An adaptive content based H.264/AVC rate control in low bit rate video

Contents lists available atScienceDirect

International Journal of Electronics and

Communications (AEÜ)

j o u r n a l h o m e p a g e :w w w . e l s e v i e r . d e / a e u e

An adaptive content based H.264/AVC rate control in low bit rate video

Fu-Chuang Chen, Yi-Pin Hsu

∗

Institute of Electrical Control Engineering, National Chiao Tung University, P.O. BOX 25-92, Hsinchu 30099, Taiwan, ROC

a r t i c l e i n f o

Article history:

Received 8 December 2009 Accepted 28 July 2010 Keywords:

Laplacian based model Rate control JVT-G012

Histogram based detector

a b s t r a c t

In this paper, we analyze a rate control formula, which is Laplacian based model, then propose a content based equation for an enhanced JVT-G012 algorithm. Based on the simple and fast characteristics of the JVT-G012 algorithm, which is very suitable for real time embedded systems, we integrate the equation into the JVT-G012 algorithm to create an efficient quantizer. However, active local motion will produce error distortion and influence the determination of the quantization parameter. To solve this problem, a modified histogram based detector is proposed. Combining this detector with the equation refines the predicted quantization parameter more precisely. It is shown both theoretically and experimentally that a well-designed quantizer not only improves video quality, but also maintains the designed bit rate budget. In addition, our algorithm can achieve superior results to existing algorithms-up to 0.72 dB without large computation loading.

© 2010 Elsevier GmbH. All rights reserved.

1. Introduction

Rate control plays an important role in the H.264/AVC video encoder, which is the newest standard from the ITU-T Video Coding Experts Group and the ISO/IEC Moving Pictures Experts Group such as MPEG-2 Visual[1], H.263[2], MPEG-4 Visual[3], and H.264/AVC [4]. Several coding parameters must be determined, such as the macro block (MB) type, quantization parameter (QP), and multi-ple reference frames. As a result, it is able to achieve much higher rate control efficiency than conventional video coding methods. For real time video applications, such as video surveillance and video conferencing, bandwidth is an important consideration. Usually, rate control regulates the coded bit stream by modifying the QP. Because the bandwidth is small and expensive, a low bit rate envi-ronment requires more accurate bit allocation and sufficient image quality. Without rate control, underflow or overflow will occur between the encode bit rate and available channel bandwidth. In other words, an encoder without rate control would be difficult to use. It is worthwhile to point out that the rate control algorithm has been developed in two parts. The Lagrange method is used to con-struct a new R-D model in[5,6], which propose an R-D optimized rate control algorithm with an adaptive initial QP determination scheme that differs from the JVT-G012 algorithm quadratic form [7]. Although these models can possible obtain the optimum QP, the correlation between two frames is calculated to modify the

ini-∗ Corresponding author.

E-mail addresses:[email protected](F.-C. Chen), [email protected](Y.-P. Hsu).

tial QP and the computational complexity is enormous. In addition, a near optimization searching method is proposed in[8]that is also based on the Lagrange method to pick up a suitable QP for achiev-ing a more accurate bit rate budget. However, the complexity of the encoder increases proportionally with the number of iterations, and it depends on the initial QP and relative parameters, such as the Lagrange multiplier, block mode, and reference frames. Because the Lagrange method has a complex computation and cannot suit real time applications, the histogram of difference (HOD) is used to detect which frames are skipped to save encoding time; moreover, the HOD value is taken to modify the bit allocation and update the Lagrange multiplier in[9,10]. Secondly, the quadratic form, which is assumed as statistical distribution, is used as the main skeleton for more accurately modifying the mean absolute difference (MAD) prediction in[11,12]and to find a suitable QP. Both[11]and[13] propose an adaptive MAD prediction method. The frame complex-ity is the main consideration, and using the empirical MAD ratio to change the QP is their main contribution. Experiments confirm that the estimation error of the quadratic model can be significantly reduced to perform a simplified and efficient form in[12]. The sec-ond order term is dropped in the original quadratic model without sacrificing much performance, and the first order term is modified by a model parameter to approximate the original quadratic model and the order of the power can be changed in different frame types, such as I-, P-, and B-frame.

Here, methods with large computational complexity are not used because real time embedded systems have limited resources. Fortunately, there is a simple and fast quadratic form that is an efficient closed form, which was recommended and implemented in the H.264/AVC reference software JM 15.0[14]. Although the QP 1434-8411/$ – see front matter © 2010 Elsevier GmbH. All rights reserved.

determination technique used in the JVT-G012 algorithm is state of the art, it is not sufficiently accurate for low bit rate video coding, and does not address the problem of local motion. Even if the JVT-G102 algorithm is selected as the main QP prediction method, the input frame must be assumed to have a specified statistical distribu-tion. Determining this distribution function becomes an important issue because the frame content varies. To use the simple and fast model, we propose an equation constructed based on an analytic rate distortion function and fed into the JVT-G102 algorithm to refine the initial QP as a middle result. To avoid high computational complexity, the middle QP will be refined again as a final QP by a modified HOD.

The rest of this paper is organized as follows. In Section2, we briefly review the original quadratic rate control model for QP com-putation. An enhanced equation will be analyzed and verified with the original quadratic model, and a modified HOD is added to solve the problem of local motion in Section3. We then evaluate the pro-posed algorithm and exhibit the simulation results in Section4. Finally, Section5gives the conclusions.

2. Overview of the statistically based rate distortion model In[15], a mathematical rate distortion was derived and proved in detail. To fit this theorem, however, the selection of a suitable distribution model is important. In previous approaches, the Lapla-cian model was derived as the rate distortion equation – this has also been used in recent research [7,11,12], thus, this model is worth consideration. From[15,16], the relation between rate and distortion can be formulated as

R = ln 1

1 × D (1)

where R, and D represents the bit rate budget, variance and dis-tortion respectively. Eq.(1)can be expanded by a Taylor series to build the new model. It is modified by substituting QP and MAD for D and, respectively. The new model, following[17], is formulated as R = X1× MAD QP + X2× MAD QP2 (2)

Thus, the QP can be obtained when the coefficients and target bit rate are given; this formula is adopted by the JVT-G102 algorithm. Because it is simple and fast, it is suitable for real time imple-mentation. However, past approaches, such as[9,10,18], show that JVT-G102 cannot provide a suitable solution to accurately estimate the QP of a P-frame at a low bit rate. Even if these approaches solve this problem and provide acceptable results, the computation loading is too large.

3. Proposed adaptive content based algorithm

3.1. Analyze and verify a Laplacian based model on the JVT-G012 algorithm

Although the JVT-G102 algorithm cannot accurately obtain a QP in low bit rate video immediately, the result is quite close to the real requirements. Even if a two pass algorithm in[12]is proposed to enhance the JVT-G102 algorithm, the computation is still large. Consequently, we propose an efficient equation for the JVT-G102 algorithm that will improve performance and provide a low com-putational burden. Eq.(1)shows that the relation between rate and distortion can be demonstrated, and the relation between rate and QP is formulated in(2). We can modify(1)to create a new equation as

D = 1

 × eR (3)

Table 1

Performance of two methods in term of average PSNR, and bit rate.

Methods JVT-G012 JVT-G012+(3) Gain

Test video sequence: Mother and Daughter (24 kbps)

Average PSNR (dB) 35.33 35.28 −0.05

Bit rate (kbps) 24.54 24.60 –

Test video sequence: Hall (32 kbps)

Average PSNR (dB) 35.99 36.13 +0.14

Bit rate (kbps) 32.18 32.14 –

Test video sequence: Highway (48 kbps)

Average PSNR (dB) 37.08 36.93 −0.15

Bit rate (kbps) 48.70 49.17 –

In(3), we assume that the distortion D is variable when bit rate R and variance are given by current MB, the initial distortion D is predicted by the JVT-G102 algorithm. Note that D can be replaced by the mean square error or mean absolute difference for a fast and simple implementation. A well defined relation[19]is constructed as follows:

D ∝ QP (4)

Thus, we can adjust QP under the predicted term of the tar-get budtar-get to satisfy the equation. But, multiple iterations are not acceptable for real time application in resource limited systems. Nevertheless, the first result in(3)can be used as guidance to refine the initial QPinitial; then we can combine(3) with the JVT-G102

algorithm to create a more efficient rate control algorithm. The JVT-G102 algorithm calculation produces the QPinitial,(3)converts

it to an inequality to increase or decrease the QPinitial. In general, if

D > 1/( × eR_{) then QP = QP}

initial− 1; otherwise QP = QPinitial+ 1. The

QPinitialdetermination for an I-frame is not considered; this only

focuses on the P-frame level.

The beginning experimental results, including the average peak signal to noise ratio (PSNR) and bit rates, are given inTable 1. The simulated conditions are described in Section4. It focuses on three parts: the bit rates, the PSNR, and enhanced PSNR. Although this method can improve image quality, only partial test video sequences are efficient.

Fig. 1shows a frame by frame PSNR comparison; in addition, we analyze the drop of PSNR in the visual domain. From the 80th frame to the 99th inFig. 1(c), the PSNR is influenced by the add adjustment equation,(3). In the original video sequence, the object in the frame has active local motion from back to front (Fig. 1(c)). For this reason, a noticeable result that can be addressed is the local motion problem and (3), which can provide smooth frames but does not apply to active local motion. Unfortunately,(3)refines the QPinitialfrom the JVT-G012 algorithm in the wrong direction.

3.2. Proposed rate control algorithm

Even if multiple iterations are avoided, a single QP determina-tion may be not accurate enough to compare with the real QP and active local motion can reduce the PSNR when(3)is applied. The QP must be decreased to improve the PSNR due to active local motion; the QP should be also increased to improve the bit rate budget when the local image is inactive and motionless. An efficient and fast detection method becomes important. The histogram based meth-ods are considered first, as they are useful and directly detect image content variation[20]. To maintain low complexity computation and local motion, block level processing and HOD which is defined at frame level operation are considered. A detailed description of HOD is given in[20]. The modified calculation equation, HODB, is given as

HODB(bn, bn−1)=



i>|TH|

0 10 20 30 40 50 60 70 80 90 100 31 32 33 34 35 36 37 38 39 40 41 Frame Number PSNR (dB)

Highway QCIF at 48Kbps and 30 fps

JVT−G012 JVT−G012+(3) (c) 0 10 20 30 40 50 60 70 80 90 100 33 34 35 36 37 38 39 40 41 Frame Number PSNR (dB)

Hall QCIF at 32Kbps and 30 fps

JVT−G012 JVT−G012+(3) (b) 0 10 20 30 40 50 60 70 80 90 100 30 32 34 36 38 40 42 Frame Number PSNR (dB)

Monther and Daughter QCIF at 24Kbps and 30 fps

JVT−G012 JVT−G012+(3)

(a)

Fig. 1. PSNR comparison between JVT-G012, and JVT-G012 with Eq.(3)for three standard test sequences: (a) “Mother and Daughter”, (b)“Hall”, and (c) “Highway”.

where i is the index of the histogram bin, the histogram of differ-ence block is from hod block(i), Npixelis the number of pixels and

TH is a threshold for determining the proximity of the position to zero. bn and bn−1indicate the current block and previous block,

respectively. Some excellent results are demonstrated and proven in[18]. From a physical viewpoint, the HODB value is generally increasing monotonically. Generally speaking, the block contains more information if the HODB is nonzero. Although the descrip-tive algorithm in section A decreases the PSNR during acdescrip-tive local motion, it can enhance the PSNR in smooth areas. Thus, we try to

keep the property and use the HDOB to avoid the influence of active local motion.

In JVT-G012, the bit allocation only focuses on predefined target bits and buffer status to compensate for bit cost. To meet the real time embedded system, two concepts of data-reuse and lower com-putation should be considered for bit allocation. Thus, we propose a modified equation in(6)that uses the HODB value for target bit adjustment without large computation. The target bit prediction, Tall, is defined in[4,14].

R = (0.9 + HODB × ϕ) × Tall (6)

The parameterϕ, obtained from experiments, is set to be ϕ =



5, 0.9 < HODB ≤ 1

0.9, HODB ≤ 0.9 (7)

After, Laplacian based model and reused HODB information to pro-vide efficient modification to improve the QPinitialfrom JVT-G012.

The full algorithm can be described as follows.

Step 1: Let R be the target bit budget for current frame, Rifor ith MB

and i = 1. N is the total number of MBs in the frame after assigning the Laplacian model.

Step 2: Obtain the HODB value according to the previous recon-structed and current frame in the co-located MB. Adjust the Ri

adaptively by(6).

Step 3: With the JVT-G102 algorithm, use Riand the predicted

MAD to get the initial QP, QPinitial, by(2).

Step 4: Select (3) as an adjustment equation and transfer to an inequality. Calculating variances . If D > 1/( × eRi_{) then} QPtmp= QPinitial− 1; otherwise QPtmp= QPinitial+ 1.

Step 5: Select (5) as an adjustment equation. Calculate the histogram hod block(i) and set TH to 32. If HODB > 0 then QPfinal= QPtmp− 3; otherwise QPfinal= QPtmp+ 3.

Step 6: Use QPfinalto encode the current MB i and set i = i + 1. If

i≤ N, go to step 2 to encode next MB. Otherwise stop the encoding procedure.

Overall, an efficient formula can be summarized as in(8). From the viewpoint of a real time embedded system,(8)can be imple-mented in parallel because the input data is independent and has a low computational complexity.

QPfinal=

⎧

⎪

⎪

⎪

⎪

⎪

⎪

⎪

⎪

⎨

⎪

⎪

⎪

⎪

⎪

⎪

⎪

⎪

⎩

QPinitial+ 4, D < 1  × eR



HODB≤ 0 QPinitial+ 2, D > 1  × eR



HODB≤ 0 QPinitial− 2, D < 1  × eR



HODB> 0 QPinitial− 4, D > 1  × eR



HODB> 0 (8)

For a clearer description,Fig. 2depicts integrating our algorithm into the JVT-G012 block diagram. In addition to the shadow blocks in rate control in the proposed functions, all other functions are built in the original H.264/AVC encoder system. The main goal of rate control is to provide a more accurate QP for quantization. The computation complexity of QP selection should be limited to meet real time applications. The implementation also considers the memory requirement in the embedded system due to limited hard-ware resource.Fig. 2shows directly calculating the HODB without extra buffer allocation and sending the HODB value to the bit alloca-tion adjustment and QP adjustment, after deciding the final target bit rate and final QP. Thus, the proposed algorithm is very suitable to integrate on the current H.264 encoder embedded system.

Fig. 2. Integration of JVT-G012 and our proposed algorithm block diagram in H.264/AVC.

4. Experimental results

The adaptive context based rate control algorithm with a Lapla-cian based model and a HOD method was incorporated into the JM H.264/AVC reference software[14]. The evaluation was conducted with the first 100 frames of six QCIF test sequences. For the low bit requirement, 24, 32, and 48 kbps were selected. The test target bit rate is 24 kbps for “Grandmother”, “Mother and Daughter”, and “Miss-America”, 32 kbps for “Akiyo”, “Bridge-far”, and “Hall”, and 48 kbps for “Carphone”, and “Highway”. Each sequence is coded at 30 fps with an IPPP structure, which indicates only one I-frame in first frame and P-frame in residual frames, thus the influence of I-frames can be reduced and ignored. There are five reference I-frames and the search window is set to 15. Context adaptive variable length coding, rate distortion optimization, and rate control are enabled. The residual parameters are fixed equally. As a reference, the per-formance of the JVT-G102 algorithm and[13]were compared with our proposed algorithm. Although three layer rate control algo-rithm is presented in[13], only using the block complexity ratio to modify the Lagrange multiplier layer provides the best mode deci-sion and QP selection. Our algorithm also uses block complexity to select a suitable QP, thus,[13]becomes a fair comparison.

Fig. 3(a)–(c) shows a frame by frame comparison in PSNR terms for three standard test sequences: “Mother and Daughter”, “Hall”, and “Highway”. The effectiveness of our algorithm is further indi-cated by the fact that our algorithm improved the video quality more than JVT-G102 algorithm and[13]. Indeed, our algorithm enhanced the PSNR to a greater degree than the two algorithms overall. ComparingFigs. 1(c) and 3(c)from the 80th to the 99th frame shows that the degree of PSNR enhancement improved significantly and avoids the local motion problem. Moreover, com-paringFigs. 1(c) and 3(c)from the 15th to the 25th frame shows that the enhanced PSNR is maintained, even when the HOD method is applied. The PSNR is close to the two algorithms under smooth con-ditions and superior when active local motion is produced. Detailed results are given inTable 2. Our algorithm is close to the required bit rate budget, which means that the algorithm can achieve the PSNR enhancement and still match the required bit rate. To eval-uate the bit rate mismatch quantification further, a useful formula

Table 2

Performance of three algorithms in term of average PSNR, PSNR Std. deviation, bit rate andR.

Methods Average PSNR (dB) PSNR std. deviation Bit rates (kbps) R

Test video sequence: Grandma (24 kbps)

JVT-G012 36.17 2.91 24.21 0.88

Wang[13] 35.83 2.87 24.21 0.88

Our 36.55 2.26 24.19 0.79

Gain +0.38/+0.72 – – –

Test video sequence: Mother and Daughter (24 kbps)

JVT-G012 35.33 3.83 24.54 2.25

Wang[13] 35.01 3.94 25.60 6.67

Our 35.48 3.61 24.54 2.25

Gain +0.15/+0.47 – – –

Test video sequence: Miss-America (24 kbps)

JVT-G012 39.39 2.31 24.44 1.83

Wang[13] 38.94 2.69 25.25 5.21

Our 39.53 1.30 24.28 1.17

Gain +0.14/+0.59 – – –

Test video sequence: Akiyo (32 kbps)

JVT-G012 39.70 1.35 32.12 0.38

Wang[13] 39.25 1.54 32.28 0.88

Our 39.77 1.16 32.19 0.59

Gain +0.07/+0.52 – – –

Test video sequence: Bridge-far (32 kbps)

JVT-G012 39.52 0.08 32.10 0.31

Wang[13] 39.67 0.10 32.22 0.69

Our 39.96 0.46 32.35 1.09

Gain +0.44/+0.29 – – –

Test video sequence: Hall (32 kbps)

JVT-G012 35.99 2.09 32.18 0.56

Wang[13] 35.82 2.16 32.20 0.63

Our 36.16 1.93 32.17 0.53

Gain +0.17/+0.34 – – –

Test video sequence: Carphone (48 kbps)

JVT-G012 34.18 1.52 48.45 0.94

Wang[13] 33.80 2.11 49.89 3.94

Our 34.21 1.26 48.25 0.52

Gain +0.03/+0.41 – – –

Test video sequence: Highway (48 kbps)

JVT-G012 37.08 1.95 48.70 1.46

Wang[13] 36.84 1.55 49.10 2.29

Our 37.41 1.07 48.94 1.96

0 10 20 30 40 50 60 70 80 90 100 31 32 33 34 35 36 37 38 39 40 41 Frame Number PSNR (dB)

Highway QCIF at 48Kbps and 30 fps

JVT−G012 Wang Our (c) 0 10 20 30 40 50 60 70 80 90 100 33 34 35 36 37 38 39 40 41 Frame Number PSNR (dB)

Hall QCIF at 32Kbps and 30 fps

JVT−G012 Wang Our (b) 0 10 20 30 40 50 60 70 80 90 100 30 32 34 36 38 40 42 Frame Number PSNR (dB)

Monther and Daughter QCIF at 24Kbps and 30 fps

JVT−G012 Wang Our

(a)

Fig. 3. PSNR comparison among JVT-G012, Wang[13], and our proposed algorithm for two standard test sequences: (a) “Mother and Daughter”, (b) “Hall”, and (c) “Highway”.

[6]is used in the paper. R =Rtest− Rbudget

Rbudget × 100%

(9) where Rtestis the bit rate of the test algorithm and Rbudgetis the

target bit rate. A smallR means that test algorithm can achieve the required bit rate. The gain inTable 2denotes the comparison of our algorithm to the JVT-G102 algorithm and[13]. In our algo-rithm, the performance with respect to video quality was enhanced by an average of 0.49 dB and a maximum of 0.72 dB compared to

Table 3

Comparisons of main computation amounts in two different algorithms per MB.

Wang[13] Our # of addition 256 256 # of subtraction 256 256 # of multiplication 1 0 # of division 3 1 # of square root 1 0

[13]and by an average of 0.21 dB and a maximum of 0.44 dB com-pared to the JVT-G012 algorithm. Not only did the average PSNR improve dramatically but the standard deviation is lower than that of the existing algorithms. Finally, all PSNR standard deviations are shown and our proposed algorithm maintains a lower deviation than the other two algorithms for test video sequences. Compar-ingTables 1 and 2shows more excellent results. Although(3)can improve the QP prediction, active local motion will lead to inaccu-rate QP, which decrease the PSNR and waste the bit inaccu-rate budget. To correct the QP prediction, the HODB method is applied to modify QP quickly and accurately.

The comparison of calculation units presents the low com-putation property of the proposed algorithm. Both[13] and our proposed algorithms are built on the JVT-G012, which is com-pared in terms of addition, subtraction, multiplication, division and

Fig. 4. Visual comparison of the original (top), JVT-G102 (middle), Wang[13](third), and the proposed algorithm (bottom). These frames are the 81th and 33th in the test video sequences “Mother and Daughter” and “Hall”, respectively.

Fig. 5. Visual comparison of the original (top), JVT-G102 (middle), Wang[13](third), and the proposed algorithm (bottom). These frames are the 95th and 78th in the test video sequences “Highway” and “Grandma”, respectively.

square root operations. Thus, the MB level describes and analyzes extra computation in[13]and in our proposed algorithm. In[13], the adaptive Lagrangian multiplier in the ith MB is the main con-tribution and computation. The formula is expressed as follows. i= i×

_MAD frame MADMB

1 4 (10) where MADframeand MADMBare the actual MAD in the previous

frame and co-located MB respectively. In(10), the numbers of mul-tiplication and square root operations are equal to 1. The addition and subtraction operation for actual MAD calculation are equal to 256 respectively, because the linear regression prediction does not obtain MAD, moreover, the division, operation is 3. The HODB computation is also the main contribution and computation in the proposed algorithm without square root operation. The number of subtraction operations is 256 for two MB operations and the number of addition operations is 256 for histogram accumulation. Table 3presents the comparison of five mathematical operations with two different algorithms.

The division and square root operation has large computa-tion loading, thus avoiding or reducing the operacomputa-tion requires implementing in hardware or in the firmware level[21].Table 3 clearly shows that the proposed algorithm is less than [13] in division and square root. Since the digital signal processor (DSP) based system produces multiple multiply-accumulates per cycle, implementing our algorithm in the DSP-based embedded system

reduces the addition and multiplication operations. The histogram function has also been recently built into the DSP core, such as TI TMS320DM6437 [22], due to this very popular function for image and video processing. Thus, HODB computation is not necessary.

Figs. 4 and 5 shows actual video sequences to demonstrate that our algorithm suits real applications. The three algorithms (JVT-G102,[13], and our proposed algorithm) are demonstrated simultaneously when the test video sequences “Mother and Daugh-ter”, “Hall”, “Highway”, and “Grandma” are selected. From visual comparison, our proposed algorithm can select a more precise QP to provide an acceptable visual image quality. It is clear that our proposed algorithm can achieve a higher PSNR than the existing algorithms and reduces the bit rate mismatch ratio.

5. Conclusion

This paper has presented an efficient rate control algorithm for a H.264/AVC video encoder. Through rate control model analy-sis, an adjustment equation is constructed, based on a statistical model. However, during active local motion the QP will be refined in an incorrect direction. Considering frame complexity and then adaptively enhancing the QP prediction accuracy of the JVT-G102 algorithm, using the HOD method, not only produces a superior result during smooth frames but also generates outstanding results during active local motion. Experimental results show that our algorithm requires no pre-encoding processing and can achieve effective improvements in PSNR-it provided more enhancement than either[13]or the JVT-G102 algorithm. In addition, all algo-rithms also can be kept on a limited bit rate budget. Thus, our algorithm is suitable for H.264/AVC video rate control even if ported into a real time embedded system.

Acknowledgement

We would like to thank the reviewers for their valuable recom-mendation.

References

[1] ITU-T; JTC1, I.: Generic coding of moving pictures and associated audio informa-tion – part 2: video. ITU-T Recommendainforma-tion H.262 – ISO/IEC 13818-2 (MPEG-2) (1994).

[2] ITU-T: Video coding for low bitrate communication. ITU-T Recommenda-tion H.263; version 1, November 1995; version 2, January 1998; version 3 (2000).

[3] JTC1, I.: Coding of audio-visual objects – part 2: visual. ISO/IEC 14496-2 MPEG-4 visual version 1, April 1999; Amendment 1 version 2, February 2000; Amend-ment 4 streaming profile (2001).

[4] Wiegand T, Sullivan G. Draft ITU-T recommendation H.264 and final draft inter-national standard of joint video specification (ITU-T Recommendation H.264| ISO/IEC 14496-10 AVC). Joint Video Team of ISO/IEC JTC1/SC29/WG11 and ITU-T SG16/Q.6 Doc. JVT-G050, Pattaya, Thailand; 2003.

[5] Liu Y, Li ZG, Soh YC. A novel rate control scheme for low delay video com-munication of H.264/AVC standard. IEEE Trans Circuit Syst Video Technol 2007;17:68–78.

[6] Wang H, Kwong S. Rate-distortion optimization of rate control for H.264 with adaptive initial quantization parameter determination. IEEE Trans Circuit Syst Video Technol 2008;18:140–4.

[7] Li Z, Pan F, Lem KP, Feng G. Adaptive basic unit layer rate control for JVT. In: Proc. Joint Video Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, JVTG012, 7th meeting. 2003.

[8] An C, Nguyen TQ. Iterative rate-distortion optimization of H.264 with constant bit rate constraint. IEEE Trans Signal Process 2008;17:1605–15.

[9] Song H, Kuo C-CJ. Rate control for low-bit-rate video via variable-encoding frame rates. IEEE Trans Circuit Syst Video Technol 2001;11:512–21. [10] Song H, Kuo C-CJ. A region-based H.263+ codec and its rate control for low bit

rate video. IEEE Trans Multimedia 2004;6:489–500.

[11] Jiang M, Ling N. Low-delay rate control for real-time H.264/AVC video coding. IEEE Trans Multimedia 2006;8:467–77.

[12] Kwon D-K, Shen M-Y, Kuo C-CJ. Rate-control for H.264 video with enhanced rate and distortion models. IEEE Trans Circuit Syst Video Technol 2007;17: 517–29.

[13] Wang M, Yan B. Lagrangian multiplier based joint three-layer rate control for H.264/AVC. IEEE Signal Process Lett 2009;16:679–82.

[14] H.264/AVC Reference Software JM15.0 [Online]. Available from: http://iphome.hhi.de/suehring/tml/.

[15] Viterbi A, Omura J. Principle of digital communication and coding. New York: McGraw-Hill Electrical Engineering Series; 1979.

[16] Berger T. Rate distortion theory: a mathematical basis for data compres-sion. New Jersey: Prentice-Hall Information and System Sciences Series; 1971.

[17] Chiang T, Zhang Y-Q. A new rate control scheme using quadratic rate distortion model. IEEE Trans Circuit Syst Video Technol 1997;7:246–50.

[18] Zhou S, Li J, Fei J. Improvement on rate-distortion performance of H.264 rate control in low bit rate. IEEE Trans Circuit Syst Video Technol 2007;17: 996–1006.

[19] Yuan W, Lin S, Zhang Y, Luo H. Optimum bit allocation and rate control for H.264/AVC. In: Proc. Joint Video Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, JVTO016, 15th meeting. 2005.

[20] Lee J, Dickinson W. Temporally adaptive motion interpolation exploiting temporal masking in visual perception. IEEE Trans Image Process 1994;3: 513–26.

[21] Kuo C-H, Chang L-C, Fan K-W, Liu B-D. Hardware/sofrware codesign of a low-cost rate control scheme for H.264/AVC. IEEE Trans Circuit Syst Video Technol 2010;20:250–61.

[22] TMS320DM6437 Digital Media Processor (Rev. D), 2008 [Online]. Available from:http://www.ti.com.

Fu-Chuang Chen received BS degree in power mechanical

engineering from National Tsing-Hua University, Taiwan, in 1983. He received MS and PhD degrees in electrical engi-neering from Michigan State University, USA, in 1987 and 1990 respectively. He was in military service from 1983 to 1985. He is currently a professor in electrical control engineering at National Chiao Tung University, Hsinchu, Taiwan. His research interests are in noise and distortion modeling, with applications to sigma-delta modulators and other related areas.

Yi-Pin Hsu received the MS degree in the Department of

Mechanical Electrical from National Taiwan Normal Uni-versity, Taipei, Taiwan, in 2005. He is now a PhD candidate in electrical control engineering at National Chiao Tung University, Hsinchu, Taiwan. His research interests are in the video rate control and DSP-based consumer product application.