IMPROVED INTER MODE DECISION FOR H.264/AVC

USING WEIGHTED PREDICTION

Amrita Ganguly and Anil Mahanta

Dept. of Electronics and Communication Engineering, Indian Institute of Technology Guwahati, Guwahati, India

Keywords:

H.264/AVC, Inter mode decision, Edge histogram, Video coding.

Abstract:

H.264/AVC video coding standard outperforms former standards in terms of coding efﬁciency but at the ex-

pense of higher computation complexity. Of all the encoding elements in H.264, inter prediction is computa-

tionally most intensive and thus adds to the computational burden for the encoder. In this paper, we propose

a fast inter prediction algorithm for JVT video coding standard H.264/AVC. Prior to performing the motion

estimation for inter prediction, characteristics like stationarity and homogeneity of each macroblock is deter-

mined. The macroblocks correlation with neighboring macroblocks in respect of predicted motion vectors and

encoding modes are studied. Weights are assigned for these parameters and the ﬁnal mode is selected based

upon these weights. The average video encoding time reduction in the proposed method is 70% compared to

the JVT benchmark JM12.4 while maintaining similar PSNR and bit rate. Experimental results for various

test sequences at different resolutions are presented to show the effectiveness of the proposed method.

1 INTRODUCTION

The H.264/AVC is the latest video coding standard

developed by Joint Video Group (JVT) of ITU-T

Video Coding Experts Group (VCEG) and ISO/IEC

MPEG Video Group (JVT.G050r1, 2003). It offers

better compression efﬁciency and greater ﬂexibility in

compressing, transmitting and storing video. There

are many advanced techniques that signiﬁcantly im-

prove the performance of the H.264/AVC video cod-

ing standard but at the cost of higher computational

overhead. The efﬁcient inter prediction using the vari-

able block size motion estimation (ME) and compen-

sation increase the coding complexity. H.264 permits

the use of different block sizes, 16× 16 pixels called

a macroblock (MB) down to 4× 4 pixels.

The H.264/AVC standard supports both intra and

inter prediction processes. In the inter prediction pro-

cess, there are seven block sizes P

16x16

, P

16x8

, P

8x16

8x8

, P

8x4

, P

4x8

and P

4x4

that are used by H.264 be-

sides the SKIP and the INTRA modes (Richardson,

2003), (Wiegand et al., 2003). The block sizes are

shown in Fig.1. For each MB, all the modes are tried

and one which gives the least rate distortion (RD) cost

is selected for encoding. The computation of the RD

cost requires the availability of the reconstructed im-

age and the actual bit count. This necessitates that the

encoding and decoding processes are completed for

P 16x16

P 16x8

P 8x16

P 8x8

P 8x4

P4x8

P 4x4

Figure 1: Inter prediction block sizes for a MB.

every mode. Thus the computational requirements for

the mode selection process is very high.

Several approaches have been proposed to reduce

the complexity and time for the inter mode decision

(Jing and Chau, 2004; Kim et al., 2006; Kim et al.,

2004; Lee et al., 2006; Liu et al., 2009; Park and

Capson, 2008; Shen et al., 2008; Wang et al., 2007;

Wu et al., 2005; Zeng et al., 2009; Ganguly and Ma-

hanta, 2010; Huang et al., 2010). In (Kim et al., 2006)

adaptive MB selection is used for mode decision. In

(Lee et al., 2006) the fast motion estimation is based

on the successive elimination algorithm (SEA) using

sum norms to ﬁnd the best estimate of the motion vec-

tors and to implement efﬁcient calculations for vari-

able blocks. In (Wu et al., 2005) edge histogram pa-

rameters have been used for fast inter mode decision.

Ganguly A. and Mahanta A..

IMPROVED INTER MODE DECISION FOR H.264/AVC USING WEIGHTED PREDICTION.

DOI: 10.5220/0003517300730078

In Proceedings of the International Conference on Signal Processing and Multimedia Applications (SIGMAP-2011), pages 73-78

ISBN: 978-989-8425-72-0

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

In this paper, the characteristics of a MB is ﬁrst an-

alyzed based on the stationarity, homogeneity, pre-

dicted motion vectors (MV) and the encoding modes

of neighboring MBs. All these parameters are ﬁrst

determined for each MB. Weights are then assigned

to these MBs depending upon the values of these four

parameters. The decision on the ﬁnal encoding mode

for the MB is taken based upon the weights of all

these parameters. The proposed method gives a sim-

ple model for arriving directly at the encoding mode

without performing rigorous ME. Results show that

the proposed method speeds up the encoding process

by 70% when compared to the JM12.4 software with

negligible loss in quality and coding efﬁciency.

The paper is organized as follows. The next sec-

tion introduces the proposed fast mode decision algo-

rithm. Section 3 illustrate the results. The conclusions

are drawn in Section 4.

2 PROPOSED FAST DECISION

ALGORITHM

In real video sequences the distribution of modes

is not uniform among the MBs (Richardson, 2003).

It depends upon the characteristics of the video se-

quences. Regions with homogenous motion, smooth

motion of moving background and static background

use larger block size motion compensation. For re-

gions with high detail and complex motion, smaller

block sizes to represent motion gives better coding

efﬁciency. Changes between video frames may be

caused by object motion, camera motion, uncovered

regions and lighting variations. The neighboring

video frames have large similarities between them.

ME and compensation attempts to reduce the tempo-

ral redundancy by exploiting these similarities. Natu-

ral video sequences contain many regions with homo-

geneous motion that result in a large number of MBs

being encoded with larger block sizes. The block size

is also dependent on the QP. For large value of QP,

more MBs tend to be encoded with larger block sizes.

Since video sequences have different motion com-

plexity, the sequences have been divided into three

different classes: Class A having low and simple mo-

tion, Class B with medium to high motion and Class

C with high and complex motion as given in Table 1.

2.1 Determination of Stationarity

Stationarity refers to the stillness between consecu-

tive frames in the temporal direction. Regions having

similar motion in consecutive frames are also consid-

ered temporally stationary. The MBs which are tem-

Table 1: Different Classes of Sequences.

Type CIF Sequence QCIF Sequence

News, Mother and daughter Suzie, Claire

Class A Container, Hallmonitor Missamerica

Foreman, Coastguard Foreman, Silent

ClassB Harbour, Ice Crew

Mobile, Flower Mobile, Football

Class C Tempete, Stefan Soccer

porally stationary usually get encoded in the SKIP or

in the P

16x16

mode whereas MBs with large motion

get encoded with smaller block sizes. The simplest

method of temporal prediction is to use the previous

frame as the prediction for the current frame.

Figure 2: Difference image of frame 10 and 11 of Mother

and Daughter sequence.

Fig.2 shows the difference frame formed by sub-

tracting frame 10 from frame 11 of the Mother and

Daughter (MaD) sequence. Regions which have little

or no motion have zero or low frame difference resid-

ual values represented by the grey regions and are

encoded with larger block sizes. The light and dark

grey areas correspond to positive and negative differ-

ences representing higher motion activity and hence

use smaller block sizes. For each MB in the current

frame (MB

), the residual block (R

) is obtained

from the collocated MB in the previous frame (MB

)

(i, j) = MB

(i, j)− MB

(i, j), i, j = 1, . .. , 16

(1)

From full search ME it is observed that at low values

of QP, MBs encoded in SKIP mode have very low val-

ues of residuals in R

. As QP increase, MBs in the

SKIP mode have larger values of residuals in R

Fig.3 shows that when encoded with SKIP mode, a

large number of residuals in a MB have values that are

below 2. It is also observed that the MBs having large

motion have high valued residuals in R

. Another

observation is that the values of residuals for QP val-

ues upto 24, the MBs encoded in the SKIP mode have

more than 90% residuals with absolute values that are

usually below 2. In the proposed method, these prop-

SIGMAP 2011 - International Conference on Signal Processing and Multimedia Applications

0 1 2 3

100

150

200

250

Value of Residuals for MBs encoded in SKIP mode

Number of Pixels per MB

Figure 3: Residuals in R

for News sequence (CIF).

erties are exploited. For each MB a weight is assigned

depending upon the QP and the values of the residuals

in R

. MBs with low values of residuals are given

lower weights indicating a region with little motion

whereas MBs with large values of residuals are as-

signed higher weights indicating high to complex mo-

tion. Thus a threshold TH

for QP values upto 24 is

deﬁned and is taken to be equal to 2. For higher values

of QP, the threshold is deﬁned as



if QP ≤ 24

ﬂoor[

(QP− 24) + TH

] if QP > 24

Hence with the increase of QP the threshold also

increases. In the proposed method, for each MB the

fraction of the residues N in R

that are below the

is determined where N is given by

N =

Number of residues in R

below TH

Total number of residues in R

A difference frame weight DF

is assigned to

each MB based on the value of N and is given in Table

2 below.

Table 2: Difference Frame Weights DF

N range > 0.9 0.8-0.9 0.7-0.8 0.6-0.7 < 0.6

0 1 2 3 4

2.2 Determination of Homogenous MB

Natural videos have many homogeneous regions. Ho-

mogeneous regions have similar spatial properties.

These regions in most cases get encoded with larger

block size. Regions with more complex texture get

encoded with smaller block sizes. If the homogeneity

of a MB is detected early, then decision on the encod-

ing mode can be taken. There exist many techniques

for detecting homogenous regions. In this paper the

homogenous region detection is based on the edge

information as video object have strong edges. The

edge based homogeneous region detection has been

proposed in (Wu et al., 2005).

The edge map is created for each frame using the

Sobel operator. Each pixel in the block will be as-

sociated with an edge vector containing edge direc-

tion and amplitude. The edge vector is deﬁned as

i,j

= {dx

i,j

, dy

i,j

} where

i,j

= p

i−1,j+1

+ 2 × p

i,j+1

+ p

i+1,j+1

− p

i−1,j−1

(2)

−2× p

i,j−1

− p

i+1,j−1

i,j

= p

i+1,j−1

+ 2 × p

i+1,j

+ p

i+1,j+1

− p

i−1,j−1

(3)

−2× p

i−1,j

− p

i−1,j+1

Here dx

i,j

and dy

i,j

are the degrees of differences

in vertical and horizontal directions. The amplitude

of the edge vector and the sum of the amplitudes are

computed as

Amp(D

i,j

) =



i,j



i,j



, (4)

Amp

16,16

∑

i=1,j=1

Amp(i, j), (5)

Homogenous regions will have a small value of S

Amp

whereas this value increases as the MB becomes less

homogeneous. The value of S

Amp

is plotted and is

shown in Fig. 4 where the edge histogram statistics

are shown for the Mobile sequence. It clearly shows

that for the SKIP mode the value of S

Amp

is below

5000. For large block sizes, S

Amp

is generally be-

low 30000. For smaller block sizes, this value is sig-

niﬁcantly large. In the proposed algorithm, weights

are assigned based on the values of S

Amp

which are

obtained from the study of the histogram characteris-

tics for different sequences. Some ranges are deﬁned.

Lower weight is given for the MB if S

Amp

is low and

higher weight for larger values of S

Amp

. The weights

assigned are given in Table 3.

Table 3: Homogeneous MB Weights Hom

Amp

< 5000 5001-15000 15001-20000 20001-25000 > 25000

Range

Hom

0 1 2 3 4

2.3 Determination of Predicted MV

For a MB, the predicted MVs (pmv) are obtained

from the MVs of the neighboring MBs (JVT.G050r1,

2003). Motion vectors for neighboring partitions are

often highly correlated. Thus the calculated pmvs

give a good indication of the amount of possible mo-

tion in the MB. Higher pmvs signal the probability

of higher motion for the MB and vice versa. In the

proposed paper, a pmv weight (pmv

) is introduced

where weights are assigned for an MB depending

IMPROVED INTER MODE DECISION FOR H.264/AVC USING WEIGHTED PREDICTION

1 2 3 4 5 6 7 8 9 10

1000

2000

3000

4000

5000

6000

7000

MBs encoded in SKIP mode

Amplitude of Histogram

(a) SKIP mode.

0 50 100 150

5000

10000

15000

20000

25000

30000

MBs encoded in P16x16, P16x8 and P8x16 modes

Amplitude of Histogram

(b) P

16x16

16x8

8x16

mode.

0 50 100 150 200

20000

40000

60000

80000

100000

120000

MBs encoded in P8x8, P8x4 and P4x8 and P4x4 modes

Amplitude of Histogram

8x8

8x4

4x8

4x4

mode.

Figure 4: Histogram Amplitude of Frame 2 of Mobile sequences for different encoding modes.

upon the pmv for the MB. For low values of pmvs

smaller weight is assigned and for higher values larger

weight is assigned and is given in Table 4 where pmv

and pmv

are predicted MVs in in the horizontal and

vertical directions respectively.

Table 4: Predicted MV Weights pmv

max|pmv

| or max|pmv

| 0 1-2 3-4 5 > 5

or both

pmv

0 1 2 3 4

2.4 Determination of Neighboring

Mode for MB

It is observed that the modes of the neighboring MBs

are often correlated.

C D

Mode Value Assigned

SKIP 0

P16x16 1

P16x8 2

P8x16 3

P8x8 4

P8x4 5

P4x8 6

P4x4 7

INTRA 8

Figure 5: Neighboring MBs: C is the current MB.

Referring to Fig.5, let C be the current MB and A, B

and D be the neighboring MBs that have already been

encoded. In the proposed algorithm, the likely mode

for C is predicted from the modes of A, B and D and

the following relation is used

Neigh

MODE

= median{A

MODE

, B

MODE

, D

MODE

}

Depending upon the value of Neigh

MODE

for a MB,

a neighboring mode weight (NM

) is deﬁned. If

Neigh

MODE

for the MB indicate large block size par-

tition then smaller weight is assigned to NM

and

higher weight is assigned if block partition is small.

The NM

for the MB is as given in the Table 5.

Table 5: Neighboring Mode Weights NM

SKIP P

16x16

16x8

8x8

4x4

Neigh

MODE

8x16

8x4

4x8

INTRA

0 1 2 3 4

2.5 Overall Algorithm

For each MB, the different weights are ﬁrst deter-

mined and the Total

is obtained for each MB.

Total

= [

Hom

pmv

]

From the weights in Total

, a ﬁnal weight Final

is obtained for each MB as follows:

• If at least three weights in Total

are equal to say

x where x=0,1,2,3,4 then

Final

=x;

• If any two weights in Total

are equal then

Final

=ceil (median (Total

));

• If all the weights in Total

are unequal then

Final

=max(Total

);

Based on this value of Final

, the decision on the

encoding mode is taken. A low value of Final

for

a MB suggest that the MB is homogeneous with little

motion and will be encoded with larger block size. A

high low value of Final

indicate higher motion and

complexity and will be encoded using smaller block

sizes. The mode selection for each MB is given in

Table 6.

Table 6: Final Encoding Modes for MBs.

Final

0 1 2 3 4

Mode SKIP P

16x16

16x8

8x8

4x4

8x16

8x4

4x8

INTRA

SIGMAP 2011 - International Conference on Signal Processing and Multimedia Applications

3 RESULTS AND DISCUSSIONS

This section compares the results of the proposed al-

gorithm with the previously reported Wu. et al.’s al-

gorithm (Wu et al., 2005). Results are presented as

improvements over the standard H.264/AVC bench-

mark JM12.4. The experiments were carried out

using some common video sequences of different

classes at the CIF and QCIF resolution. The conﬁgu-

ration used is the baseline proﬁle, motion search range

of ±16, sequence type IPPP and one reference frame.

Only the ﬁrst frame is intra coded and QP used are

24, 28, 32 and 36 as per the recommended simulation

conditions in (Sullivan and Bjontegaard, 2001). Com-

parisons are made in terms of distortion and percent-

age differences in rate and time taken for encoding.

To evaluate the average encoding performance over a

range of QPs, the differences in PSNR (∆PSNR) in

dB and bitrate (∆Rate (%)) are calculated according

to numerical averages between RD curves as given by

Bjontegaard (Bjontegaard, 2001).

3.1 Distortion and Compression Ratio

Comparisons

Table 7 lists the performance of the proposed algo-

rithm in comparison to JM12.4 implementation and

Wu et al.’s algorithm (Wu et al., 2005). The results

are arranged for different classes of sequences. Class

A sequences have simple motion, Class B sequences

have medium to high motion and Class C sequences

have high to complex motion. The trend in the results

shows that for the sequences, there is only a marginal

loss in the PSNR performance. There is an aver-

age 0.03 dB loss in PSNR in the proposed method.

Also the average bitrate increase is 1.5%. The results

demonstrate the effectiveness of the proposed algo-

rithm.

3.2 Comparison with JM12.4 Modes

Table 8 shows the percentage of MBs that are en-

coded (using the proposed method) in the same mode

as the JM12.4 for different QP. The results show that

the proposed method is effective as it has been able

to maintain the same ﬁnal encoding modes as the

JM12.4 to a large extent.

3.3 Computational Speedup

Table 7 shows the percentage reduction in encoding

time ∆T(%) for sequences of different classes. The

time saving obtained depends upon the type of se-

quence. An increased saving is noted for Class A se-

Table 7: Performance Comparison For Different Sequences.

Class Sequence Performance Comparison

Proposed Wu et al.’s (Wu et al., 2005)

∆PSNR ∆Rate ∆T ∆PSNR ∆Rate ∆T

(%) (%) (%) (%)

CIF News -0.05 2.29 88.42 0.02 2.01 39.23

Class A MaD 0.02 2.86 81.38 0.01 0.85 43.21

Container 0.01 1.32 60.33 0.05 1.55 46.18

Hall 0.09 0.40 76.84 0.06 0.90 34.67

Foreman 0.08 0.82 72.79 0.07 1.22 34.90

Class B Coastguard -0.05 0.66 65.07 0.05 0.54 26.30

Ice 0.05 3.14 85.17 0.07 1.29 45.68

Harbour 0.06 1.88 65.37 0.05 1.10 21.56

Flower 0.08 0.19 76.43 0.05 2.98 36.85

Class C Stefan 0.08 1.59 70.59 0.02 1.46 32.25

Tempete 0.02 0.82 71.88 0.09 1.02 27.21

Mobile 0.04 1.88 62.41 0.09 1.69 12.35

Average 0.03 1.48 73.05 0.05 1.38 33.36

QCIF Claire -0.01 -2.81 90.69 -0.01 -0.95 47.35

Class A MissAmerica -0.07 -1.40 89.87 0.02 0.91 48.23

Suzie 0.04 1.21 77.73 0.05 0.49 42.91

Foreman 0.08 1.04 67.73 0.04 1.29 30.25

Class B Silent 0.03 2.40 82.46 0.09 0.79 42.62

Crew 0.08 2.17 63.86 0.05 1.65 19.64

Football 0.10 1.50 65.30 0.05 1.86 32.42

Class C Mobile 0.14 1.62 28.09 0.07 1.50 15.32

Soccer 0.04 1.17 63.67 0.03 3.03 20.19

Average 0.04 0.76 69.93 0.04 1.17 33.21

MaD: Mother and Daughter

∆PSNR(+/-): picture quality loss/gain measured in dB

∆Rate(+/-): bitrate increase/decrease measured as a %

∆T(+/-): encoding time saving/loss measured as a %

Table 8: % of MBs encoded in the same mode w.r.t FME.

Sequence Quantization Parameter

(CIF) 24 28 32 36

News 82.10 75.12 84.12 86.01

MaD 66.12 87.86 89.6 90.66

Container 76.50 86.01 89.30 89.15

Hall 59.11 68.69 72.45 83.33

Foreman 46.11 58.71 64.90 76.23

Coastguard 62.33 67.83 78.84 79.47

Ice 67.42 73.99 71.97 72.98

Harbour 46.21 52.66 55.08 63.33

Flower 69.70 70.71 69.44 72.14

Stefan 22.73 57.32 66.31 68.99

Tempete 52.27 59.75 68.23 70.43

Mobile 29.80 34.34 49.39 66.19

quences where for some sequences 90% time saving

is noted whereas time saving obtained for Class C se-

quences is comparatively low. This is due to the fact

that Class A sequences have low motion complexity

and hence a large number of MBs get encoded with

larger block sizes. The saving in time is achieved as

the decision on the ﬁnal mode for encoding is taken

prior to the ME and for each MB at the most only

three modes are searched. However,for all sequences,

the proposed algorithm exhibits a good computational

saving regardless of the QP setting.

4 CONCLUSIONS

In this paper, an improved mode decision algorithm

for H.264/AVC video coding standard has been

proposed based on the weights assigned for different

characteristics of the MB. Stationarity based weights

are obtained from frame difference residuals. Ho-

IMPROVED INTER MODE DECISION FOR H.264/AVC USING WEIGHTED PREDICTION

mogeneity based weights are obtained from edge

histograms parameters. The pmv weight and the

neighboring mode weights are obtained from the cor-

relation the the MB with neighboring MBs. Results

of simulations carried out on different sequences

demonstrate that there is very little degradation of the

PSNR and the bitrate performance in the proposed

algorithm despite a large saving in encoding time

and computation. The average encoding time saving

is around 70%. The proposed method achieves

almost the same coding performance in terms of

picture quality and compression ratio as that of the

H.264/AVC standard and improves on Wu et al.’s

(Wu et al., 2005) algorithm. Hence, for a variety

of sequences with varying motion activities, the

proposed algorithm gives a consistent performance

on encoding time reduction, computational saving

and coding efﬁciency.

REFERENCES

Bjontegaard, G. (2001). Calculation of average PSNR

difference between RD-curves. ITU-T VCEG, Doc.

VCEG-M33, pages 1–2.

Ganguly, A. and Mahanta, A. (2010). Fast Mode Decision

Algorithm for H.264/AVC using Edge Characteristics

of Residue Images. In Proceedings of ACM , ICVGIP,

pages 512–517.

Huang, Y.-H., Ou, T.-S., and Chen, H. H. (2010). Fast Deci-

sion of Block Size, Prediction Mode and Intra Block

for H.264 Intra Prediction. IEEE Trans. on Circuits

and System for Video Technology, 20(8):1122–1132.

Jing, X. and Chau, L. P. (2004). Fast Approach for

H.264 Inter Mode Decision. Electronic Letters,

40(17):1050–1052.

JVT.G050r1 (2003). Joint Video Team of ITU-T and

ISO/IEC JTC 1, Draft ITU-T Recommendation and

Final Draft International Standard of Joint Video

Speciﬁcation (ITU-T Rec. H.264 — ISO/IEC 14496-

10 AVC), document JVT-G050r1 2003. Technical

Report document JVT-G050r1, Joint Video Team of

ITU-T and ISO/IEC JTC 1.

Kim, D., Kim, J., and Jeong, J. (2006). Adaptive Mac-

roblock Mode Selection for Reducing the Encoder

Complexity in H.264. In ACVIS, number 4179 in

LNCS, pages 396 – 405.

Kim, G. Y., Moon, Y. H., and Kim, J. H. (2004). An Early

Detection of All-Zero DCT Block in H.264. In Int.

Conf. Image Process (ICIP), volume 1, pages 453–

456.

Lee, Y. L., Lee, Y. K., and Park, H. (2006). A Fast Mo-

tion Vector Search Algorithm for Variable Blocks. In

ACVIS, number 4179 in LNCS, page 311 322.

Liu, Z., L.Shen, and Zhang, Z. (2009). An Efﬁcient Inter-

mode Decision Algorithm Based on Motion Homo-

geneity for H.264/AVC. IEEE Trans. on Circuits and

System for Video Technology, 19(1):128–132.

Park, I. and Capson, D. W. (2008). Improved Inter Mode

Decision Based on Residue in H.264/AVC. In Interna-

tional Conference on Multimedia & Expo, pages 709–

712.

Richardson, I. E. G. (2003). H.264 and MPEG-4 Video

Compression: Video Coding for Next-generation Mul-

timedia. John Wiley and Sons.

Shen, L., Liu, Z., Zhang, Z., and Shi, X. (2008). Fast In-

ter Mode Decisions Using Spatial Property of Motion

Field. IEEE Trans. on Multimedia, 10(6):1208–1214.

Sullivan, G. and Bjontegaard, G. (2001). Recom-

mended Simulation Common Conditions for H.26L

Coding Efﬁciency Experiments on Low-Resolution

Progressive-Scan Source Material. ITU-T VCEG,

Doc. VCEG-N81, pages 1–3.

Wang, H., Kwong, S., and Kok, C.-W. (2007). An Efﬁcient

Mode Decision Algorithm for H.264/AVC Encoding

Optimization. IEEE Trans. on Multimedia, 9(4):882–

888.

Wiegand, T., Sullivan, G. J., Bjontegaard, G., and Luthra,

A. (2003). Overview of the H.264/AVC Video Cod-

ing Standard. IEEE Trans. on Circuits and System for

Video Technology, 13(7):560–576.

Wu, D., Pan, F., Lim, K. P., Wu, S., Li, Z. G., Lin, X., Ra-

hardja, S., and Ko, C. C. (2005). Fast Intermode De-

cision in H.264/AVC Video Coding. IEEE Trans. on

Circuits and System for Video Technology, 15(6):953–

958.

Zeng, H., Cai, C., and Ma, K. (2009). Fast Mode Decision

for H.264/AVC Based on Macroblock Motion Activ-

ity. IEEE Trans. on Circuits and System for Video

Technology, 19(1):1–11.

SIGMAP 2011 - International Conference on Signal Processing and Multimedia Applications