoiding early network partitions through load balancing and provision of error resilience.

III. MULTI-STREAM MULTIPATH DISTRIBUTED RELIABILITY

In section II, we discussed various video coding techniques that can be potentially adapted for use with WSNs. One thing is common in all of them, data can be partitioned into various streams. These streams can either be self sufficient in that only one of them can be used to decode the information (such as in MDC) or hierarchical in the sense that we need one of them in order to decode the second one. We propose a multi-stream multipath distributed reliability (MMDR) framework that is generic in nature such that any technique used for encoding, till the time it has streams, we are good. The framework uses a three prong strategy:

978-1-4244-2734-5/09/$25.00 ©2009 IEEE 208

>Paper # 13< 3

Source encode the video data into multiple streams and channel code the resulting streams Assign the video to multiple paths for ensuring load balancing and error resilience Use partial decoding at intermediate nodes through distributed progressive error recovery framework (discussed subsequently) to recover from channel induced errors

Figure 2 gives the details of the proposed MMDR setup in WSNs. In the initial phase, the video data is source encoded such that we attain multiple streams. These streams are then channel coded before assigning to various paths. In the subsequent subsections, we present the details on the components of the proposed architecture.

A. Prioritized Data Partitioning The source encoded video is partitioned into M streams.

based on the significance of data towards decoding of video. e.g. assuming that we are using H.264 [17] for coding video, the encoded video frames are split based on whether they are intra-coded (I-frames) or inter-coded (P-frames). These streams are then channel encoded before they are fed to a path assignment block.

B. Path Assignment We assume that certain multipath routing scheme such as

directed diffusion is already working underneath that returns multiple paths to the destination [8] [11]. These schemes also return information on the level of reliability for each path. Based on this information, K paths are chosen and partitioned M streams of data are assigned to selected paths. A special case in MMDR is when K M during path assignment phase. In such a scenario, packets from multiple streams may share the same path and intermediate nodes might need information on significance of packets. Thus, any protocol designed based on MMDR should build that capacity within the packet streams. We restrict our discussion here for cases when

K M .For the example H.264 data, since intra-coded I frames

contain greater amount of data than their P counterparts, they may require further splitting over further paths for better load balancing, though, P frame streams may also requiring be split, depending on number of paths available for reaching back to destination and group of pictures (GOP) size used for encoding video. Intelligent data partitioning schemes are thus required that partition the frames at their boundaries such that despite losses in datasets for one frame other frames remain unaffected. For the scope of this work, we assume that is the case.

C. Distributed Progressive Error Recovery In [6], we presented a distributed progressive error recovery

algorithm (D-PERA) that provides reliability to sensor data through partial decoding of LDPC encoded packets at intermediate nodes. Under D-PERA, intermediate nodes over the end-to-end path carries out finite decoding less than that required for complete decoding, before forwarding the information to next hop. A relevant question is then how much processing each sensor node should undertake that maximizes the net throughput at destination node without fast depleting its energy resources. D-PERA addresses this question and achieves significantly better performance than conventional end-to-end channel coding schemes [6].

Using LDPC codes, D-PERA employs a statistical model for variation of error rate with number of decoding iterations first presented by Qaisar et al. in [5]. Under the proposed statistical model, the expected bit error rate in an LDPC encoded information packet received from a channel with error probability , after l decoding iterations is given as :

( ) ( )( , ) ( ) ( )ind indl lind ind indf l e e (1)

where min max , max0 l l , ind is the corresponding index value with respect to min and ( )ind , ( )ind ,

( )ind and ( )ind are the statistical coefficients [5].

Figure 2. System architecture for multistream multipath distributed reliability framework for video delivery in WSNs

978-1-4244-2734-5/09/$25.00 ©2009 IEEE 209

>Paper # 13< 4

Figure 3. Pairing of nodes in a multi-hop multi-path wireless sensor network with associated channel error probabilities

Nodes collaborate, in a pair-wise fashion, in order to distribute processing budget over the end-to-end path. Nodes pair in Lead-Follower relationship with each Lead node having the responsibility to undertake iteration assignment. The lead node asks its follower for its associated channel error probability F using which it estimates the iteration assignment for the follower. Follower nodes periodically report their associated channel probability F to the Lead nodes. Iteration assignment is carried out again if variation

new old in either of F or L exceeds certain preset

threshold TH .D-PERA framework can be easily be adapted for multipath

multi-stream video transport. Assuming to be the summation operator, and l the iteration assignment vector for a node pair, we are given an iteration assignment budget

( )pipairl to be distributed among a pair of nodes over a

particular path such that resulting distortion pipairD is

minimized.( , )* ( , )

( )

00

pi

pi

pair L L F F

L F pair

L

F

Minimize D f l f l

SuchThat l l l

ll (2)

where 1 2 1 2 2 1* (1 ) (1 ) and L , F are the associated channel error probabilities for the symmetric channels associated with each node.

We use Newton’s optimization [19] in order to solve the problem which has advantages in terms of computational complexity and algorithmic convergence when provided with good initial values. We use proportional assignment for

initialization such that: ( )LL pair

L Fl l . Any leftover

node from the pairing process is assigned( )

2pairl

iterations.

In case of multiple paths leading to the base stations, an

inherent question can be as to how much budget ( ) pairlshould be assigned to each pair for a certain path. The decision to choose ( ) pairl depends on the significance level of the

data being transmitted, quality of the path and the energy capabilities of individual sensor nodes. For instance, if a certain path is carrying intra-coded I-frames that are critical at the decoder on the receiving end for obtaining their predictive inter-coded counterparts, we may consider higher ( ) pairlfor that particular path as compared to others. Although PVC achieves high coding efficiency by exploiting the temporal correlation between adjacent frames, it makes the reconstruction of a frame depend on the successful reconstruction of its reference picture. Without effective error protection, a lost packet in a frame can cause not only error within this frame, but also errors in many following frames, even when all the following frames are correctly received. Thus, selection of good ( ) pairl values may prove critical in

PVC. Similar arguments can be made regarding base layer in case of layered coding or I-frames for DVC. In this work, we have left the decision for path pair-budget selection in the partial decoding budget allocator, on the network designer, though, further rate-distortion formulations can be formulated depending on the desired level of reliability at the destination and energy capabilities of individual nodes.

D. Unequal Error Protection The proposed architecture can be optionally combined with

Unequal Error Protection (UEP), where, the channel coding rate is varied based on various node and network parameters. This does not come without few challenges though. In most of UEP literature, the packet size n is kept constant, but information bits k are varied based on [25]. Again, there are M streams and K M channel codes maybe used to encode them. but, issue is, for each variation in information bits, we may need a different channel code to encode it. Thus, since sensor networks are randomly deployed, tradeoff in sensor networks is greater amount of storage for each LDPC code that might be required at each node depending on type of code used to encode that particular stream.

IV. SAMPLE APPLICATION SIMULATION SETUP

In most of the prevalent video coding architectures, video encoding is the primary computationally extensive task with complexity dominated by the motion-search operation. For the scope of this work, we assume that sensors have the capability to encode video though, efficient source coding of video still remains a wider research problem. In addition, we further assume that we are not facing motion estimation complexity issues that maybe encountered in coding over video sensor nodes and confine our focus towards data reliability aspects using progressive error recovery within the network.

We consider a sensor node transmitting to a base station

33p

11 p 14p

13p12 p

21p 23p22 p 24 p

31p

32p34 p

35 p

978-1-4244-2734-5/09/$25.00 ©2009 IEEE 210

>Paper # 13< 5

with multiple paths leading to the base station obtained through some underlying routing scheme such as directed diffusion [8]. The transmission range of each node is set to

2r m with min 0.005 and max 0.08 . We assume each node equipped with Atmel Atmega128L processor transmitting at transmit power 1TP mW with 1:2000 ratio between transmit and computation power [21]. We take an LDPC code with degree distribution polynomial

6 2 1( )= 0.207 +0.271 +0.522 . Each frame is encoded with length n LDPC code. We set information length 1024kbits with channel coding rate R . The average number of ones in a parity check matrix of an LDPC code is taken as

11

0

( )d [24]. Fossorier et al. [20] give a comparison of

mathematical operations required for one LDPC decoding iteration for various LDPC decoding algorithms including sum product algorithm [23]. We use log-domain sum product algorithm for decoding LDPC encoded data due to its advantages in terms of computational complexity. The total energy for end-to-end transmission is the transmission cost in transmitting n bits over end-to-end multi-hop path, whereas, energy estimate for a single path in MMDR involves both transmit and computation energy for in-network processing. We set number of paths 4K with multiple realizations of the multi-hop network for each run such that average end-to-end equivalent error probability av remains constant. All the results are averaged over 100 runs. We use two paths in order to send packets belonging to I frame and two for P frame. We set GOP size equal to 6. Both I & P-frame data is equally divided on each corresponding path such that despite losses in datasets for one frame other frames remain unaffected. On receive side, a resequencer collects all the packets, and assigns them to corresponding streams, which are then fed to channel and source decoding modules to obtain the reconstructed video ( Figure 2. ).

Table: H.264 encoder parameters Parameter ValueSequence Foreman.qcif

YUV Format 4:2:0 QP 28

Profile IDC High Entropy Coding CABAC Sequence Type IPP

Error Concealment On

V. RESULTS AND DISCUSSION

In this section, we present some preliminary simulation results that illustrate the efficiency of the proposed MMDR architecture in WSNs with use of an H.264 video encoder [1].

Figure 4. MMDR vs end-to-end transmission with 0.5R , 4K ,

4N , ( ) 8pairl a) 0.12av b) 0.125av

Figure 4. shows the performance of MMDR framework as compared to end-to-end channel coding. We see that MMDR outperforms end-to-end channel coding by significant margin, achieving reliable communication with much higher energy efficiency for average end-to-end equivalent error probabilities as high as 0.125av with a pair budget of

( ) 8pairl .

Figure 5. shows the transmitted and reconstructed frame for both end-to-end channel coding and MMDR framework under different channel conditions with different pair-wise iteration budgets. The results clearly indicate much improved performance under MMDR for same energy levels used during end-to-end transmission.

(a)

(b) Figure 5. Frame # 2(P) for Foreman QCIF sequence with R 0.569 (a) i- No losses ii- With 10% losses ( av 0.10 ) and end-2-end scheme iii- With

av 0.10 & MMDR scheme and ( ) 6pairl (b) i- No losses ii- With

11.75% losses ( av 0.1175 ) and end-2-end scheme iii- With 11.75% losses

& D-PERA scheme with ( ) 8pairl

5 5.5 6 6.5 7 7.5 8 8.5 9 9.50

0.02

0.04

0.06

0.08

0.1

0.12

0.14

Energy(J)

Exp

ecte

d B

ER

MMDR 0.12end-to-end 0.12MMDR 0.1250end-to-end 0.1250

978-1-4244-2734-5/09/$25.00 ©2009 IEEE 211

>Paper # 13< 6

Figure 6. Average YPSNR for both MMDR with ( ) 10pairl and End-

to-End for R=0.5 LDPC encoded H.264 video.

Figure 6. uses peak signal-to-noise ratio (PSNR) in order to measure quality of video sequences for different end-to-end equivalent channel error probabilities. For a channel coding rate R=0.5 and average end-to-end equivalent error probability as high as 0.14, the plots clearly establish that MMDR outperforms end-to-end channel coding by considerable margins.

The results indicate that MMDR provides reliability to video streams in WSNs without fast depleting the energy resources of video sensors. They further depict that little bit of processing at intermediate nodes can greatly enhance video quality at the destination node. Use of path diversity gives the added flexibility in terms of added lifetime for WSNs. The rate-distortion curves obtained in Figure 4. and Figure 6. provide insight onto fine tuning the functional demands from the network in terms of adjusting the expected bit error rate/PSNR at the destination node for amount of energy spent in transmitting the bits reliably over end-to-end path.

VI. CONCLUSION

Motivated by the stringent requirements due to resource constrained nature of sensor nodes, in this work, we have proposed an architecture to provide reliability to video in WSNs. We have described the architectural platform, theoretical foundations, as well as the bridge from theory to video practice, and presented promising experimental results based on real-world video sequence that establishes the efficacy of our proposed solution over prevalent reliability schemes in WSNs. The extension of the proposed framework to DVC, SVC or MDC can be achieved in a similar fashion.

ACKNOWLEDGMENT

The authors would like to thank Sohraab Soltani and Rami Halloush at Wireless & Video Communications Laboratory at Michigan State University for their valuable insights during implementation phase of the proposed framework.

REFERENCES

[1] JM reference software, http://iphome.hhi.de/suehring/tml/[2] Vivek K. Goyal, Multiple Description Coding: Compression Meets the

Network, IEEE Signal Processing Magazine, September, 2001. [3] Qaisar, S. B.; Radha, Hayder, "OPERA: An Optimal Progressive Error

Recovery Algorithm for Wireless Sensor Networks," 4th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc Communications and Networks (SECON), June 2007

[4] Qaisar, S. B.; Radha, Hayder, "Optimal Progressive Error Recovery for Wireless Sensor Networks using Irregular LDPC Codes," 41st Annual Conference on Information Sciences and System (CISS '07), March 2007

[5] Qaisar, S. B.; Karande, S.; Misra, K.; Radha, H., "Optimally Mapping an Iterative Channel Decoding Algorithm to a Wireless Sensor Network," IEEE International Conference on Communications,( ICC '07), June 2007

[6] Qaisar, S.B.; Radha, Hayder, “Low Complexity Distributed Reliability for Wireless Sensor Networks”, 42nd Annual Conference on Information Sciences and System (CISS’08), March 2008.

[7] Wang, Y; Reibman, Amy R; Lin, S, Multiple Description Coding for Video Delivery, Proceedings of the IEEE, Vol. 93, No.1, January 2005.

[8] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva, “Directed diffusion for wireless sensor networking,” IEEE/ACM Trans. Networking, vol. 11, pp. 2-16, Feb. 2002

[9] “Video Transport Over Ah Hoc Networks: Multistream Coding With Multipath Transport” IEEE Journal on Selected Areas in Communications, Vol. 21, No. 10, December 2003

[10] D. Sidhu, R. Nair, and S. Abdallah, “Finding disjoint paths in networks,” in Proc. ACM SIGCOMM, Zurich, Switzerland, Sept. 1991, pp. 43–51.

[11] Deepak, G., Govindan, R., Shenker, S and Estrin, D., Highly-resilient, energy-efficient multipath routing in wireless sensor networks, ACM MOBIHOC 2001

[12] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva, “Directed diffusion for wireless sensor networking,” IEEE/ACM Trans. Networking, vol. 11, pp. 2-16, Feb. 2002

[13] I.F. Akyilidiz et al., A survey on wireless multimedia sensor networks, Computer Netw. (2006), doi: 10.1016/j.comnet.2006.10.002

[14] Crossbow Imote2 Multimedia Board – IMB 400 for Wireless Sensor Networks, Product File, 2008.

[15] Taal J. and Lagendijk I., Asymmetric Multiple Description Coding using Layered Coding and Lateral Error Correction,

[16] Puri, R., Majumdar, A., Ishwar, P and Ramchandran, K., Distributed Video Coding in Wireless Sensor Networks, IEEE Signal Processing Magazine, July, 2006

[17] Radha, H. and Schaar, M., The MPEG-4 Fine-Grained Scalable Video Coding Method for Multimedia Streaming Over IP, IEEE Transactions on Multimedia, Vol.3, No.1, March, 2001

[18] ITU-T, ISO/IEC JTC1. ITU-T Recommendation H. 264-ISO/IEC14496-10 AVC, Advanced Video Coding for Generic Audiovisual Services [S], 2003.

[19] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Cambridge Univ. Press, 2004.

[20] Marc P. C. Fossorier, Miodrag Mihaljevi´c, and Hideki Imai, “Reduced Complexity Iterative Decoding of Low-Density Parity Check Codes Based on Belief Propagation”, IEEE Transactions on Communications,Vol. 47, No. 5, May 1999

[21] Ramesh Govindan, IEEE COMSOC Wireless Sensor Networks Tutorial, http://www.comsoc.org/freetutorials/nsc/

[22] Sason, I. and Urbanke, R., Parity-Check Density Versus Performance of Binary Linear Block Codes Over Memoryless Symmetric Channels, IEEE Transactions on Information Theory, Vol. 49, No. 7, July 2003.

[23] Richardson T. J., Shokrollahi A., and Urbanke R., Design of provably good low-density parity-check codes, IEEE Transactions on Information Theory, vol. 47, pp. 619–637, Feb. 2001.

[24] Sason, I. and Urbanke, R., Parity-Check Density Versus Performance of Binary Linear Block Codes Over Memoryless Symmetric Channels, IEEE Transactions on Information Theory, Vol. 49, No. 7, July 2003.

[25] M. van der Schaar and H. Radha, “Unequal packet loss resilience for fine-granular-scalability video,” IEEE Trans. on Multimedia, vol. 3, no. 4, pp. 381–393, Dec. 2001.

0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.1510

15

20

25

30

35

40

Channel Error Probability

PS

NR

(dB

)

YPSNR MMDRYPSNR MMDR

978-1-4244-2734-5/09/$25.00 ©2009 IEEE 212