2009 IEEE Symposium on Industrial Electronics and Applications(ISIEA 2009),October 4-6,2009,Kuala Lumpur,Malaysia

A Re-quantization Noise Reduction Method in MPEG-2 to H.264 Intra Transcoding Takeshi YOSHITOME, Kazuto KAMIKURA NTT Cyber Space Laboratories NTT Corporation

Nobuhiko KITAWAKI Graduated School of Systems and information Engineering University of Tsukuba

Abstract | An MPEG-2 to H.264 intra transcod-

ing method is proposed. This method uses the encoding information from an MPEG-2 stream and keeps as many DCT coecients of the original MPEG-2 bitstream as possible. Experimental results show that the proposed method improves PSNR by about 0.76-1.27 dB compared with a typical conventional method. Keywords |Transcoding; MPEG-2; H.264 I.

Transcoding with First Encoding Information

A data ow example of transcoding with rst encoding information is shown in Fig.1. For simpli cation purposes, the macro-block (MB) size is 2x2. In the rst encoding, an input MB that includes the pixel values (A, B, C, and D) is processed by orthogonal transformation. Then, the transformed DCT coecients (47, 31, 25, and 18) are quantized by a quantization step 1=10. This

978-1-4244-4683-4/09/$25.00 ©2009 IEEE

MC prediction error

DCT coefficient

Quantized DCT coefficient

Encoder A B #1 C D ME1

A’ B’ 47 31 C’ D’ DCT1 25 18 Q1

4 3 2 1

Decoder a b #1 c d

a’ b’ c’ d’

4 3 2 1

(

(frame)

(mv)

40 30 20 10

=10)

Bitstream #1

Almost same because mv, Dct-type, Q of 1st enc are reused ME2 (mv)

Encoder #2

a’ b’ DCT2 40 30 Q2 4 3 (frame) ( =10) c’ d’ 20 10 2 1 Bitstream #2

Introduction

Many kinds of video services use a video compression tool to reduce the size and the transmitting cost of the video contents. In recent years, it has become increasingly important to convert video streams from MPEG2[1] to H.264[2]. For transcoding, some papers[3, 4] have reported that the use of an MPEG-2 motion vector and compression mode information helps to reduce the processing time of H.264 re-encoding. A direct mathematical matrix conversion from the 8x8 real DCT of MPEG-2 to the 4x4 integer DCT of H.264 has also been reported[5]. We have focused our attention on the fact that the complete re-use of encoding information, such as picture type, motion vector, and macro-block type, makes it possible to suppress re-quantization noise to almost zero[6]. This noise suppression mechanism is e ective when the compression standards of the rst and second encoding are identical, such as for MPEG-2/MPEG-2 or H.264/H.264 conversion. In our work, we attempted to expand this approach to apply it to MPEG-2/H.264 transcoding. In this paper, we rst discuss the use of this noise suppression mechanism when the same compression standards are used. Next, we show the di erences between the MPEG2 and H.264 standards, analyze how they a ect the noise reduction, and propose an MPEG-2/H.264 transcoding method. Finally, we describe simulation results we obtained and compare the transcoding noises generated by our method and a conventional method. II.

Input MB

Decoder a b #2 c d

a’ b’ c’ d’

40 30 20 10

4 3 2 1

No quantization noise at 2nd quantization

Figure 1: Second encoding with rst encoding information quantization process adds quantization noise to the input MB. In the second encoding, every type of encoding information, such as MB type, motion vector, DCT mode, and quantization step, is re-used. This re-use of encoding information does not produce any signi cant di erence between motion compensation errors of the rst encoding and those of the second encoding; nor does it produce any notable di erence between the DCT coecients of the two encodings. This means that the re-quantization noise added in the second encoding phase is almost zero. If a better motion vector that could decrease the motion compensation errors (a', b', c', and d') were used in the second encoding, the re-quantization noise in the second encoding would increase. Unfortunately, there is no bit reduction because the output bitstream size of the rst and second encodings are almost the same when these encodings make use of the same compression tool. But transcoding using di erent compression tools, such as from MPEG-2 and H.264, may reduce the bitstream size because the performance of H.264's entropy coding (CABAC or CAVLC) is higher than that of MPEG-2's entropy coding (VLC). III.

Differences between MPEG-2 and H.264

As depicted in Fig.1, there is a close resemblance between the decoded images of the rst and second decoders when the compression standards of the rst and second encodings are exactly the same. To reenact the e ect of

94

Authorized licensed use limited to: NTT Yokosuka. Downloaded on January 3, 2010 at 22:24 from IEEE Xplore. Restrictions apply.

2009 IEEE Symposium on Industrial Electronics and Applications(ISIEA 2009),October 4-6,2009,Kuala Lumpur,Malaysia

this noise reduction mechanism, we must control H.264 behavior to imitate MPEG-2 behavior as much as possible when the rst encoder is MPEG-2 and the second encoder is H.264. The two encoders are similar but are not upper compatible. The di erences between them are listed in Table1. These di erences should be absorbed to recreate this noise reduction mechanism. The "o" means that H.264 can recreate the same MPEG-2 function in the table. For example, if the motion vector resolution of MPEG-2 is 0.5 pixels, that of H.264 is also 0.25 pixels. The "?" means that H.264 can process the function only in a similar but not identical way that MPEG-2 does. For example, a certain MPEG-2 quantization step cannot be expressed exactly using H.264 because the former is de ned by a linear interval and the latter is de ned by a logarithmic interval. The following four functions are not exactly the same between MPEG-2 and H.264: (1) Internal picture motion compensation, (2) Orthogonal transformation, (3) Available quantization step, and (4) Frame/Field MB speci cation. In the next section, we describe these di erences in detail and show a way to overcome them. Table 1: Major di erences between MPEG-2 and H.264

MPEG-2 H.264 H.264 substitute Block size 16x16 4x4, 4x8, 8x4, 8x8 OK 16x8 8x16,16x8,16x16 Vector resolution 0.5pix 0.25pix OK Intra prediction no yes ? Orthogonal DCT integer DCT ? Transformation Size of orthogonal 8x8 8x8(*1) OK Transformation 4x4 entropy coding VLC CAVLC OK CABAC Speci cation of yes yes(*1) OK Quantization matrix Speci cations of each MB pair MBs ? Field/FrameMB Interval of linear logarithmic ? Quantization step (*1):only High-Pro le A. Internal motion compensation Although there is no intra prediction function in MPEG-2, H.264 has such a function to increase compression eciency. Adjoining pixels in the upper, left, and upper-left blocks are referred to in this function. Because all prediction modes use the prediction in H.264, it is impossible to exactly equalize the H.264 and MPEG-2 intra MBs. The best way to imitate the latter is to use the DC prediction mode in H.264. If this is used, the orthogonal transformation will change only the DC coecient. None of the coecients except the DC one are a ected by the di erence between the MPEG-2 and H.264 predictions. Accordingly, we decided to use the DC prediction mode in our method.

Table 2: Available 1 in MPEG-2 and in H.264

1 H.264 q MPEG-2 0 scale type 1 0.625 O -0.6875 O 0.8125 OO --10.875 O O 1.125 OO --1.25 OO --1.375 1.625 OO O21.75 O 2.25 OO --2.5 O- -O-32.75 3.25 O OO O3.5 44.5 O OO -55.5 O O- O66.5 O OO -78 OOO O 910 OO OO 11 O- O12 O 13 OO O14 O 16 OO OO OO 18 20 OO OO OO 22 B.

1 H.264 q MPEG-2 0 scale type 1 24 O O 26 OO OO O28 30 32 O-- OOO O-34 36 O- OO O38 40 O- OO O42 44 O- OO O46 48 -- OO O50 52 O OO O54 56 O- OO O58 60 -- OO -62 64 O- -O68 72 OO -OO 80 88 O OO 96 104 OO -OO 112 128 OO --144 160 OO --176 208 OO --224 - -

Available quantization step size There are 51 available quantization steps in H.264. The minimum value is 0.625 and the maximum value is 224. In contrast, there are 31 available quantization steps in MPEG-2. The q scale value, which is stored in the MPEG-2 picture header, can specify the di erent mapping from the quantization code to the quantization step. Available quantization steps in MPEG-2 and in H.264 are shown in Table.2. A 1 value of "O" means there is availability. There are six cases in which MPEG-2 can specify 1 and H.264 can not when q scale type= 1. When q scale type= 0, 13 cases exist. In these cases, the nearest quantization step of H.264 is used in place of that of MPEG-2. The ratio di erence between the MPEG2 and H.264 quantization steps is 8.3% for all 13 cases when q scale type= 1. When q scale type= 0, the ratio di erence ranges from 3.2% to 8.3%. The negative effects caused by this di erence will be described in detail in Section. V. The amount of this di erence is within the acceptable level. C. Orthogonal transformation In MPEG-2 compression, the DCT, the quantization and de-quantization, and the inverse DCT are described in the following Eqs. (1), (2), and (3) : Y = HXH t (1) Yd = DequantM P EG ( QuantM P EG (Y )) (2) X^ = H t Yd H (3)

95 Authorized licensed use limited to: NTT Yokosuka. Downloaded on January 3, 2010 at 22:24 from IEEE Xplore. Restrictions apply.

2009 IEEE Symposium on Industrial Electronics and Applications(ISIEA 2009),October 4-6,2009,Kuala Lumpur,Malaysia

where X is the input image, Y is the DCT coecient, Yd is the de-quantized DCT coecient that was quantized, X^ is the decoded image, H is the transfer matrix, DequantM P EG (), QuantM P EG () are the quantization and de-quantization functions of MPEG-2, and H t is the transposed matrix of H . The (i,j) component of H is described in Eq.4 as follows: Hi;j





= i cos j + 21 i N 8



< :

q

q

; i=0

1

N

(4)

; i>0

2

N

Z = T XT

(5)

t

(6)

X^

(7)

= T t Zd T

Z0;0 = Y0;0 =

B B T =B B @

1 q 2 1 10 1 0 2 q 2 0 5

1 q 2 1 0 10 1 0 2 q 2 5

11

q2 2 0 5 1 q 2 1 0 10

a

B 3c B 2 B b B 5c B 4 T =B B a B 3c B B 4b @ 2 3c 8

a

a

5c 4 b

0

2 3c 8

0 0

a

3c 2

0 0

b

3c 4

a

C C C C A

3c 4 0 0

3c 8

b

b

2 3c 2

a

a

(8)

0

0 0

0

3c 8

b

0

r

a

0

b

2 3c 2

5c 4

b

3c 4

a

5c 4

3c 8

0

3c 4

5c 4

0 0

r

a

3c 4 0 0

2 3c 2

a

0 0

b

2 3c 2

b

0

3c 8

b

5c 4

a

0

5c 4

a

0

3c 2

C C

0

5c 4

C C

0

3c 4

C C A

b

2 3c 8

a

0

3c 2

b

bC C

aC C b

0

3c 4

1

0

2 3c 8

r

1 ; b = 1 ; c = 32 where a = (9) 8 5 289 When N=8, the (i,j) component of T is Ti;j , and from Eqs. (4) and (9), the rst and fth columns of H are equal to the same position of T shown in Eq.(10). T0;j

= H0;j

1 8

for j = 1; 2; 5; 6

r

= 18 for j = 0::7

(10)

1

1@ A 8 i2 ;j2 xi;j X

0

2

2

2

2

i 2;j 1

8 <

1 : 2

xi;j +

0

2

xi;j A

2

i ;j 2

X 0

2

2

xi;j

i 1;j 2

X

2

1

X

0

0

xi;j A

i 2;j

X 1 Z4;4 = Y4:4 = @ 8 i2 1;j2 1 xi;j X

1

X

0

1 X x Z4;0 = Y4:0 = @ 8 i2 ;j2 1 i;j

where

with N=8 is shown in Eq.(9). 0

0

for j = 0; 3; 4; 7

1 X x Z0;4 = Y0;4 = @ 8 i2 1;j2 i;j

with N=4 is shown in Eq.(8). 1 q2 2 5 1 q 2 1 10

1

q8

0

Zd = DequantH 264 ( QuantH 264 (Z ))

0

q

Let the (i,j) components of X , Y , and Z be Xi;j , Yi;j , and Zi;j and from Eqs.(1) and (5), Y0;0 , Y0;4 , Y4;0 , and Y4;4 be equal to Z0;0 , Z0;4 , Z4;0 , and Z4;4 as shown in Eq.(11): 0

In H.264, if we assume that Z is the integer DCT and T is the integer DCT matrix, DequantH 264() and QuantH 264 () are the quantization and de-quantization functions of H.264, and Zd is the de-quantized integer DCT coecient that was quantized, then the integer DCT, the quantization and de-quantization, and the inverse integer DCT are described in Eqs. (5), (6), (7) below:

T

= H4;j = :

1

where i =

T

T4;j

8 <

2

1

xi;j A

i 2;j 2

= (0; 1; 2; 3; 4; 5; 6; 7) = (0; 3; 4; 7) = (1; 2; 5; 6)

(11)

Equation11 indicates that the four components (Y0;0 ; Y0;4 ; Y4;0 ; and Y4;4 ) generated by MPEG-2's 8x8 DCT are identical to the four components (Z0;0 ; Z0;4 ; Z4;0 ; and Z4;4 ) generated by H.264's 8x8 integer DCT. To compare the other 60 components of Y and Z , an example image is processed by two consecutive stages using mathematical operations. In the rst stage, the image is processed by the DCT, the quantization and dequantization, and the IDCT shown in Eq.(1) - Eq(3). In the second stage, the output image of the rst stage is processed by following the di erent functions. 1. true-DCT/quantize/de-quantize/true-IDCT (Eq.1-Eq.3, PSNR is shown in Fig. 2) 2. 4x4int-DCT/quantize/de-quantize/4x4int-IDCT (Eq.5-7 with N=4 . PSNR is shown in Fig. 3) 3. 8x8-DCT/quantize/de-quantize/8x8int-IDCT (Eq.5-7 with N=8 . PSNR is shown in Fig. 4)

96 Authorized licensed use limited to: NTT Yokosuka. Downloaded on January 3, 2010 at 22:24 from IEEE Xplore. Restrictions apply.

2009 IEEE Symposium on Industrial Electronics and Applications(ISIEA 2009),October 4-6,2009,Kuala Lumpur,Malaysia

A

MPEG2 DCT/Q/ IQ/IDCT

B

A

50

C

MPEG2 DCT/Q/ IQ/IDCT

PSNR(dB)

PSNR(dB)

50

MPEG2 DCT/Q/ IQ/IDCT

H.264 4x4DCT/Q/ IQ/4x4IDCT

B

55 C

50

45

45

PSNR(A, C)

PSNR(A, B)

35

PSNR(A, C)

5

10

15

20

25

30

35

B

C

PSNR(B, C)

PSNR(A, C)

PSNR(A, B)

35

30 0

H.264 8x8DCT/Q/ IQ/8x8IDCT

40

PSNR(A, B)

PSNR(A, C)

35

30

MPEG2 DCT/Q/ IQ/IDCT

45

40

40

A

PSNR(dB)

55

55

30 0

5

10

15

20

25

30

35

0

5

10

15

20

25

30

35

Figure 2: PSNR of trueDCT/trueDCT Figure 3: PSNR of trueDCT/4x4DCT Figure 4: PSNR of trueDCT/8x8DCT re-encoding re-encoding re-encoding In the second stage, the quantization and dequantization without dead-zone are used because only the di erence between the orthogonal transformations of MPEG-2 and H.264 re ects the PSNR results. From Fig. 2 to Fig. 4, A means the input image, B means the decoding image of the rst encoding, C means the decoding image of the second encoding, and PSNR(A, B) means the PSNR value between A and B. Figure 2 shows that PSNR (A, B) equals PSNR (A, C). This indicates that the repetition of orthogonal transformations of MPEG-2 does not decrease the PSNR, as mentioned in Section.II. We omitted PSNR (B, C) from Fig. 2 because the requantization noise showed the lowest value and PSNR (B, C) rose to over 200 dB. When the second encoding is the 4x4 integer DCT of H.264 (Eq.(5) - Eq.(7),N=4), the PSNR (A, B) is almost equal to the PSNR (B, C) shown in Fig.3. This indicates the amount of quantization noise of the second encoding is almost the same that of the rst encoding. The total quantization noise doubles and the nal PSNR (A, C) is lower than PSNR (A, B) by about 3 dB. On the other hand, when the second encoding is the 8x8 integer DCT of H.264 (Eq.(5) - Eq.(7),N=8), PSNR (A, B) < PSNR (B, C) and the di erence becomes larger as 1 becomes larger, as shown in Fig.4. It appears that a large 1 decreases the PSNR reduction caused by the di erence between the 8x8 DCT and the real 8x8 DCT. The di erences between PSNR (A, B) and PSNR (A, C) are 2.0, 1.6, 1.0, 0.6, and 0.2 when 1 = 2, 4, 8, 16, and 32. Incidentally, PSNR (B, C) is 43.16 dB when 1=32; the PSNR of each coecient in this case is shown in Eq.(12). All four components (0,0), (4,0), (0,4), and (4,0) are over 200 dB. This result substantiates Eq. (11). The PSNR value of the other 60 components ranges from 35 to 52 dB. These components determine the total noise level. P SNR of F req:domain1=32 (B; C ) = 0211:3

47 3 47 1 40 8 212 1 B 47 1 46 5 47 1 43 9 54 2 B 46 3 46 9 48 0 45 4 52 3 B 39 2 43 9 45 2 45 2 50 5 B B213 2 55 5 55 9 50 8 217 5 B 39 8 44 1 45 0 44 9 49 6 @ 36 6 40 1 41 8 43 4 45 2 34 8 40 0 42 4 43 9 47 1 D. DCT type speci cation :

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

:

39 6 44 1 45 5 45 9 49 1 48 2 46 2 47 9 :

:

:

:

:

:

:

:

37 3 39 8 41 8 43 0 42 8 45 9 48 0 49 8 :

:

:

:

:

:

:

:

35 61 39 9C 43 0C 44 9C C (12) 47 0C 48 5C 50 5A 52 9 :

:

:

:

:

:

:

:

Although a DCT type can be speci ed for each MB independently in MPEG-2, H.264 speci es the DCT type

Table 3: Percentage of the DCT type agreement between upper and lower MB

No 18 10 14 15 16 19 22 23 28 46 47

Scene Cognac and Fruit Walk through the Sq. Streetcar Yacht Harbor Yachting Whale Show Opening Ceremony Marching in Green Leaves Summertime Tanning Sprinkling Picture Cuts Ave

A 84.9 83.8 84.4 84.3 89.2 87.3 81.8 82.5 92.5 89.6 85.8 86.0 86.0

MPEG-2 B C encoder D 84.5 84.0 89.7 85.9 82.8 89.7 85.3 82.7 82.0 85.5 84.8 82.9 84.3 83.3 89.6 88.2 87.6 88.1 86.7 80.6 82.3 88.9 82.5 83.6 93.9 89.1 91.5 88.8 87.9 89.7 85.5 84.1 86.0 86.8 84.6 89.6 87.0 86.7 86.9

E 84.8 83.5 81.2 82.7 84.7 87.6 82.5 81.9 91.1 88.0 85.0 84.3 84.8

for each pair of MBs that adjoin each other above and below. When an upper MB is speci ed as the frame DCT and its lower MB is speci ed as the eld DCT in MPEG2, the DCT type of one MB is di erent from its DCT type speci ed in H.264, whether the DCT type of the pair of MBs is frame or eld type in H.264. When the DCT type of MPEG-2 is di erent from that of H.264, the 8x8 pixel group of MPEG-2 DCT does not agree with those of H.264 DCT. The quantization noise reduction in the second encoder, mentioned in Section.II, decreases when many disagreements are generated. We therefore measured the disagreement quantity using the MPEG-2 bit-streams generated by ve commercial encoders. The percentage for which the upper MB has the same DCT type as the lower MB is shown in Table.3. The input images are 12 sequences included in the standard HDTV sequences[7]. The average of the DCT type agreement between upper and lower MBs is about 86%. For the remaining 14%, single agreement is obtained between MPEG-2 and H.264 for upper or lower MBs, so the overall MB agreement obtained is 93%. In other words, the quantization noise reduction is not e ective for only 7% of the MBs. IV.

Proposed Method

The proposed MPEG-2 to H.264 conversion rules are listed in Table.5. The high H.264 pro le enables the use of an 8x8 integer DCT that has the same DCT size of MPEG-2. The quantization matrix of H.264 is set to the same coecient as that of MPEG-2's quantization matrix. The quantization step is selected to be the value

97 Authorized licensed use limited to: NTT Yokosuka. Downloaded on January 3, 2010 at 22:24 from IEEE Xplore. Restrictions apply.

2009 IEEE Symposium on Industrial Electronics and Applications(ISIEA 2009),October 4-6,2009,Kuala Lumpur,Malaysia 34.0

Table 4: DCT type mapping between MPEG-2/H.264

H.264 transcoding r R i up low up low up low M M M M M M B B B B B B Frame Frame Frame I8 I8 I8 I8 I8 I8 O O O O O O Frame Frame Field I8 I8 I8 re I8 I8 O X O X X O Frame Frame Field I8 I8 re I8 I8 I8 X O X O O X Field Field Field I8 I8 I8 I8 I8 I8 O O O O O O re: re-judge I8: I8x8

I

16 16

=16 14 18

32.5

14

13

up low M M B B Frame I8 I8 O O Field re I8 X O (a) Cheer Leader (b) Bus Field I8 re Figure 5: PSNR comparison of four proposed methods O X Field I8 I8 used the modi ed jm12.1, which can input the rst encodO O 18

16

13

32.5

16

18

20

32.0

32.0

16

20

High Pro le I8x8 DC the same that MPEG-2's matrix nearest of MPEG-2's step Field (upper and lower MB were Field DCT in MPEG-2) Frame ( otherwise )

Simulation

A. Conditions We used experimental simulation to compare our proposed method with a conventional method. The conventional method we used is the ordinary MPEG-2/H.264 transcoding method that does not use the rst encoding information, and does intra MB mode selection. The simulation conditions were as follows. TM5 was used for the rst MPEG-2 encoding under the following conditions: The quantization step 1 was xed, the q scale type was equal to one, and the quantization matrix was the MPEG2 default setting matrix. The jm12.1 was used for the conventional second H.264 encoding under the following conditions. The pro le was high-pro le, the RD optimization was o , the UseHadamrd was on, and the quantization matrix was the H.264 default setting matrix. The conventional second encoding had no restrictions with regard to selecting MB mode and intra prediction mode, so its output bit-streams included the all-MB mode and DC prediction, and also included all available DCT sizes. We

26

R

i

I conv.

240

method r

22

31.0

R

26

31.0

30.5 220

20

22

method r

22

28

18

31.5

20

22

31.5

nearest to that of MPEG-2's quantization step. The intra prediction mode is set to the DC mode to imitate the MPEG-2 standard. The H.264 standard can select one of the intra MB modes from among I4x4, I8x8 and I16x16 and each DCT matrix size is either 4x4, 8x8, or 4x4. The proposed method uses the I8x8 mode because its DCT matrix size is the same as that of MPEG-2. The DCT type mapping from MPEG-2 to H.264 is the method r previously shown in Table.4 in Section.III(D). V.

bus1q16 33.0

16

33.0

Table 5: Proposed MPEG-2/H.264 transcoding method

Pro le MB type Intra pred. mode Quantization matrix Quantization step MB pair

=16

PSNR (dB)

method MB position pair MBtype DCT coe . Frame Field pair MBtype DCT coe . Field Frame pair MBtype DCT coe . Field Field pair MBtype DCT coe . O : good preservation X : altered

16 33.5

PSNR (dB)

First MPEG2 upper lower MB MB Frame Frame

33.5 16

cherq16

260

280 MacroBlock Bits (bit)

300

320

i

30.5

30.0 180

I conv.

190

200

210

220 230 MacroBlock Bits (bit)

240

250

260

270

ing information, for our proposed transcoding method. The quantization step 1 was xed for the simulations of both the conventional and proposed methods. B. Results The PSNR of the conventional method and the proposed method are shown in Fig.6(a)-6(c). The quantization step of the rst encoding, 11 , was set to 16, 26, or 36. That of the second encoding, 12 , varied from 2 to 64. The x-axis means PSNR and the y-axis means macroblock bits. The solid line indicates the proposed method and the dotted line indicates the conventional method. These gures show that the proposed method has two peaks when 12 =11 and 12 =0:511 . At these points, the PSNR of the proposed method is higher than that of the conventional method from about 1.0-1.9 dB. The PSNR of the proposed method drops when 0:511 < 12 < 11 and 11 < 12 < 2:011 . This PSNR drop is similar to that reported in MPEG-2/MPEG-2 transcoding. Figure.6(a) shows that the PSNR of the proposed method is higher than that of the conventional method when 12 <22. In Fig.6(b), the PSNR of the proposed method is higher than that of the conventional method when 12 <36. In Fig.6(c), the PSNR of the proposed method is higher than that of the conventional method when 12 < 52. As we mentioned in Sec.III(B), if the rst encoding MPEG-2 uses the quantization step 11 that is not allowed in H.264, an approximate quantization step must be used in H.264 transcoding. However, the di erence between the true value and the approximation value is as high as 8.3%. The range in which the proposed method is superior to the conventional method exceeds this 8.3% error in Fig.6(a) - (c). This indicates that the proposed method is superior to the conventional method even if the approximate quantization step is used in H.264 transcoding. The compression ratio and PSNR of the proposed method when the quantization step 12 is equal to 11 are listed in Table.6. The conventional PSNR column is the interpolation value when the conventional method generates the same bitstream in terms of size. These interpolation values are calculated using the simulation results of the conventional method. The PSNR of the proposed method is about 0.76-1.27 dB better than that of the conventional method when the quantization

98 Authorized licensed use limited to: NTT Yokosuka. Downloaded on January 3, 2010 at 22:24 from IEEE Xplore. Restrictions apply.

2009 IEEE Symposium on Industrial Electronics and Applications(ISIEA 2009),October 4-6,2009,Kuala Lumpur,Malaysia

Table 7: Lowest comp. ratio when proposed method is superior to conventional method

Table 6: Comp. ratio and PSNR of proposed method

Input MPEG-2 image 1 PSNR MB size (dB) (kbit) Cheer 16 31.37 324.679 Leader 26 29.84 238.483 36 28.59 191.898 16 33.62 278.037 Bus 26 30.90 199.317 36 29.14 157.199 Mobile 16 30.78 455.700 and 26 27.85 329.502 Calendar 36 25.91 257.195

Proposed transcoder Conv. PSNR MB size comp. PSNR PSNR di . (kbit) ratio (dB) (dB) (dB) 297.031 0.915 31.04 30.28 0.76 205.979 0.864 29.56 28.63 0.93 154.483 0.805 28.39 27.40 0.99 241.855 0.870 33.14 31.87 1.27 159.621 0.801 30.58 29.63 0.95 115.591 0.735 28.92 28.10 0.82 424.557 0.932 30.05 29.02 1.03 295.313 0.896 27.32 26.28 1.04 219.698 0.854 25.53 24.60 0.93

32

1 lower limit size 16 233.98 26 154.79 36 16 194.96 Bus 26 140.07 36 100.47 Mobile 16 366.62 and 26 252.94 Calendar 36 188.32

30

MPEG-2(

cherq26

=16, 324.7bit) 16

31

18

14 13

11 10

9

8

MPEG-2(

28

14 20 18

26

13

29 PSNR (dB)

29

28

26

40

36

28

27

Proposed (method r)

160

200

240 280 320 MacroBlock Bits

360

400

18 32

28 26

440

160

200 240 MacroBlock Bits

280

14

13

22

28 44

52

Proposed (method r)

Conventional 26 120

16

20

Proposed (method r)

40

Conventional 27 120

=36, 191.1bit)

36

27

28

MPEG-2(

cherq36

16

32

20

MPEG comp. ratio size limit 324.679 0.721 238.483 0.649 191.898 278.037 0.701 199.317 0.703 157.199 0.639 455.700 0.805 329.502 0.768 257.195 0.732

29

11 10

22

22

30

=26, 238.4bit)

7

PSNR (dB)

cherq16

PSNR (dB)

image Cheer Leader

Conventional 320

26 120

160

200

240

MacroBlock Bits

1 = 16 1 = 26 1 = 36 Figure 6: PSNR of proposed method and conventional method (Cheer Leader)

step 12 is equal to 11 . The re-compression ratio of the H.264 bitstream size to the MPEG-2 bitstream size is from 0.801 to 0.932. As 12 becomes larger than a certain value, the PSNR of the proposed method becomes lower than that of the conventional method. Although the amount of this limit value depends on the input images we tested, the proposed method is superior to the conventional method for our tested simulations when the re-compression is over 0.805. VI.

Conclusion

We proposed a re-quantization noise reduction method for MPEG-2 to H.264 intra transcoding. This method uses encoding information from the MPEG-2 stream and keeps as many DCT coecients of the original MPEG-2 bitstream as possible. We analyzed the di erence between MPEG-2 and H.264 regarding the intra prediction and quantization steps and orthogonal transformation to reduce the re-quantization noise in H.264. Experimental results showed that the proposed method improves PSNR by about 0.76-1.27 dB compared with the conventional method. The advantages of the proposed method are that it not only results in high PSNR but also that it does not require complex calculation to select an MB mode and an intra prediction mode.

References

ciated audio information\, Nov. 1994. [2] \Draft ITU-T recommendation and nal draft international standard of joint video speci cation (ITU-T Rec.H.264/ISO/IEC 14486-10 AVC\, in Joint Video Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, JVT-G050,2003 [3] Xiaoan Lu, et al "Fast mode decision and motion estimation for H.264 with a focus on MPEG-2/H.264 transcoding" in Proc. of ISCAS 2005, 23-26 May 2005, pp1246-1249, 2 [4] Gao Chen, et al "Ecient block size selection for MPEG-2 to H.264 transcoding" Proc. of the 12th annual ACM international conference on Multimedia [5] Joo-Kyong Lee, et al "Quantization/DCT Conversion Scheme for DCT-domain MPEG-2 to H.264/AVC Transcoding" IEICE Trans. Commun., E87-B, 7 (July 2004) [6] P. Guilotel, et al "Adaptive Encoders: The New Generation of MPEG-2 Encoders" SMPTE journal (April 2000) [7] High vision digital standard image data http://www.ite.or.jp/shippan/test chart index.html

[1] ISO/IEC IS 13818-2, ITU-T Recommendation H.262, \Generic coding of moving pictures and asso-

99 Authorized licensed use limited to: NTT Yokosuka. Downloaded on January 3, 2010 at 22:24 from IEEE Xplore. Restrictions apply.

A Re-quantization Noise Reduction Method in MPEG-2 ...

the 4x4 integer DCT of H.264 has also been reported[5]. We have focused our attention on the fact that the com- plete re-use of encoding information, such as picture type, motion vector, and macro-block type, makes it possible to suppress re-quantization noise to almost zero[6]. This noise suppression mechanism is ...

397KB Sizes 0 Downloads 155 Views

Recommend Documents

Wavefront Noise Reduction in a Shack-Hartmann ...
techniques that can increase both speed (e.g., steady-state solution [12]) and numerical precision (e.g., U-D factorisa- tion [13]) of the Kalman filter. Other solutions include sophis- ticated centroiding algorithms (iterative [10], correlation-base

Use of adaptive filtering for noise reduction in ...
software solutions to the adaptive system using the two main leaders of adaptive LMS (least mean square) ... environment is variable in time and its development.

Zwicker Tone Illusion and Noise Reduction in the Auditory System
May 1, 2003 - in noisy surroundings is given as an illustration. ... effect in 1964, now called the Zwicker tone. ..... [16] I. Nelken and E.D. Young, J. Basic Clin.

Noise reduction in multiple-echo data sets using ...
Abstract. A method is described for denoising multiple-echo data sets using singular value decomposition (SVD). .... The fact that it is the result of a meaningful optimization and has .... (General Electric Healthcare, Milwaukee, WI, USA) using.

Three-Dimensional Anisotropic Noise Reduction with Automated ...
Three-Dimensional Anisotropic Noise Reduction with. Automated Parameter Tuning: Application to Electron Cryotomography. J.J. Fernández. 1,2. , S. Li. 1.

COLLABORATIVE NOISE REDUCTION USING COLOR-LINE MODEL ...
pose a noise reduction technique by use of color-line assump- .... N is the number of pixels in P. We then factorize MP by sin- .... IEEE Conference on. IEEE ...

Recurrent Neural Networks for Noise Reduction in Robust ... - CiteSeerX
duce a model which uses a deep recurrent auto encoder neural network to denoise ... Training noise reduction models using stereo (clean and noisy) data has ...

A Survey of Noise Reduction Methods for Distant ...
H.3.1 [Information Storage and Retrieval]: Content. Analysis ... relation extraction paradigms can be distinguished: 1) open information ... While open information extraction does ..... to the textual source on which it is most frequently applied,.

Use of adaptive filtering for noise reduction in communications systems
communication especially in noisy environments. (transport, factories ... telecommunications, biomedicine, etc.). By the word ..... Companies, 2008. 1026 s.

pdf-1491\noise-reduction-techniques-in-electronic-systems-by ...
pdf-1491\noise-reduction-techniques-in-electronic-systems-by-henry-w-ott.pdf. pdf-1491\noise-reduction-techniques-in-electronic-systems-by-henry-w-ott.pdf.

A Network Traffic Reduction Method for Cooperative ...
[11], our work in [9] operates across the physical layer and the medium access layer. In this paper, we extend our previous work by developing a network traffic ...

A Network Traffic Reduction Method for Cooperative ...
Wireless positioning has been providing location-based ser- vices in ... Let us consider a wireless network with two types of ..... Cambridge University Press,.

Joint ICI and Noise Reduction in OFDM Using a New ... - IEEE Xplore
transmitter and the receiver or Doppler spread. Carrier frequency offset causes intercarrier interference (ICI) and ICI degrades the system performance and ...

Hyperspectral image noise reduction based on rank-1 tensor ieee.pdf
Try one of the apps below to open or edit this item. Hyperspectral image noise reduction based on rank-1 tensor ieee.pdf. Hyperspectral image noise reduction ...

Speckle Noise Reduction of Medical Ultrasound ...
tors are named by abbreviation in Table III based on the estimation approach ...... 3, pp. 156-163, 1983. [25] A. N. Evans, M. S. Nixon, “Mode filtering to reduce ultrasound ... and image restoration. Mansur Vafadust received his B.sc. degree.

impulse noise reduction using motion estimation ...
requires a detailed knowledge of the process, device models and extreme care during layout. The main type of capacitors involved are gate, miller and junction capacitors related to input and output stage of the transconductors connected to the integr

Noise and Air Pollution Reduction Measures.pdf
alternative power sources such as battery power. • Maintenance, servicing and testing done during business hours to avoid disrupting sleep and weekend.