SMOOTHED REFERENCE PREDICTION FOR IMPROVING

SINGLE-LOOP DECODING PERFORMANCE OF H.264/AVC

SCALABLE EXTENSION

So-Young Kim and Woo-Jin Han

DM R&D Center, Samsung Electronics 416, Maetan3-dong, Youngtong-gu, Youngtong-dong

Kyounggi-do 442-742, Korea

Keywords: Video coding, video signal processing, smoothing methods, digital filters and prediction methods.

Abstract: It is well-known that multi-layer extension of H.264/AVC shows good spatial scalability performance

mainly due to its efficient inter-layer prediction techniques. Although single-loop decoding is a kind of

technique to reduce the decoder-side computational complexity by performing only one motion

compensation to decode multi-layer data, its limited use of inter-layer prediction sometimes degrades the

performance especially for fast-motion sequences. In this paper, smoothed reference prediction technique is

proposed to improve the single-loop decoding performance by replacing base-layer information with

current-layer information and simple block-based smoothing function. Experimental results show that the

proposed method can improve the coding efficiency with all benefits of single-loop decoding mode. In

addition, the proposed method was adopted to scalable extension of H.264/AVC standard Working Draft.

1 INTRODUCTION

Scalable video coding has received considerable

attention for many multimedia applications in terms

of both coding algorithms and standard activites.

Among many technologies used in scalable video

coders, scalable extension of H.264/AVC (Schwarz

and Hinz, 2004., Schwarz and Marpe, 2004.,

Schwarz and Marpe, 2005., Reichel and Schwarz,

2005) has been considered as one of the best

compromise between the coding efficiency and the

excellent scalability features. The scalable extension

of H.264/AVC is based on a multi-layer structure to

allow the spatial, temporal and SNR scalabilities.

Generally, it is well-known that the multi-layer

has some penalty on the coding efficiency compared

to the single-layer structure mainly due to the

redundant representation of information (Schwarz

and Marpe, 2005). To minimize the multi-layer

overhead, several kinds of inter-layer prediction

techniques have been exploited in the H.264/AVC

scalable extension. More specifically, motion,

texture, and residual data can be predicted from

already coded layers. With the inter-layer prediction

techniques, it was reported that the H.264/AVC

scalable extension can provide comparable coding

efficiency compared to the current state-of-the-art

video coder (Schwarz and Marpe, 2005).

A major drawback of the multi-layer structure

lies in its heavy complexity requirements. Especially,

when the more the number of layers are, the larger

total complexity requirements is applying to the real

world applications. Single-loop decoding technique

(Schwarz and Hinz, 2005) was proposed to reduce

the decoding complexity of multi-layer structure by

allowing the inter-layer texture prediction only when

the corresponding base-layer macroblock is one of

the intra-type, which does not need motion

compensation. In other words, the decoder of

H.264/AVC scalable extension performs motion

compensation only once in the top-most layer even

in the bit-stream with a multiple number of layers.

Although the decoding complexity can be reduced

significantly, the restriction of the inter-layer texture

prediction sometimes shows a non-negligible

degradation up to 0.7 dB especially in the fast-

motion sequences. Figure 1 shows a general coder

structure for scalable extension of H.264/AVC with

two spatial layers. The redundant information such

as motion and texture between consecutive layers is

used for inter-layer prediction. Inter-layer texture

prediction and residual prediction are explited for

effective inter-layer prediction.

Kim S. and Han W. (2007).

SMOOTHED REFERENCE PREDICTION FOR IMPROVING SINGLE-LOOP DECODING PERFORMANCE OF H.264/AVC SCALABLE EXTENSION.

In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 67-71

DOI: 10.5220/0002135800670071

 SciTePress

Motion compensation

& intra prediction

Motion compensation

& intra prediction

Transform

Entropy coding

Inter-layer prediction

• intra

• motion info.

• residual

motion

texture

motion

texture

Multiplex

H.264/AVC compatible

bit-stream of base layer

Scalable

bit-stream

Motion compensation

& intra prediction

Motion compensation

& intra prediction

Transform

Entropy coding

Inter-layer prediction

• intra

• motion info.

• residual

motion

texture

motion

texture

Multiplex

H.264/AVC compatible

bit-stream of base layer

Scalable

bit-stream

Motion compensation

& intra prediction

Motion compensation

& intra prediction

Transform

Entropy coding

Inter-layer prediction

• intra

• motion info.

• residual

motion

texture

motion

texture

Multiplex

H.264/AVC compatible

bit-stream of base layer

Scalable

bit-stream

Figure 1: A coder structure for the scalable extension of

H.264/AVC.

In this paper, we compare the characteristics of the

inter-layer texture and residual prediction techniques

and propose a modified version of inter-layer

residual prediction to compensate the coding penalty

from the restriction of the inter-layer texture

prediction. This smoothed reference prediction

technique was adopted to Working Draft in scalable

extension of H.264/AVC standardIt will be shown

that the major coding penalty of the single-loop

decoding results from the blocking artifacts due to

the mismatch between two different layers and

adaptive usage of simple smoothing function can

help to reduce this effect.

Section 2 describes the basic concept of the inter-

layer texture and residual predictions, Section 3

presents the proposed smoothed reference prediction

technique. Experimental results and our conclusions

arising from our work are depicted in Section 4 and

section 5.

2 INTER-LAYER TEXTURE AND

RESIDUAL PREDICTIONS

Inter-layer texture prediction exploits the fact that

the decoded texture of the lower layer is generally

similar to the corresponding current texture. It uses

the decoded macroblock of the lower layer as a

predictor of the current macroblock. Only the

difference between them is coded and transmitted as

][

bcT

OUOR

−

(1)

where R

is the residual signal using the inter-layer

texture prediction. O

and O

are original signals of

the current and lower layers, respectively. The

upsampling filter, U(⋅), is optioally applied when the

spatial resolutions of two layers are different.

On the contrary, inter-layer residual prediction

exploits redundancy of the residual signals between

two layers. Thus, the decoded residual signal of the

lower-layer is used as a predictor of the current

residual signal.

(a) (b)

Figure 2: Visual example of (a) single-loop decoding and

(b) multi-loop decoding.

The rationale behind the inter-layer residual

prediction lies in the similarity of the motion fields

including both reference frames and motion vectors

in multi-layer structure. In that case, the resultant

residual signals of two layers are also similar thus

the inter-layer residual prediction can improve the

coding efficiency. The residual signal of the inter-

layer residual prediction, R

can be defined as

][

bcR

RURR

−

(2)

where Rc is the residual signal of the current layer

without inter-layer prediction and Rb is the

reconstructed residual signal of the lower layer,

respectively. It should be noted that the actual

implementation of the upsampling filter U is

different to (1) and the same notation is used for the

simplicity's sake.

Although the inter-layer texture and residual

predictions seem to be very different in nature, the

following section shows some similarities between

two inter-layer prediction techniques and the inter-

layer residual prediction can be modified to improve

the performance in the regions where the inter-layer

texture prediction is not allowed.

3 SMOOTHED REFERENCE

PREDICTION

In single-loop decoding mode, the inter-layer texture

prediction is not allowed if the lower-layer

macroblock is coded using the motion compensation

process, that is the intra mode macroblocks of lower

layers are reconstructed only at a decoder. In this

case, inter-layer residual prediction or inter

prediction not considering lower layer information is

used instead of inter-layer texture prediction, which

sometimes makes strong high-pass artifacts as

shown in Figure 2. Figure 3 depicts well how many

macroblocks in enhancement layer are coded by

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

Multi-loop decoding – optimal selection

Single-loop decoding – suboptimal selection

Multi-loop decoding – optimal selection

Single-loop decoding – suboptimal selection

Figure 3: The intra and inter coded macroblocks in multi-

loop decoding and single-loop decoding mode(green parts

are the inter coded macroblocks).

normal intra prediction in H.264 and inter-layer

texture prediction, and the green means the

macroblocks coded by inter mode. It is shown that

the restriction not to use inter coded macroblock in

lower layer for inter-layer texture prediction in

single-loop decoding reduces the ratio of intra coded

macroblocks even where intra mode can give the

best R-D performance.

To understand the situation further, (1) can be

rewritten as

][

bbcT

RPDUOR +⋅−=

(3)

where D(

⋅

) is a in-loop deblocking operator defined

in H.264/AVC. Similarly, (2) can be rewritten as

[]

()

bccR

RUPOR +−=

(4)

Comparing (3) and (4) shows two differences:

the first one is the temporal predictor. The inter-

layer texture prediction uses the temporal predictor

of the base-layer and the inter-layer residual

prediction uses that of the current-layer. In addition,

the inter-layer texture prediction exploits the in-loop

deblocking operator as well as the upsampling

operation on the reconstructed signal.

Figure 4: Four predictors using (a) inter prediction (b)

inter-layer residual prediction (c) inter-layer texture

prediction (d) smoothed reference prediction.

The first difference which uses the base-layer

temporal prediction cannot be fixed due to the

single-loop constraints. On the other hand, the

second difference related to the deblocking and

upsampling filters can be minimized if we apply

some filters to (4). The main idea of the proposed

method is to apply the suitable filter to the inter-

layer residual prediction to give the inter-layer

texture prediction benefits.

The most straightforward choice of the filter

may be the upsampling and in-loop deblocking

filters used in (3). It gives

[]

()

bccS

RUPMDUOR +⋅

⋅

−

(5)

where R

is a new residual signal and M(

⋅

) is a

downsampling filter used for generating the lower

resolution layer. However, the upsampling and

downsampling operations generally require frame-

based processing. The additional complexity due to

the resampling and deblocking filters is also

significant. Since the predictor used in the inter-

layer texture prediction can be thought as the low

pass approximate of the current layer, we use

simple bi-linear smoothing filter instead of the three

filter combinations, U

⋅

), defined as

{

}

4/2)1()(2)1()(' +++

−

nxnxnxnx

(6)

where x(n) is n-th pixel value in the signal. (6) is

applied to both horizontal and vertical direction for

all samples inside the current macroblock. For the

macroblock boundaries, only top and left

macroblock boundaries are modified to allow the

block-based processing.

We call this kind of prediction as the smoothed

reference since it can be realized as the smoothed

version of the inter-layer residual prediction. The

final equation of the proposed method is given by

[]

()

bccS

RUPSOR +

−

(7)

where S(

⋅

) is a bi-linear smoothing filter. It should be

noted that (6) requires only 3 additions and 2 shifts

thus it can be implemented very efficiently.

Furthermore, the actual complexity burden is much

less since the new prediction is only used where

single-loop constraint has significant penalty with

respect to suitable rate-optimization criteria.

Figure 4 shows four different predictors

generated by inter prediction, inter-layer residual

prediction, inter-layer texture prediction and

smoothed reference prediction, respectively. As

shown in the figure, the inter prediction and the

inter-layer residual prediction sometimes show very

strong high-pass artifacts subjectively related to the

blocking artifacts. The prediction signal by new

smoothed reference prediction is remarkably similar

(a) inter prediction.

(b) inter-layer

residual prediction.

prediction.

(d) smoothed reference

prediction.

SMOOTHED REFERENCE PREDICTION FOR IMPROVING SINGLE-LOOP DECODING PERFORMANCE OF

H.264/AVC SCALABLE EXTENSION

to that of the inter-layer texture prediction, which

justifies the reason to use the smoothing filter.

Table 1 shows a comparative result of the ratio

of macroblock modes for single-loop, multi-loop and

single-loop with smoothed reference prediction.

As shown in the table, in the single-loop decoding

mode, the ratio of inter-layer texture prediction is

significantly small whereas that of inter-layer

residual prediction is large. All other macroblock

types are used with similar relative ratios. It

indicates that the inter-layer residual prediction is

used for most macroblocks that violates the single-

loop constraint instead of the inter-layer texture

prediction. When the smoothed reference prediction

is used, the relative distribution of macroblock types

is very similar to the multi-loop case. This is

because the smoothed reference prediction is used

for the macroblocks suffering from single-loop

constraint instead of the inter-layer residual

prediction.

Table 1: Macroblock types for Football CIF.

Prediction type Single-loop Multi-loop

Smoothed

reference

Inter-layer

texture

(+smoothed

reference)

12% 29% 28%

Directional intra 1% 2% 1%

Inter-layer

residual

42% 24% 26%

o inter-layer

prediction

45% 45% 45%

4 EXPERIMENTAL RESULTS

4.1 Experimental Condition

The proposed method was implemented in the

reference software of H.264/AVC scalable extension

developed in the MPEG/JVT as an extension of

H.264/AVC. Football CIF and Soccer 4CIF

sequences were used for the performance

verification since these sequences have relatively

large coding penalty when single-loop decoding is

used. For the scalability test points, we use 5 layers

for Football CIF and 6 layers for Soccer 4CIF as

defined in Table 2.

Bit-rates (kbps)

0 200 400 600 800 1000 1200

Y-PSNR (dB)

Single-loop

Multi-loop

Single-loop + smoothed reference

(a) Football sequence.

Bit-rates (kbps)

0 500 1000 1500 2000 2500 3000 3500

Y-PSNR (dB)

Single-loop

Multi-loop

Single-loop + smoothed reference

(b) Soccer sequence.

Figure 5: PSNR curves for Football and Soccer sequence.

Table 2: Scalability test points used in the experiments

sequence.

4.2 PSNR Results

Figure 5 shows PSNR graphs for Football and

Soccer sequences. In Football sequence, the

maximum PSNR degradation from single-loop

decoding constraint is about 0.5 dB. With the

proposed method, the performance gap is reduced to

0.2 dB. The performance trend is similar in

soccer sequence. The performance gap is reduced by

a half with the proposed method. It indicates that

there is further room to be improved in the single-

loop decoding with smoothed reference prediction.

The proposed method actually reduces the

performance penalty due to the single-loop decoding

Football Soccer

QCIF@7.5Hz, 128kbps QCIF@15Hz, 96kbps

QCIF@15Hz, 192kbps QCIF@15Hz, 192kbps

CIF@15Hz, 384kbps CIF@30Hz, 384kbps

CIF@15Hz, 512kbps CIF@30Hz, 768kbps

CIF@30Hz, 1024kbps 4CIF@30Hz, 1536kbps

4CIF@60Hz, 3072kbps

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

constraint while still allowing the single-loop

decoding concepts, which means much lower

computational complexity compared to the multi-

loop decoding case.

4.3 Visual Examples

Figure 7 and 8 show visual examples of the single-

loop decoding mode and the single-loop decoding

mode with the proposed method. Not only blocking

artifact but high frequency component of prediction

signals is removed. Comparing Figure 3 and 6, it is

exploited that the ratio of intra coded macroblocks

by R-D cost is improved almost similar to the ratio

in multi-loop decoding mode.

Multi-loop decoding – optimal selection

Single-loop decoding – smoothed reference

Multi-loop decoding – optimal selection

Single-loop decoding – smoothed reference

Figure 6: The intra and inter coded macroblocks in multi-

loop decoding and single-loop decoding with smoothed

reference prediction (green parts are the inter coded

macroblocks.

5 CONCLUSION

In this paper, we proposed a new prediction

technique, which modified the inter-layer residual

prediction to compensate the performance penalty

of the single-loop decoding by adding block-based

bi-linear smoothing function to the inter-layer

residual prediction process.

(a) Single-loop decoding mode.

(b) Single-loop decoding with smoothed

reference prediction.

Figure 7: Visual examples: (a) single-loop decoding mode

(b) single-loop decoding with smoothed reference

prediction for 7

frame of football.

From experimental results, it was shown that the

performance penalty due to the single-loop decoding

constraint could be reduced by the proposed method

and the improvements of subjective quality was also

meaningful. In addition, we expierimented several

filters with three filter coefficients as a smoothing

filter, however no filters were superior to bi-linear

smoothing filter. Finally, the proposed technique

was adopted the sclable extension of H.264/AVC

standard Working Draft.

a) Single-loop decoding mode.

(b) Single-loop decoding with smoothed

reference prediction.

Figure 8: Visual examples: (a) single-loop decoding mode

(b) single-loop decoding with smoothed reference

prediction for 36

frame of foreman.

REFERENCES

Schwarz, H., Hinz, T., Kirchhoffer, H., Marpe, D.,

Wiegand, T., Oct, 2004. Technical Description of the

HHI proposal for SVC CE1. ISO/IEC

JTC1/SC29/WG11, M11244.

Schwarz, H., Marpe, D., Wiegand, T., Dec, 2004. MCTF

and Scalability Extension of H.264/AVC. Proc. of

PCS, San Francisco, CA, USA.

Schwarz, H., Marpe, D., Schierl, T., Wiegand, T., July,

2005. Combined Scalability Support for the scalable

Extensions of H.264/AVC. Proc. of ICME

Reichel, J., Schwarz, H., Wien, M(eds.)., Jan, 2005.

Scalable Video Coding – Working Draft 1.

Joint Video Team (JVT), Doc. JVT-N020, Hong Kong, CN

Schwarz, H., Hinz, T., Marpe, D., Wiegand, T., Sep, 2005.

Constrained Inter-Layer Prediction for Single-Loop

Decoding in Spatial Scalability. Proc. of ICIP,

Genova, Italy.

W. J. Han and S. Y. Kim, Jul, 2005. Smoothed reference

prediction for single-loop decoding. Joint Video Team

(JVT), Doc. JVTO085, Poznan, Poland.

W. J. Han and S. Y. Kim, Jan, 2006. Modified IntraBL

design using smoothed reference. Joint Video Team

(JVT), Doc. JVTR091, Bangkok, Thailand.

SMOOTHED REFERENCE PREDICTION FOR IMPROVING SINGLE-LOOP DECODING PERFORMANCE OF

H.264/AVC SCALABLE EXTENSION