WO1995009487A1 - High efficiency encoding/decoding device - Google Patents

High efficiency encoding/decoding device Download PDF

Info

Publication number
WO1995009487A1
WO1995009487A1 PCT/JP1994/001546 JP9401546W WO9509487A1 WO 1995009487 A1 WO1995009487 A1 WO 1995009487A1 JP 9401546 W JP9401546 W JP 9401546W WO 9509487 A1 WO9509487 A1 WO 9509487A1
Authority
WO
WIPO (PCT)
Prior art keywords
rounding
coefficient
inverse
orthogonal transform
quantization
Prior art date
Application number
PCT/JP1994/001546
Other languages
English (en)
French (fr)
Inventor
Yoshihiro Murakami
Atsuo Yada
Hiromi Yoshinari
Haruo Togashi
Satoshi Miyazawa
Takuya Kitamura
Original Assignee
Sony Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sony Corporation filed Critical Sony Corporation
Priority to EP94927072A priority Critical patent/EP0671816B1/en
Priority to US08/428,255 priority patent/US5604494A/en
Priority to DE69423734T priority patent/DE69423734T2/de
Publication of WO1995009487A1 publication Critical patent/WO1995009487A1/ja

Links

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/40Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using video transcoding, i.e. partial or full decoding of a coded input stream followed by re-encoding of the decoded output stream
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/44Decoders specially adapted therefor, e.g. video decoders which are asymmetric with respect to the encoder
    • H04N19/45Decoders specially adapted therefor, e.g. video decoders which are asymmetric with respect to the encoder performing compensation of the inverse transform mismatch, e.g. Inverse Discrete Cosine Transform [IDCT] mismatch
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/60Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding

Definitions

  • the present invention relates to a high-efficiency coding / decoding apparatus, and in particular, to a high-efficiency coding / decoding apparatus which repeats a digital signal such as a video signal to perform high-efficiency coding.
  • the present invention relates to a high-efficiency coding / decoding device capable of suppressing accumulation of errors during decoding.
  • BACKGROUND ART In recent years, as a high-efficiency encoding / decoding device that compresses and encodes a digital signal and then decodes it, for example, a digital video signal is compression-encoded and recorded on a recording medium. A digital VTR that decodes a signal reproduced from this recording medium has appeared.
  • a video signal is generally compressed and recorded as described below, and further reproduced and expanded in the following manner.
  • the time-axis digital video data supplied to the signal recording system is first subjected to orthogonal transform such as discrete cosine transform (DCT ′).
  • DCT ′ discrete cosine transform
  • the video data on this frequency axis is quantized and further compressed, for example, by performing variable-length coding.
  • the compressed video data is recorded on a magnetic tape as a recording medium.
  • the compressed video data recorded on the recording medium is reproduced.
  • This reproduced data is expanded by variable-length decoding, and further subjected to inverse quantization.
  • the inversely quantized data is subjected to an inverse discrete cosine transform (IDCT) as an inverse orthogonal transform, so that it is restored from the frequency axis to a video image on the time axis again. After that, the restored video data will be retrieved.
  • IDCT inverse discrete cosine transform
  • Some of the digital VTRs that perform such compression encoding of a video signal use, for example, a predictive encoding method between frames / fields.
  • a predictive encoding method between frames / fields.
  • the local decoded image for performing the above-mentioned predictive coding on the encoding side and the decoded image on the decoding side match.
  • the calculation method of the inverse transform (inverse orthogonal transform) at the time of local decoding on the encoding side and the inverse transform (inverse orthogonal transform) on the decoding side, and the rounding on the encoding side and the decoding side If the above methods are different, there is a problem that a mismatch as described later may occur.
  • the quantization representative value is odd except for 12048.
  • the odd number is used for the quantization representative value in order to solve the problem that even if the IDCT standard is satisfied, a mismatch between IDCTs of different designs can occur.
  • the quantization representative value is positive / negative symmetric except for +2 047 Z—2048.
  • the step size is 2 ⁇ QUANT.
  • the above-mentioned mismatch is caused by the difference in the calculation method and rounding method in the inverse transform, but the same applies even when the calculation method and the rounding method in the inverter are not different. Problems may occur.
  • the above-mentioned misalignment may cause image degradation in the multi-generation characteristics during direct digital dubbing. is there.
  • pattern emphasis occurs due to monotone increase or decrease of the amplitude level of a specific image pattern.
  • FIG. 1 shows a configuration in a case where the above-mentioned direct digital dubbing is performed in this digital VTR.
  • an input video signal supplied via a terminal 103 is recorded on a magnetic tape in a digital VTR 100.
  • the output terminal of the digital VTR 100 is connected to the input terminal of the digital VTR 101, and the output terminal of the digital VTR 101 and the input terminal of the digital VTR 100 are connected. And are connected. By repeating recording and reproduction in each digital VTR 100, 101, multi-dubbing is performed.
  • the output terminal of the digital VTR 100 is also connected to the monitor 102, so that the monitor 102 can be used to observe changes in image quality due to multi-dubbing. it can .
  • each of the digital VTRs 100 and 101 in FIG. 1 is a digital VTR of a component record that employs bitrate reduction.
  • the bit rate reduction method is a conversion code + variable length coding
  • the DCT is used as a conversion basis function.
  • each digital VTR 1001 supports a system of 10-bit video, and therefore, the calculation accuracy of the transform-inverse transform (DCT-IDCT) sufficiently satisfies this video accuracy.
  • the calculation word length is adjusted as described above.
  • the configuration for performing direct digital dubbing shown in FIG. 1 can be simplified as shown in FIG.
  • data from one digital VTR is supplied to terminal 103 as input data, and this input data is supplied to DCT circuit 111 by DCT circuit 111.
  • DCT is performed.
  • the coefficient data from the DCT circuit 111 is quantized by a requantizer (quantizer, inverse quantizer) 112, and the output is sent to the IDCT circuit 113.
  • a rounding error E re occurs in the output from the requantizer 1 1 2.
  • a rounding error E rs also occurs in the output from the IDCT circuit 113, and the output of the IDCT circuit 113 is sent to the DCT circuit 111. Infinity rounding described later is used.
  • the output of the IDCT circuit 113 is sent from a terminal 104 to a monitor or the like.
  • the multi-dubbing in FIG. 2 can be expressed as shown in FIG. 3 by another expression.
  • FIG. 3 shows, for example, a case where dubbing is performed twice (that is, a case where conversion-inverse conversion is performed three times).
  • the two dubbing operations are performed by connecting three sets of a DCT circuit 111, a quantizer, an inverse quantizer 112, and an IDCT circuit 113, in series. Is equivalent to.
  • coefficients other than the DC coefficient are uniformly quantized.
  • the quantization step of the DC coefficient is defined as qdc
  • the quantization step of the AC coefficient is defined as qac.
  • Quantization level q x x 'Q (coefficient / q x x)
  • the rounding method includes, for example, simple rounding (forward rounding) and infinity rounding. The differences are described below.
  • the boundary point is always rounded up because the decision is made only with the rounding bit. Therefore, as shown in C of FIG. 5, the error distribution is always biased, including ⁇ / 2.
  • the boundary point is just half of the bit to be rounded, that is, ⁇ 0.5, where asymmetry appears in the error distribution.
  • the triangle in the figure indicates that the value is not included, and the symbol in the figure indicates that the value is included. That is, in A and B in FIG. 6, when the value is larger than one ⁇ / 2 and less than ⁇ 2, the value is rounded to 0, and when the value is more than 2 and less than 3 ⁇ / 2, 1 ⁇ Rounded to 2 && 3 ⁇ ⁇ 2 or more and less than 5 mm Z 2 is 2 ⁇ 2 — ⁇ Rounded, and is larger than 1 3 ⁇ ⁇ 2 and less than 2 mm — 1 ⁇ Rounded to 2, 13 m larger than 15 ⁇ / 2 It indicates that when Z2 is less than 2, it is rounded to —2 ⁇ 2— b .
  • P () of C in Fig. 6 also indicates the probability.
  • the method uses infinity rounding and sets the quantization step size to 2.
  • the result looks like D ⁇ 4, 4, 4 3 ⁇ , and the second and subsequent encoding / decoding (that is, applying dabin will change the divergence direction in the divergence direction) In other words, the rounding errors do not accumulate.
  • the present invention is intended to significantly reduce image degradation in, for example, multi-generation characteristics during direct digital dubbing. It is an object of the present invention to provide a high-efficiency encoding / decoding device capable of performing the above-described operations.
  • the high-efficiency encoding / decoding device has been proposed in order to achieve the above-mentioned object, and has a structure in which an analog signal is converted into an analog signal.
  • Orthogonal transform means for orthogonally transforming the evening signal; and inverse orthogonal transform means for inversely transforming the orthogonally transformed digital signal, and at least one of the orthogonal transform and the inverse orthogonal transform is performed.
  • rounding in the even direction or rounding in the odd direction is used.
  • the high-efficiency encoding / decoding device of the present invention further includes a quantizing means for quantizing the digitized signal after the orthogonal transform, and a digital signal before the inverse orthogonal transform. And inverse quantization means for inverse quantization.
  • FIG. 1 is a diagram showing a configuration of a digital VTR at the time of direct digital dubbing.
  • FIG. 2 is a block circuit diagram showing a simplified configuration of direct digital dubbing.
  • FIG. 3 is a block circuit diagram for explaining a flow when dubbing is performed twice.
  • FIG. 4 is a diagram conceptually showing the relationship between the input / output word length and the significant digits of the normalized simulation model in DCT_IDCT.
  • FIG. 5 is a diagram for explaining forward rounding.
  • FIG. 6 is a diagram for explaining rounding toward infinity.
  • FIG. 7 is a block circuit diagram showing a schematic configuration of the high-efficiency encoding / decoding device according to the embodiment of the present invention.
  • FIG. 8 is a diagram for explaining rounding in the even-numbered direction.
  • Fig. 9 is a diagram showing the range of error when the quantization step size is set to three.
  • FIG. 10 is a diagram showing simulation results of multi-dubbing when using DCT and forward rounding.
  • FIG. 11 is a diagram showing a simulation result of multi-dubbing when DCT and rounding in infinity are used.
  • FIG. 12 is a diagram showing a simulation result of multi-dubbing when using DCT and rounding in the even direction.
  • Figure 13 shows the DCT with 12-bit arithmetic precision and the multi-dubbing simulation with rounding in the even direction, rounding toward infinity, and rounding in the positive direction. It is a figure which shows a utilization result.
  • FIG. 14 is a diagram showing simulation results of multi-dubbing using DWT and forward rounding.
  • Figure 15 shows the simulation results of multi-dubbing when DWT and rounding to infinity are used.
  • FIG. 16 is a diagram showing simulation results of multi-dubbing when DWT and rounding in the even direction are used.
  • Figure 17 shows the simulation results of multidubbing when DWT is used when the arithmetic precision is 12 bits, rounding in the even direction, rounding to infinity, and rounding in the positive direction.
  • FIG. 18 is a diagram showing a group (1) of image enhancement patterns.
  • FIG. 19 is a diagram showing a group (2) of image enhancement patterns.
  • FIG. 20 is a diagram showing a group (3) of image enhancement patterns.
  • FIG. 21 is a diagram showing a group (4) of image enhancement patterns.
  • the high-efficiency encoding / decoding device of the present invention further includes the orthogonal transform.
  • a quantizer 11 for quantizing the converted digital signal and an inverse quantizer 21 for inversely quantizing the digital signal before the inverse orthogonal transform are provided.
  • a digital signal obtained by analog-to-analog conversion of an analog signal is supplied to a terminal 1, and this signal is supplied to a quadrature converter 1 for performing, for example, a discrete cosine transform (DCT) process.
  • DCT discrete cosine transform
  • the digital signal that has been orthogonally transformed from the time axis to the frequency axis by the orthogonal transformer 10 is quantized by the quantizer 11 and further variable by the variable-length encoder 12. It is compressed by being subjected to long coding.
  • the compressed digital data is recorded on a recording medium by the recording means 13.
  • the compressed digital data recorded on the recording medium is reproduced by the reproducing means 23.
  • This reproduced data is decompressed by the variable length decoder 22 and subjected to inverse quantization by the inverse quantizer 21.
  • the inversely quantized data is subjected to an inverse orthogonal transform (for example, an inverse discrete cosine transform: IDCT) by an inverse orthogonal transformer 20 so that the time axis is again shifted from the frequency axis to the time axis.
  • IDCT inverse discrete cosine transform
  • the problem of image degradation occurs as described above.
  • the above-mentioned orthogonal transform for example, DCT and inverse orthogonal transform, for example, at least one of the IDCT and the rounding in the even direction or the odd direction is performed. I use it. Note that the rounding in the even-numbered direction is, as shown in FIG. 8, rounded down and rounded up so that the result of rounding the boundary point always becomes an even value.
  • a mark in the figure indicates that the value is not included, and a mark in the figure indicates that the value is included.
  • a and B in Fig. 8 when the value is not less than one ⁇ / 2 and not more than ⁇ / 2, round to 0, and when it is larger than ⁇ 2 and less than 3 ⁇ / 2, 1
  • When rounded to b when it is 3 ⁇ ⁇ 2 or more and 5 ⁇ / 2 or less, it is 2 ⁇ 2-b. Therefore, if it is larger than 13 lm / 2-less than lm / 2, use l'2 -b.
  • P () in C in Fig. 8 indicates the probability.
  • the CCT of the 2 ⁇ 2 pixel data and the corresponding 2 ⁇ 2 IDCT represented by C 10 C 1 are
  • the quantization step size for example, 3 is used as a value excluding the power of 2 above, and two dubbing, that is, input-decoding processing is performed three times for the input .
  • the change over time in this case is as follows.
  • the data does not change no matter how many times the connection is continued after the first time.
  • at least one of the orthogonal transform (discrete cosine transform: DCT) and the inverse orthogonal transform (IDCT) is rounded in the even direction (or rounded in the odd direction). ) Is described in detail below.
  • the rounding in the even-numbered direction or the rounding in the odd-numbered direction will be described without using the rounding in the infinite direction as in the above-described conventional example.
  • the rounding in the even direction is mainly described, but the rounding in the odd direction is basically the same as the rounding in the even direction.
  • DCT discrete cosine transform
  • the DC coefficient 2 is a change due to the processing of the space plane.
  • the AC coefficient in the above (1-3) behaves as if it were a catalyst, and after the second time (conversion-inverse conversion second time, dubbing first time), the same size is obtained. Repeat the click. Thus, this repetition causes the entire image sample to monotonically increase.
  • the input in this specific example 2 is
  • AC4 AC5 AC6 AC7 -113 -113-112 -111 AC8 AC9 AC10 AC11 -113 -113 -112 -112 AC12 AC13 AC14 AC15-112-112 -112 -112
  • the DC coefficient of the DCT output is 1499. Dividing this DC coefficient by 2 (requantization) results in —2 24.5. The AC coefficient is also divided by 2 (requantization).
  • the change in the DC coefficient is a change due to the processing of the coefficient surface.
  • the change in the DC coefficient is -2 in both the coefficient plane and the space plane, and is equivalent to one quantization step.
  • a new component due to the processing of the spatial plane appears in AC 8 of the AC coefficient.
  • the transform and inverse transform have sufficient computational accuracy so that there is no accumulation of errors except at the boundary points in rounding. Therefore, only the boundary points in rounding are involved in the accumulation cycle.
  • the requantization step is recursively performed on the coefficient plane. Bring a shift. If the change in the former process includes a component other than the boundary point of the space plane, the increase of the boundary point due to the rounding of the space plane returns to the coefficient plane boundary point again due to the symmetry of DCT — IDCT. The storage cycle is cut off. Therefore, an accumulation cycle occurs only when the boundary point of the space plane is the coefficient plane and that of the coefficient plane is that of the space plane.
  • the spatial output is shifted unipolarly by the DC coefficient, and as a result, the processing of boundary points by the rounding of the spatial plane is converted to unipolarity. As a result, the average value of the error is shifted. This is a major factor causing a monotonous increase (decrease).
  • the accumulation cycle occurs during rounding, which is an integer operation (non-linear processing), and the coefficients of the coefficient plane and the spatial plane samples are well related.
  • This is a very special case involved, low stochastic cases, and this phenomenon will always occur in any picture, as evidenced by the low stochastic case.
  • the image enhancement pattern confirmed in the actual image is the pattern under the first, second, and third conditions described in the above-mentioned condition of the storage cycle.
  • the cause of the above-mentioned accumulation cycle is that a method of balancing the error distribution by positive and negative polarities (that is, rounding to infinity) is employed.
  • a bit rate reduction system using transform coding there is no close correspondence other than the DC coefficient between the polarity of the coefficient plane and the polarity of the space plane.
  • the effect of the processing of the boundary points is always converted to unipolar in the spatial plane, and as a result, the distribution of the error does not tilt in the direction of the balance, but has a bias in one direction, which accumulates. It is done.
  • the image pattern emphasized by the monotonous increase and decrease becomes a geometric pattern, and greatly impairs the image quality of the multi-generation.
  • a method of balancing the error distribution based on even values for the output rounding of the space plane, that is, rounding in the even direction is adopted.
  • the rounding of the spatial plane output is based on an odd value based on the odd value. It is also possible to adopt a method of balancing the distribution of
  • FIG. 10 shows a video signal when dubbing is repeated on a straight-line when the DCT operation accuracy is 12 bits or 13 bits or 14 bits.
  • the state of degradation of the SZN of the Y signal and the PR and PB signals (simulation results after multiple dubbing) is shown.
  • the evaluation image has a compression ratio of 1/2.
  • DCT and IDCT operations are performed on floating-point numbers, and rounding to integers uses a function based on the IEEE754 standard. Rounding in the forward direction is performed.
  • FIG. 11 shows a simulation result in the case of using the infinite direction rounding
  • FIG. 12 shows a simulation result in the case of using the even direction rounding in this embodiment. Shows the results of the review.
  • DCT is used.
  • -Bit rate reduction by wavelet conversion has better image quality than DCT.
  • the wavelet transform based on the so-called Haar basis can be realized by addition and subtraction of integers, and can be frequency-decomposed after block division as in DCT, so that the hardware scale is small. it can .
  • the DCT part in the above-described embodiment is replaced with an 8 ⁇ 8 discrete wavelet transform (DWT) of 10 by the Haar basis to simulate.
  • DWT discrete wavelet transform
  • the arithmetic precision of DWT is a maximum of 14 bits, and no rounding error occurs in this case.
  • DWT and IDWT also use the same rounding method as quantization, there are three places of rounding. These roundings are compared for rounding in the above directions, rounding to infinity, and rounding in even directions.
  • FIGS. 14 to 17 show the Y, PB, and PR signals of the video signal, and when the DWT calculation accuracy is 12 bits, 13 bits, and 14 bits, This shows the simulation results when dubbing is repeated on the straight.
  • Fig. 14 shows the simulation using positive rounding
  • Fig. 15 shows the simulation using infinity rounding
  • Fig. 16 shows the simulation using even rounding.
  • FIG. 17 shows simulation results of rounding in the even direction, rounding to the infinity direction, and rounding in the positive direction in the case of 12 bits in which the arithmetic precision of DWT is small.
  • the even-direction rounding (or the odd-direction rounding) is used in the transform coding having sufficient calculation accuracy, the singularity in the rounding is reduced.
  • the accumulation of the effects of processing certain boundary points can be prevented.
  • the monotonic increase and decrease in the amplitude of the fixed image pattern can be prevented.
  • the present invention is applicable not only to a digital VTR but also to a system employing bit encoding by transform coding, so that the present invention can be applied to other similar systems such as an audio system. Can also be applied.
  • a pattern in which the image is enhanced in the accumulation cycle is given as an example.
  • image enhancement pattern a pattern in which the image is enhanced in the accumulation cycle.
  • the variation of the pattern is, for example, the group (1) to (1) shown in Fig. 18.
  • the group (4) shown in 21 is conceivable. In the above groups (1) to (4),
  • the input coefficient value format is
  • the accumulation cycle is a special case where the rounding, which is an integer operation (non-linear processing), the coefficients of the coefficient plane, and the spatial plane sample are well related.
  • the patterns of the groups (1), (2), and (3) shown in FIGS. 18 to 21 can be mentioned.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Discrete Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Television Signal Processing For Recording (AREA)
  • Compression Or Coding Systems Of Tv Signals (AREA)
  • Signal Processing For Digital Recording And Reproducing (AREA)

Description

明 細 書 高能率符号化復号化装置 技 術 分 野 本発明は、 高能率符号化復号化装置に関 し、 特に 、 例えばビデオ 信号などのディ ジタ ル信号を繰 り 返 して高能率符号化復号化する際 の誤差の蓄積を抑え得る よ う な高能率符号化復号化装置に関する も のである 。 背 景 技 術 近年は、 ディ ジタ ル信号を圧縮符号化 し、 その後復号化する高能 率符号化復号化装置 と し て 、 例えばディ ジタル ビデオ信号を圧縮符 号化 して記録媒体に記録 し , この記録媒体か ら再生 した信号を復号 化す るよ う なディ ジ 夕ル V T Rが登場 し て きて いる 。
上記ディ ジタ ル V T R にお いては、 一般に、 以下のよ う に して ビ デォ信号を圧縮 し て記録 し 、 さ ら に再生 して伸張するよ う になさ れ て い る。
すなわち 、 図示は省略するが、 信号記録系 (符号化側 ) に供給さ れた時間軸のディ ジ タル ビデオデータは、 先ず、 例えば離散コサイ ン変換 ( D C T ' 等の直交変換が施される こ と に よ り周波数軸のデ —夕 に変換される。 この周波数軸の ビデオデータ は、 量子化され、 更に例えば可変長符号化等が施さ れる こ と によっ て圧縮さ れる。 当 該圧縮さ れた ビデオデー タ は記録媒体と しての磁気テープに記録さ れる 。
ま た、 信号再生系 I 復号化側) では、 上記記録媒体に記録された 上記圧縮さ れた ビデオデータが再生される。 この再生データ は、 可 変長復号化によ っ て伸張さ れ、 さ ら に逆量子化が施される 。 当該逆 量子化さ れたデ一夕 は、 逆直交変換 と し て逆離散コサイ ン変換 ( I D C T ) が施さ れる こ と に よ って上記周波数軸か ら再び時間軸の ビ デォデ一夕 に復元さ れ、 その後この復元さ れた ビデオデ一夕が取 り 出 さ れる こ と になる 。
なお、 このよ う な ビデオ信号の圧縮符号化を行う ディ ジタル V T R に は、 例えばフ レーム間/フ ィ 一ル ド間の予測符号化方式を用 い る も のがある。 この よ う なディ ジタ ル V T R にお いては、 符号化側 で上記予測符号化を行う ための局部復号画像と復号化側の復号画像 が一致する必要があ る。 この と き符号化側での局部復号の際の逆変 換 (逆直交変換) と復号化側における逆変換 (逆直交変換) での演 算方法及び符号化側 と復号化側にお ける丸めの方法が異な る と 、 後 述す るよ う な ミ スマ ヅチが起こ り 得る問題がある。 このため、 例え ば国際電信電話諮問委員会 ( CCITT : Comite Consultatif Interna tionale Te 1 eg r aph i que et Telephonique)における勧告 H . 2 6 1 ( D C T を用 いたテ レ ビ会議/電話用低速動画像符号化ァルゴ リ ズ ム ) では、 表 1 に示すよ う に、 量子化代表値を奇数にする よ う に し て い る。
以下余白 デッ卜ゾーンつき量子化の量子化代表値
QUANT 1 0 4 • R q • 17 1
1 o Q 1
J
量子
Γノ 7 ックス
一 127 -255 一 509 一 765 一 1019 . -2039 -2048 . -2048 - 0Δ *tΛo 一 126 一 505 一 759 一 1011 —2048 - .ΓKJιΑ'-tΛO 一 -Q Ό 一 1 一 19 HO -QQ 1 Ό 一 1 C;
-1 一 3 - 5 - 9 -11 • -23 -27 • -51 -53 • -89 - 93
0 0 0 0 0 • 0 0 • 0 0 • 0 0
1 9 11 • 23 27 • 51 • RQ QQ
2 5 9 15 19 • 39 45 • 85 «9 • 149 155
3 7 13 21 27 • 55 63 • 119 125 • 209 217
4 9 17 27 35 • 71 81 • 153 161 • 269 279
5 11 21 33 43 • 87 99 • 〗87 197 • 329 341
56 113 225 339 451 • 903 1017 • 1921 2033 • 2047 2047
57 115 229 345 459 • 919 1035 • 1955 2047 • 2047 2047
58 117 233 351 467 • 935 1053 - 1989 2047 • 2047 2047
59 119 237 357 475 • 951 1071 • 2023 2047 • 2047 2047
60 121 241 363 483 • 967 1089 • 2047 2047 • 2047 2047
125 251 501 753 1003 • 2007 2047 • 2047 2047 - 2047 2047
126 253 505 759 1011 • 2023 2047 • 2047 2047 • 2047 2047
127 255 509 765 1019 • 2039 2047 • 2047 2047 • 2047 2047
すなわ ち この表 1 に示すよ う に、 上記勧告 H . 2 6 1 では、 量子 化代表値が一 2 0 4 8 を除いて奇数 となっ て いる。 この量子化代表 値に奇数を用 いるの は、 I D C T規格を満た して も異な る設計の I D C T間の ミ ス マ ヅ チが起こ り 得る と い う 問題を解決す るためであ る。 なお 、 こ の表 1 にお いて 、 量子化代表値は + 2 0 4 7 Z— 2 0 4 8 を除き正負対称であ る 。 ま た、 ステ ップサイ ズ = 2 x Q U A N Tである 。
と ころ で、 上記 ミ スマ ッ チは、 逆変換における演算方法や丸めの 方法が異なる こ とが原因であ るが、 逆変換器における演算方法や丸 めの方法が異な らな い場合でも同様の問題が発生するこ とがある 。 例えば、 ディ ジタ ル V T Rにお いて、 上記 ミ ス マ ヅチ と してダイ レ ク 卜ディ ジ夕ルダビン グ時のマルチジェネ レー シ ョ ン特性におけ る画像劣化の問題が生 じ る こ とがあ る。
すなわち 、 ある特定画像パター ンの振幅レベルの単調増加又は減 少に よる パター ン強調が発生する こ とである。
以下、 上記ディ ジタル V T Rにお けるダイ レ ク 卜 ディ ジタ ルダ ビ ング時のマルチジェネ レ ー シ ョ ン特性にお いて、 画像劣化が生ずる こ と につ いて 、 図面を用 いて説明する。
こ こで、 図 1 には、 このディ ジタ ル V T Rにお いて上記ダイ レ ク 卜 ディ ジタルダビン グを行う場合の構成を示す。
こ の図 1 にお いて 、 端子 1 0 3 を介 して供給さ れた入力 ビデオ信 号は、 ディ ジタル V T R 1 0 0 にお いて磁気テープに記録さ れて い る。 このディ ジタル V T R 1 0 0の出力端子とディ ジ夕 ル V T R 1 0 1 の入力端子とが接続され、 さ ら にディ ジタル V T R 1 0 1の出 力端子と ディ ジタル V T R 1 0 0の入力端子とが接続さ れて いる 。 各ディ ジタル V T R 1 0 0 , 1 0 1 にお いて記録ノ再生を繰 り 返す こ と によ り 、 マルチダビ ングが行われる こ と にな る 。 なお、 デイ ジ タル V T R 1 0 0の出力端子はモニタ 1 0 2 と も接続さ れ、 したが つて 、 このモニタ 1 0 2 に よ ってマルチダ ビング に よ る画質の変化 を観察する こ とがで きる 。
ま た、 上記図 1 の各ディ ジタル V T R 1 0 0 , 1 0 1 は、 ビッ ト レ一 卜 リ ダク シ ョ ン を採用 したコ ン ポーネ ン 卜記録のディ ジタル V T Rであ る。 こ こで、 上記ビッ ト レ ー ト リ ダク シ ョ ンの方式は、 変 換符号 +可変長符号化と し 、 変換基底関数には上記 D C Tを採用 し て い る と する。 更に 、 当該各ディ ジタル V T R l O O i 1 0 1 では 1 0 ビッ ト映像のシ ステ ム をサポー 卜 し 、 したがっ て変換一逆変換 ( D C T — I D C T ) の演算精度も この映像精度を十分満足するよ う に演算語長を と つ て調整 して いる とす る。
この図 1 に示すダイ レ ク 卜 ディ ジタルダビングを行う構成は、 簡 略化する と図 2のよ う に示すこ とができ る 。
す__ わち この図 2 にお いて 、 端子 1 0 3 には一方のディ ジタル V T Rからのデータが入力データ と し て供給さ れ、 この入力デ一夕が D C T回路 1 1 1 に よっ て D C T さ れる 。 この D C T回路 1 1 1 か らの係数デ一夕 は、 再量子化器 (量子化 , 逆量子化器) 1 1 2 によ つ て量子化され、 その出力が I D C T回路 1 1 3 に送ら れる。 この 再量子化器 1 1 2か らの出力 に丸め誤差 E r eが発生する。 ま た、 上記 I D C T回路 1 1 3 か らの出力 にも丸め誤差 E r sが発生 し、 この I D C T回路 1 1 3 の出力は上記 D C T回路 1 1 1 に送 られる, なお 、 こ こでの丸め には後述する無限大方向丸めを採用 して いる。 I D C T回路 1 1 3の出力は端子 1 0 4か ら モニタ等に送 られる。 ま た、 この図 2 における マルチダ ビングを更に別の表現に よ り 表 すと 、 図 3のよ う に示す こ とができ る。 この図 3 においては、 例え ば 2 回のダビングを行う 場合 ( すなわち変換一逆変換を 3回行う場 合) を示 して いる。 この 2回のダビングは、 D C T回路 1 1 1 と量 子化 ' 逆量子化器 1 1 2 と I D C T回路 1 1 3 とか らな る 1組の構 成を 、 直列に 3組接続 し た こ と に相当する 。
なお、 上記図 2の構成 にお いて 、 入出力の語長と有効桁の関係を 概念的に表すと 、 図 4の よ う に示す こ とができ る。
この図 4 にお いて 、 D C係数以外の係数 ( A C係数) は一様量子 化さ れて いる。 こ こで、 D C係数の量子化ステップを q d c と し、 A C係数の量子化ス テッ プを q a c とする。 ま た、 正規化 した D C T 、 I D C Tを使う 。 この と き 、 係数面の ビッ ト と再量子化ステ ツ ブの関係は、
q X X =除数
量子化レベル = q x x ' Q 〔係数/ q x x〕
とな る。 なお Q 〔 〕 は丸めを示す。
したがって 、 例えば
q d c = q a c = 1 係数面 1 2 ビ ッ ト に丸め q d c = q a c = 2 係数面 1 1 ビッ 卜 に丸め q d c = q a c = 4 係数面 1 0 ビ ッ ト に丸め であ る。
上述 した よ う なデ ィ ジ タ ル V T R にお いて 、 D C T— I D C Tの 演算では十分な精度を持ってお り 問題はな い とする と、 上記ダイ レ ク 卜 ディ ジタルダビング時の画像劣化 (特定画像パター ンの単調増 加又は減少) の発生は、 丸め誤差が蓄積する こ と に よる と考え られ る。
こ こで、 丸め方式 には 、 例えば単純丸め (正方向丸め ) や、 無限 大方向丸めな どがあ る。 以下これ ら につ いて違い を示す。
先ず、 図 5 を用 いて正方向丸め ( 単純丸め) につ いて説明する 。 この図 5 にお いて 、 図中の〇印はその値を含ま な いこ と を 、 攀印 はその値を含むこ と を示 して いる 。 すなわち 、 図 5 の A及び B にお いて 、 以上 ΔΖ 2未満の と きに は 0 に丸め、 Δ Ζ 2以上 3 Δ / 2未満の と き には 1 · 2— bに丸め、 3 Δ / 2以上 5 Δ / 2未満 の と き には 2 · 2 -bに丸め、 一 3 Δ Ζ 2以上一 Δ Ζ 2未満の と き に は一 1 · 2—bに丸め、 一 5 Δ Ζ 2以上一 3 Δ Ζ 2未満の と き には一 2 · 2—bに丸める こ と を示 して いる 。 ま た、 図 5の Cの P ( ) は 確率を示 して いる。
この正方向丸め (単純丸め ) では、 丸めを行う ビッ ト のみで判断 を行 うため、 境界点は常に切 り 上げる。 したがって 、 図 5の Cに示 すよ う に 、 Δ/ 2 を常に含み、 誤差の分布が偏る こ と になる。 なお 境界点と は、 丸めの対象 と な る ビッ 卜の丁度半分すなわち ± 0 . 5 の こ とで、 誤差の分布に非対称性が現れる と ころである 。
こ のよ う なこ とか ら 、 従来よ り 無限大方向丸めを用いて い る。 図 6 を用 いて無限大方向丸めにつ いて説明する 。
この図 6 にお いて も、 図中の〇印はその値を含ま ない こ と を、 眷 印はその値を含む こ と を示 して いる 。 すなわち、 図 6の A及び B に おいて、 一 Δ / 2 よ り大 き く Δ Ζ 2未満の と き には 0 に丸め、 厶ノ 2以上 3 Δ / 2未満の と き には 1 · 2— &に丸め、 3 Δ Ζ 2以上 5 厶 Z 2未満の と き には 2 · 2 — ^こ丸め、 一 3 Δ Ζ 2 よ り大き く ー 厶 2以下の と き には — 1 · 2 に丸め、 一 5 Δ / 2 よ り大き く 一 3 厶 Z 2以下の と き には — 2 · 2 — bに丸める こ と を示 してい る 。 ま た、 図 6 の Cの P ( ) も確率を示 して いる 。
この無限大方向丸めは、 境界点を絶対値で正負同 じになる よ う に 切 り 上げる。 したがって 、 誤差の分布は、 図 6の Cの ( a ) 〜 ( c に示すよ う に 3種類 とな り 、 当該誤差の分布は X = 0 を中心と して バラ ンス するもの と なる 。
以上、 丸めにお いては、 Δ 2の点のみが誤差の範囲のバラ ンス を崩す点である こ とが判 る。
よ り具体的に説明する 。
こ こで、
時間軸上の 2 X 2の画素のデータ を、 D 0 0 D 0
D 1 0 D 1 周波数軸上の 2 X 2の画素のデータ を C 0 0 C 0
C 1 0 C 1 で表すと する。
ま た、 この 2 x 2の画素データの D C T と これに対応する 2 x 2 の I D C Tを丸めの演算個所を少な く するために、 それぞれ次のよ う にする 。
D C T
C 0 0 = D 0 0 + D 0 1 + D 1 0 + D 1 1
C 0 1 = D 0 0 - D 0 1 D 1 0 - D l 1
C 1 0 = D 0 0 + D 0 1 - D 1 0 - D l 1
C 1 1 = D 0 0 - D 0 1 - D 1 0 D 1 1
I D C T
D 0 0 = ( C 0 0 + C 0 1 C 1 0 C l l ) / 4 D 0 1 = : C O O - C 0 1 + C 1 0 - C I 4 D 1 0 = ( C 0 0 + C 0 1 - C 1 0 - C 1 4 D 1 1 = ( C O O — C O l — C I O 十 C I 4 こ こで、 2 回のタ ビン グ すなわち入力 に対 して 3 回の符号化 -復 号化の処理を施 した場合のデータの変化は次のよ う にな る。 この と きの丸め方法には無限大方向丸めを用い 、 量子化のステ ップサイ ズ を 2 とする。
入力に例えば { 3 , 1 , 1 , 0 } を与えた場合については以下の よ う にな る。
入 力 D { 3 1 , 1 0 }
1 回 目 の符县ィ匕 ー復 ィ匕
D C T後 C { 5 3 , 3 1 } 量子化逆量子化後 C { 6 4 , 4 2 }
I D C T後丸め前 D { 4. 0, 1.0, 1. 0, 0. 0}
I D C T後丸め後 D { 4 1 , 1 0 }
2 回 目の符号化 -復号化 ( 1 回 百のダビング)
D C T後 C { 6 4 , 4 2 } 量子化逆量子化後 C { 6 4 , 4 2 }
I D C T後丸め前 D { 4 - 0, 1.0, 1. 0, 0. 0}
I D C T後丸め後 D { 4 1 , 1 0 }
3 回 目の符号化 -復号化 ( 2 回 巨のダビング)
D C T後 C { 6 4 , 4 2 } 量子化逆量子化後 C { 6 4 , 4 2 }
I D C T後丸め前 D f 4. 0, 1.0, 1. 0, 0. 0}
I D C T後丸め後 D { 4 1 , 1 0 } の よ う にな り 、 2回 目以降は何回符号化一復号化 ( すなわちダビン グ) を施 して もデータ に変化はな い 。
次に入力 に例えば { , 0 } を与える場合に は以下のよ う な る。
入 力 D { 1 , 1 , 1 , 0 }
1 回 目の符号化 -復号化
D C T後 C = { 3 1 , 1 ― 1 } 量子化逆量子化後 C = { 4 2 , 2 ― 2 }
I D C T後丸め前 D 二- { 1 - 5, 1.5, 1. 5, -0 .5}
I D C T後丸め後 D = { 2 2 , 2 ― 1 }
2回 目の符号化 -復号化 ( 1 回 Sのダ ビング)
D C T後 C = { 5 , 3 , 3 一 3 } 量子化逆量子化後 C = { 6 4 , 4 ― 4 }
I D C T後丸め前 D = { L . 5, 2.5, 2. 5 , - 1 .5}
I D C T後丸め後 D { 3 , 3 , 3 2 }
3回 目の符号化 -復号化 ( 2回 巨のダビング)
D C T後 C { 7 , 5 , 5 5 } 量子化逆量子化後 C { 8 , 6 , 6 6 }
I D C T後丸め前 D { 3. 5, 3.5, 3. 5, - 2 .5}
I D C T後丸め後 D { 4 , 4 , 4 3 } のよ う にな り 、 2回 目以降の符号化 復 化 ( すなわちダビン を施すこ とで発散方向にデ一夕が変化 し て し ま う よ う にな Ό。 い 換えれば、 丸め誤差が蓄積 して い く と にな o
このよ う に、 量子化と逆量子化における演算丸めの方法 と して無 限大方向丸めを用 い る と , あ る入力 に対 し て はダ ビ ングの度にデ一 夕が変化 し発散する 、 すなわち丸め誤差が蓄積 して い く よ う にな る 場合があ る。
そ こで、 本発明は、 上述 したよ う なこ と に鑑み、 例えばダイ レ ク ト ディ ジタルダ ビン グ時のマルチジ ェネ レー シ ョ ン特性にお いて画 像劣化を非常に少な く する こ とができる高能率符号化復号化装置を 提供する こ と を 目的 と し て いる。
発 明 の 開 示 本発明に係る高能率符号化復号化装置は、 上述の 目的を達成する ため に提案された ものであ り 、 アナ ログ信号がアナ ログノディ ジ夕 ル変換さ れたたディ ジ夕 ル信号を直交変換する直交変換手段と 、 上 記直交変換されたデ ィ ジ タ ル信号を逆直交変換する逆直交変換手段 と を有し 、 上記直交変換 と逆直交変換の少な く と も一方に偶数方向 の丸め又は奇数方向の丸めを用 いる よう に したものであ る。
こ こで、 本発明の高能率符号化復号化装置には、 さ ら に上記直交 変換後のディ ジ夕ル信号を量子化する量子化手段と 、 上記逆直交変 換前のディ ジタル信号を逆量子化する逆量子化手段と を設けるよ う にする。
本発明 によれば、 直交変換と逆直交変換の少な く とも一方で偶数 方向丸め又は奇数方向丸めを用 いる こ とで、 空間面で極性に関係な く 丸め誤差の範囲をバラ ン ス させ、 例えばダイ レ ク 卜ディ ジタルダ ビン グ時のマルチジェネ レ ー シ ョ ン特性にお いて丸め誤差の蓄積を 防ぎ、 画像劣化を非常に少な く する こ と を可能と してい る。 図面の簡単な説明 図 1 は、 ディ ジタ ル V T Rのダイ レ ク ト ディ ジタルダ ビング時の 構成を示す図であ る 。
図 2は、 ダイ レ ク 卜デ ィ ジタルダ ビン グの構成を簡略化 して示す ブロ ヅ ク 回路図であ る。
図 3は、 2回ダ ビングを行う場合の流れを説明するためのブロ ッ ク 回路図である。
図 4は、 D C T _ I D C T にお け る正規化下シ ミ ュ レ ー シ ョ ンモ デルの入出力語長と有効桁の関係を概念的に示す図であ る。
図 5は、 正方向丸めにつ いて説明する ための図である 。
図 6は、 無限大方向丸めについて説明するための図である。
図 7は、 本発明実施例の高能率符号化復号化装置の概略構成を示 すブロ ッ ク 回路図である 。
図 8は、 偶数方向丸め につ いて説明するための図であ る。
9は、 量子化ス テッ プサイ ズを 3 と した と きの誤差の範囲を示 す図であ る。
図 1 0 は、 D C T と正方向丸めを用いた と きのマルチダビングの シ ミ ュ レ ー シ ョ ン結果を示す図であ る。
図 1 1 は、 D C T と無限大方向丸めを用 いた と きのマルチダビン グの シ ミ ュ レ一 シ ョ ン結果を示す図であ る 。
図 1 2 は、 D C T と偶数方向丸めを用 いた と きのマルチダビング の シ ミ ュ レー シ ョ ン結果を示す図である 。
図 1 3 は、 演算精度 1 2 ビッ ト時の D C T と偶数方向丸め, 無限 大方向丸め, 及び正方向丸めを用 いたと きのマルチダビ ングのシ ミ ユ レ ーシ ヨ ン結果を示す図である 。
図 1 4 は、 D W T と正方向丸めを用 いた と きのマルチダビングの シ ミ ュ レ ー シ ョ ン結果を示す図であ る。
図 1 5 は、 D W T と無限大方向丸めを用 いた と きのマルチダビン グの シ ミ ュ レー シ ョ ン結果を示す図であ る。
図 1 6 は、 D W T と偶数方向丸めを用 いた と きのマルチダビング の シ ミ ュ レーシ ョ ン結果を示す図である 。
図 1 7 は、 演算精度 1 2 ビッ ト 時の D W T と偶数方向丸め, 無限 大方向丸め, 及び正方向丸めを用 いた と きのマルチダビングのシ ミ ュ レーシ ョ ン結果を示す図である。
図 1 8 は、 画像強調パタ ー ンのグループ ( 1 ) を示す図である。 図 1 9 は、 画像強調パター ンのグループ ( 2 ) を示す図である。 図 2 0 は、 画像強調パター ンのグループ ( 3 ) を示す図である。 図 2 1 は、 画像強調パター ンのグループ ( 4 ) を示す図である。 発明を実施するための最良の形態 以下、 本発明の実施例につ いて図面を参照 しなが ら説明する。 本発明実施例の高能率符号化復号化装置は、 図 7 に示すよ う に、 アナ ログ信号がアナ ログ ディ ジタル変換さ れたディ ジタル信号を 直交変換する直交変換器 1 0 と 、 上記直交変換さ れたディ ジタル信 号を逆直交変換する逆直交変換器 2 0 と を有 し、 上記直交変換と逆 直交変換の少な く と も一方に偶数方向の丸め又は奇数方向の丸めを 用 い るよ う に した ものであ る。
ま た、 本発明の高能率符号化復号化装置には、 さ ら に上記直交変 換後のディ ジタル信号を量子化する量子化器 1 1 と 、 上記逆直交変 換前のディ ジタル信号を逆量子化する逆量子化器 2 1 と を設けて い る。
この図 7 にお いて 、 端子 1 にはアナロ グ信号がアナロ グ ディ ジ タル変換されたディ ジタ ル信号が供給さ れ、 これが例えば離散コサ イ ン変換 ( D C T ) 処理等を行う 直交変換器 1 0 に送ら れる。 当該 直交変換器 1 0 で時間軸か ら周波数軸に直交変換さ れたディ ジタル 信号は、 量子化器 1 1 で量子化さ れ、 さ ら に可変長符号化器 1 2 に よ っ て可変長符号化が施される こ と によ っ て圧縮さ れる 。 当該圧縮 されたディ ジ夕ルデ一夕 は記録手段 1 3 によ って記録媒体に記録さ れる 。
ま た、 上記記録媒体に記録された上記圧縮されたディ ジタルデー 夕 は再生手段 2 3 に よっ て再生さ れる。 この再生データ は、 可変長 復号化器 2 2 によ っ て伸張され、 さ ら に逆量子化器 2 1 によ って逆 量子化が施される。 当該逆量子化さ れたデータは、 逆直交変換器 2 0 よつ て逆直交変換 ( 例えば逆離散コサイ ン変換 : I D C T ) が 施さ れる こ と によ っ て上記周波数軸から再び時間軸のディ ジタル信 号に復元さ れ、 その後こ の復元さ れたディ ジタル信号が端子 2か ら 取 り 出さ れる こ と になる 。
こ のよ う なディ ジタル V T R にお ける例えばダイ レク ト デイ ジ夕 ルダ ビング時のマルチジ ェネ レー シ ョ ン特性にお いては、 前述 した よ う に画像劣化の問題が生 じ る こ とがあ るが、 本発明ではこの画像 劣化を少な く するために 、 上記直交変換、 例えば D C T と逆直交変 換、 例えば I D C T との少な く と も一方 に偶数方向の丸め又は奇数 方向の丸めを用 いて いる 。 なお、 上記偶数方向丸め とは、 図 8 に示すよ う に 、 境界点を丸め た結果が必ず偶数値にな る よ う に 、 切 り 捨て 、 切 り 上げを行う もの であ る。
すなわち この図 8 にお いて 、 図中の〇印はその値を含ま な いこ と を、 ·印はその値を含む こ と を示 してい る 。 この図 8の A及び B に お いて、 一 Δ/ 2以上 Δ / 2以下の と き には 0 に丸め、 Δ Ζ 2 よ り 大き く 3 Δ / 2未満の と き には 1 · 2 - bに丸め、 3 ΔΖ 2以上 5 Δ / 2以下の と き には 2 · 2 - bに 。め、 一 3 厶 / 2 よ り大き く — 厶 / 2未満の と き には— l ' 2 -bに 。め、 — 5 Δ Ζ 2以上— 3 Δ Ζ 2以 下の と き には一 2 · 2 -bに 。める こ とを示 して い る 。 ま た、 図 8の Cの P ( ) は確率を示 して いる。
こ の偶数方向丸めでは 、 境界点を丸めた結果が必ず偶数値になる よ う に、 切 り捨て 、 切 り 上げを行う 。 したがって 、 誤差の分布は図 8の Cに示すよ う に 2種類で、 誤差の分布は 「各数値を 中心にバラ ンス 」 する。
よ り具体的に説明する 。
こ こで、 前述の従来例同様に、
時間軸上の 2 X 2 の画素のデータ を、 D 0 0 D 0
D 1 0 D 1 周波数軸上の 2 2の画素のデ一夕 を C 0 0 C 0
C 1 0 C 1 で表 し、 ま た、 この 2 x 2の画素データ の D C T と これ に対応する 2 x 2の I D C T を 、
D C T
C O O = D O O D 0 1 + D 1 0 D l 1 C 0 1 = D 0 0 — D 0 1 十 D 1 0 — D
C I O D O O + D 0 1 - D 1 0 - D
C 1 1 D O O — D O l — D I O + D
D C T
D 0 0 ( C O O 十 C O l 十 C I O 十 C ) / 4
D 0 1 ( C O O - C O l 十 C I O - C ) / 4 D 1 0 ( C O O + C O l - C I O - c ) / 4 D 1 1 ( C O O - C O 1 - C l O + C ) / 4 と する。
こ こで、 量子化ステッ プサイ ズに は上記 2のべき乗を除 く 値と し て例えば 3 を用 い、 2回のダビングすなわち入力に対して 3回の符 号化ー復号化の処理を施す。 この場合のデ一夕の変化は次のよ う に なる 。
入力に前述 した従来例の無限大方向丸めでは丸め誤差が蓄積 した 例えば { 1 , 1 , 1 , 0 } を与えた場合につ いて説明する。
入 力 D = { 1 , 1 , 1 , 0 }
1回 目の符号化 -復号化
D C T後 C = { 3 , 1 , 1 ,- 1 } 量子化逆量子化後 C = { 3 , 0 , 0 , 0 }
I D C T後丸め前 D = { 0.75 , 0.75, 0.75, 0.75 }
I D C T後丸め後 D = { 1 , 1 , 1 , 1 }
2回 目の符号化一復号化 ( 1 回 目のダビング )
D C T後 C = { 4 , 0 , 0 , 0 } 量子化逆量子化後 C = { 3 , 0 , 0 , 0 }
I D C T後丸め前 D = { 0.75, 0.75, 0.75, 0.75 } I D C T後丸め後 D -: { 1 1 , 1 1 }
3 回 目の符号化 -復号化 ( 2 回 巨のダビング )
D C T後 C = { 4 0 , 0 0 } 量子化逆量子化後 C - -: { 3 , 0 0 , 0 }
I D C T後丸め前 D = { 0.75, 0.75, 0.75, 0.75}
I D C T後丸め後 D = { 1 , 1 , 1 , 1 } のよ う に 、 2 回 目以降は何回継 i接 z し て もテ一夕 は変化 しない ま た、 丸めに偶数方向丸めを用 いる例 と して 、 例えば I D C τ後 の丸めに偶数方向丸めを用 いる例に いて説明する
入 力 D { 1 , 1 , 1 , 0 }
1 回 目の符号化 -復号化
D C T後 C = { 3 1 • 1 - 1
} 量子化逆量子化後 C = { 4 2 2 - 2
}
I D C T後丸め前 D = { 1. 5, 1. 5, 1. 5,一 0. 5}
I D C T後丸め後 D { 2 2 2 0 }
2回 目の符号化 -復号化 ( 1 回 巨のダ ビング )
D C T後 C { 6 2 2 一 2 量子化逆量子化後 C = { 6 2 2 一 2
I D C T後丸め前 D = { 2. 0, 2. 0, 2. 0, 0. 0}
I D C T後丸め後 D { 2 2 2 0
3 回 目の符号化 -復号化 ( 2 回 巨のダビング )
D C T後 C 6 2 2 - 2 量子化逆量子化後 C 6 2 2 一 2
I D C T後丸め前 D 2. 0, 2. 0, 2. 0, 0. 0}
I D C T後丸め後 D 2 2 2 0 のよ う に 2回目以降は何回継続接続 して もデータ は変化 しな い。 と ころ で、 本発明実施例において 、 上述 したよ う に上記直交変換 (離散コサイ ン変換 : D C T ) と逆直交変換 ( I D C T ) の少な く と も一方に偶数方向の丸め (又は奇数方向の丸め) を用 い る よ う に した こ と について以下に詳細に説明する 。
先ず、 丸め方法と して 、 前述 した従来例の よ う に無限大方向丸め を用 いずに、 上記偶数方向の丸め又は奇数方向の丸めを用 いる こ と と したこ と につ いて説明する。 なお 、 こ こでは主に偶数方向丸めに ついて述べるが、 奇数方向丸めも こ の偶数方向丸め と基本的には同 じである 。
以下、 ダイ レ ク 卜 ディ ジタルダ ビ ングを繰り返す こ と に よって丸 め誤差が蓄積 (無限大方向丸めによ る誤差の蓄積) して い く 蓄積サ ィ クルのメ カニズム につ いて 、 実画像を用 いて説明する 。
無限大方向丸めに よっ て丸め誤差が蓄積する具体例 1 について述 ベる 。 この具体例 1 では、 D C係数の量子化ステ ップ q b c = 4, A C係数の量子化ス テ ッ プ q a c = 2 と する。 これは、 全画像サン ブルが単調増加する例である。
こ こで、 入力は、
310 303 303 310
310 303 303 310
310 303 303 310
310 303 303 310
とする。 この入力を離散コサイ ン変換 ( D C T ) した D C T出力は、 DC AC1 AC2 AC3 1226 0 14 0
AC4 AC5 AC6 AC7 0 0 0 0
AC8 AC9 AC10 AC11 0 0 0 0
AC12 AC13 AC14 AC15 0 0 0 0 とな り 、 この D C T出力の D C係数は 1 2 2 6 と な る。 こ の D C係 数を 4で割る (再量子化 ) と 、 3 0 6 . 5 となる 。 また 、 A C 1 〜 A C 1 5 の A C係数につ いて は 2で割る (再量子化) 。
こ れに よ り 、
306.5 0 7 0
0 0 0 0
0 0 0 0
0 0 0 0
とな る。
( 1 - 2 )
こ れを丸める と D C係数について は 3 0 7 とな る 。 すなわち、
307 0 7 0
0 0 0 0
0 0 0 0
0 0 0 0
とな る。
こ れを対 して逆量子化を行う 。 D C係数に対 して は 4 を掛ける こ とで 1 2 2 8 とな る 。 したがって 、 上記量子化一逆量子化によって 上記入力 に対する D C係数の変化分は 2 、 すなわち丸め誤差 E r c ( = 0. 5 ) x 4 、 とな る 。 すなわち こ の D C係数の変化分は、 係 数面の処理によ る変化分 な る。 ま た、 A C係数について は 2 を掛 ける こ と で 1 4 と な る。 すなわち
1228 0 14 0
0 0 0 0
0 0 0 0
0 0 0 0 とな る。
( 1 - 3 )
こ こで、 A C係数分の
0 0 14 0
0 0 0 0
0 0 0 0
0 0 0 0 を、 逆変換 ( I D C T ) する と
3.5 -3.5 - 3.5 3.5
3.5 -3.5 -3.5 3.5
3.5 -3.5 -3.5 3.5
3.5 -3.5 -3.5 3.5 とな る。
( 1 - 4 )
ま た、 D C係数分の
1228 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0 を 、 逆変換 ( I D C T ) する と 、 307..0 307. , 0 307..0 307..0
307 , .0 307. , 0 307. .0 307. .0
307. .0 307. , 0 307. .0 307. , 0
307. .0 307. 0 307. .0 307. , 0
とな る。
( 1一 5 )
上記(1-3) すなわち A C係数分の I D C T出力 と 、 上記(1-4) す なわち D C係数分の I D C T出力を加える と 、 上記(1-3) すなわち A C係数分の I D C T出力の境界点は、 全て正極性の処理にな り 、 D C係数の変化に転化さ れる 。
この丸めによ る変化は、
0.5 0.5 0. 5 0.5
0.5 0.5 0. 0 0.5
0.5 0.5 0. 5 0.5
0.5 0.5 0. 5 0.5
とな る。
こ れを D C T する と
2 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
とな る。 すなわち 、 この D C係数の 2が空間面の処理に よ る変化分 とな る。
( 1 - 6 )
上記(1-2) すなわち D C係数変換分の 2 と 、 上記(1-5) すなわち D C係数の変化に転化さ れた後の D C T によ る D C係数変化分の 2 とか ら、 (係数面の処理によ る変化 ) 十 (空間面の処理によ る変化 = 4 とな る。 これは、 D C係数の 1 量子化ステ ップ分である 。 した がっ て、 D C係数は、
1 2 2 6 1 2 3 0
とな る。
( 1 - 7 )
なお、 上記( 1- 3 ) の A C係数の境界点の変化分を D C係数に転化 する こ と によっ て 、 この A C係数は、 全 く 同 じ値 ( = 1 4 ) で係数 面に戻っ て く る。
したがって、
1230 0 14 0
0 0 0 0
0 0 0 0
0 0 0 0
とな る。
( 1 一 8 )
このよ う に、 上記(1-3) の A C係数があたかも触媒の よ う な振る 舞い を し 、 2回目 ( 変換—逆変換の 2回 目 、 ダビングの 1 回目 ) 以 降で全く 同 じサイ ク ルを繰 り 返す。 したがっ て、 この繰 り 返 し によ つ て全画像サンブルが単調増加する よ う になる。
次に、 丸め誤差が蓄積する具体例 2 を挙げて説明する 。
この具体例 2では、 D C係数の量子化ステ ヅブ q b c = 2 , A C 係数の量子化ステ ッ ブ q a c = 2 と する 。
こ れは、 画像ブロ メ ク の 1行と 4行で計 8サンブルが単調減少す る例であ り 、 具体例 1 に比べて複雑であ る 。
こ の具体例 2での入力は、
DC AC1 AC2 AC3 -112 - 113 - 112 - 112
AC4 AC5 AC6 AC7 = -113 -113 - 112 -111 AC8 AC9 AC10 AC11 -113 -113 -112 -112 AC12 AC13 AC14 AC15 - 112 - 112 -112 - 112 と する。
2 - 1 )
この入力 を D C T した D C T出力 は、
-499.00 -1.3750 0.5000 0.5625
-0.1875 -0.3750 0.4375 0.3750
0.5000 1.1250 0.0000 0.0625
-0.4375 0.3750 -0.1875 0.3750
とな り 、 この D C T 出力の D C係数は一 4 9 9 と な る。 この D C係 数を 2で割る (再量子化 ) と 、 — 2 2 4 . 5 となる 。 ま た 、 A C係 数について も 2 で割 る (再量子化) 。
こ れに よ り 、
-224.500 0.68750 0.25000 0.28125
-0.09375 0.18750 0.28175 0.18750
0.25000 0.56250 0.00000 0.03125
-0.21875 0.18750 -0.09375 0.18750
とな る。
( 2 - 2 )
こ れを丸める と D C係数について は一 2 2 5 とな る。 すなわち 、 0 0 0 0
0 0 0
0 0 0 0
と な る。
こ れを対 して逆量子化を行う D C係数に対 して は 2 を掛ける こ とで 一 4 5 0 とな る 。 すなわち
-450 - 2 0 0
0 0 0 0
0 2 0 0
0 0 0 0
したがって 上記量子化 逆量子化に よ っ て上記入力 に対する D C係数の変化分は、 E r C X 2 = - 1 と なる。 この D C係数の変化 分は、 係数面の処理によ る変化分と なる 。
( 2 - 3 )
ま た、 逆量子化後、 D C以外では A C係数の A C 1 = — 2 、 A C
9 = 2が残る こ こで、 A C係数分の
0 -2 0 0
0 0 0 0
0 2 0 0
0 0 0 0
を 、 逆変換 ( I D C T ) する と
0.000000 0.000000 0.000000 0.000000
- 1.306563 - 0.541196 0.541196 1.306563
- 1.306563 - 0.541196 0.541196 1.306563
0.000000 0.000000 0.000000 0.000000 と な る。
( 2 - 4 )
ま た、 D C係数分の
-450 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
を 、 逆変換 ( I D C T ) する と
-112.5 -112.5 -112.5 -112.5
-112.5 -112.5 - 112.5 -112.5
-112.5 -112.5 -112.5 -112.5
-112.5 -112.5 -112.5 -112.5
とな る。 ま た、 この D C係数分の逆変換 ( I D C T ) の結果は
-112.5 -112.5 -112,5 -112.5
112.5 -112.5 -112.5 -112.5 (2-4-1) 112.5 -112.5 -112.5 -112.5
112.5 -112.5 -112.5 -112.5
112.5 -112.5 -112.5 -112.5 0 0 0 0
■112.0 -112.0 -112.0 -112.0 + -0.5 -0.5 - 0.5 -0.5
■112.0 -112.0 -112.0 -112.0 - 0.5 -0.5 - 0-5 -0.5
112.5 -112.5 -112.5 -112.5 0 0 0 0 • · (2-4-2) のよ う に表すこ とができ る
( 2 - 5 ) こ こで上記(2- 3) の A C係数分の I D C T出力 と 、 上記(2- 4 - 2 の第 2項の D C係数分の I D C T 出力すなわち
0 o 0 0
- 0.5 - 0.5 -0.5 -0.5
o
o o
- 0.5 -0.5 -0.5 -0.5
0 0 ◦ 0 0
と を加えて丸めを行う と 、
0 0 0 0
- 2 一 1 0 1
一 2 一 1 0 1
0 0 0 0
とな り 、 これを変換する と 、
0.000
0.000 0.000 0.000
1.000 0.000 0.158
0.000 0.000
と な る。
( 2 - 6 )
ま た、 上記(2- 4) の境界点の処理 (
よ る変化分は 、
-0 - 5 -0.5 -0.5 -0.5
0 0 0 0
0 0 0 0
- 0.5 -0.5 -0.5 -0.5 とな る こ れを変換する と 、
一 1 0 0 0
0 0 0 0
一 1 0 0 0
0 0 0 0
とな る。
( 2 - 7 )
上記(2-2) と (2-6) よ り 、 D C係数の変化分は、 係数面、 空間面 合わせて — 2で、 1 量子化ステ ヅブ分である 。 こ こで、 新た に、 A C係数の A C 8 に空間面の処理によ る成分が現れる 。
( 2 - 8 )
ま た、 上記(2-6) の結果よ り 、 上記(2-5) は、 そのま ま丸めが入 り 、
- 1 - 2 0 0
0 0 0 0
1 2 0 0
0 0 0 0
とな り 、 これが結果的に このサイ クルの触媒になる 。
( 2 - 9 )
以下同様に 2回目の変換を行う と 、 D C = — 4 5 1 、 A C 8 = - 1 、 これを 2で量子化する と D C = — 2 2 5 . 5 、 A C 8 = - 0 . 5 と なる 。
( 2 - 1 0 )
こ れを丸めて D C = — 2 2 6 、 A C 8 = — 1 と し 、 逆量子化する と D C = — 4 5 2 、 A C 8 = — 2 と な り 、 変化分はそれぞれ E r e x 2 = — 1 とな る 。
( 2 - 1 1 )
上記 D C と A C 8 の係数か ら触媒の成分が重な る部分を差 し引 い て 、
- 451 0 0 0
0 0 0 0
- 3 0 0 0
0 0 0 0
と し 、 こ れを逆変換する と 、
-113.5 -113.5 -113.5 -113.5
-112.0 -112.0 -112.0 -112.0
-112.0 -112.0 -112.0 -112.0
-113.5 -113.5 -113.5 -113.5
この境界点の処理に よる変換分は上記(2- 6) に示 した通 り となる 。 ( 2 - 1 2 )
上記(2- 10)と 、 上記(2- 11)の結果よ り 、 係数面、 空間面合わせて D C , A C 8 と も に — 2 で、 1 量子化ス テ ップ分の変化 と なる。 ( 2 - 1 3 )
触媒の成分は、 変換—逆変換を繰 り返 して も変化 しな い。 これ以 降、 上記(2-8) 〜 (2-13)のサイ クルを繰 り 返す。
次に、 上述 した具体例の よ う に丸め誤差が蓄積する蓄積サイ クル の条件につ いて説明する 。
先ず、 蓄積サイ クルの条件を考え る。 上述 した具体例で示 したよ う に 、 蓄積サイ クル に関する係数のかかわ り は、 非常に複雑である , こ こ では、 画像強調 パタ ー ンの洗い出 しのため幾つかの条件を求め る。 これはまた 、 本発明実施例にお ける丸め誤差の蓄積に対する対 策の考えかた ( すなわち偶数方向丸めを用 いるよ う に した こ と ) を 説明する ものであ る 。
第 1 に 、 変換と逆変換 ( D C T - I D C T ) は、 十分な演算精度 を持って い るため、 丸め にお ける境界点以外に誤差の蓄積はない。 したがっ て 、 蓄積サイ ク ル に関与するのは丸めにお ける境界点のみ であ る。
第 2 に 、 係数面の境界点の処理に よる変化分に、 空間面の境界点 の処理に よ る変化分を加え合わせる こ と で、 係数面で再起的に 1 量 子化ステ ップのシ フ 卜 を も た らす。 前者の処理の変化分に、 空間面 の境界点以外の成分が含ま れる と 、 D C T — I D C Tの対称性か ら 空間面の丸めによ る境界点の増加分が、 再度係数面境界点に戻っ て こな く な り 、 蓄積サイ クルが切れる 。 したがって 、 蓄積サイ クルが 起こ るのは、 空間面の境界点が係数面で 、 かつ係数面のそれが空間 面のそれである と きだけである。
第 3 に 、 例えば、 量子化ステ ップサイ ズ q x x を 3 と した と き に は、 図 3 に示すよ う に、 誤差の分布はデジ ッ 卜 を対象と した誤差の 範囲 に収ま る。
なお、 q x x = 4 の誤差の範囲を — Δ Ζ 2 〜 Δ / 2 と すれば、 こ の図 3 にお いて q x x = 3 の と き 、 誤差の範囲は、
( - 3 / 4 ) X ( Δ / 2 ) 〜 ( 3 / 4 ) X ( Δ Ζ 2 )
とな り 、 q x x = 2 の誤差の範囲を 一 / 2 〜 Δ ' Ζ 2 と すれば ,
( 一 3 / 2 ) X ( Δ ' ./ 2 ) ~ ( 3 / 2 ) X ( Δ ' / 2 ) とな る。 第 4に 、 蓄積サイ クル を起こすために 、 係数面の境界点処理に よ る変化分が、 空間面でどの範囲に収ま るべ きかを考える と き 、 例え ば、 係数面の境界点の誤差 0 . 5 の影響が、 空間面で 2 . 5 になつ た と する 。 空間面ではこ れを丸めて 3 . 0 とするが、 こ の と き丸め に よ る誤差は◦ . 5 であ る 。 D C T , I D C Tは、 対称で可逆であ るか ら 、 次の係数面では E = 0 . 5 / 5 = 0 . 1 とな り 、 境界点の 蓄積サイ クルが断ち切れる。
—方、 飛び越 し にな ら な いで空間面の境界点にな る と きは、 同様 の計算で E = 0 . 5 とな り 蓄積サイ クル を断ち切 ら ない 。
すなわち、
係数面 → 空間面 係数面 飛び越すと き Erc = 0.5 (ex)2.5→ 3.0 Ers = 0.5 E = 0.1 飛び越さな い と き Erc = 0.5 ( ex)0.5→ 1.0 Ers = 0.5 E = 0.5 とな る。
したがって、 係数面の誤差によ る影響が、 空間面で 1 量子化ステ ヅ プ以内 に収ま る時のみ、 蓄積サイ クルが起きる こ とが判る 。 q d c ≠ q a cの と き は、 大き いほうの量子化が、 この量子化である。 上述した具体例に示 し た実画像で得た結果は、 全てこ れに したが つ て いる 。
上述 したよ う な こ とか ら 、 D C係数に よ っ て空間面出力が片極性 にシ フ ト し、 その結果、 空間面の丸めに よ る境界点の処理が、 片極 性に転化される こ と によ り 、 誤差の平均値を シフ ト させる。 これが 単調増加 ( 減少 ) を引き起こ す大きな要因である。
上述したよ う に、 蓄積サイ クルは、 整数演算 (非線形処理) であ る丸めと き に発生 し 、 係数面の各係数、 空間面サンプルがう ま く 関 係 し あっ た非常に特殊な ケースであ り 、 確率的に低いケースである ま た 、 確率的に低い こ と を裏付ける よ う に 、 この現象は どのよ う な 絵柄でも常に起こ る と い う わ けではな く 、 実画像で確認 した画像強 調パター ンは前記蓄積サイ クルの条件にお いて述べた第 1 , 第 2, 第 3 の条件にお けるパタ ー ン とな る 。
上記蓄積サイ クルが起き る原因は、 正負極性によ って誤差の分布 をバラ ン ス させる方法 ( すなわち無限大方向丸め ) を採用 した こ と にあ る。 変換符号化を利用 した ビッ 卜 レー ト リ ダク シ ョ ン システム では、 係数面の極性と空間面の極性間に は D C係数以外に密接な対 応関係はな く 、 あ る係数の組み合わせに よ る境界点の処理の影響が 空間面で常に片極性に転化さ れる場合があ り 、 その結果、 誤差の分 布がバラ ンスする方向に傾かず、 一方向に偏 り を持ち、 これが蓄積 さ れるのである。
こ のよ う に、 上記蓄積サイ クル に よっ て特定画像パタ ー ンの単調 増加、 減少のサイ クルが構築され、 変換の演算精度によ らずに完全 再 成が実現できな いよ う になる。
さ らに、 単調増加, 減少によって強調さ れる画像パタ ー ンは、 幾 何学的パター ン にな るので、 マルチ ジェ ネ レーシ ョ ンの画質を大ぃ に損なう 。
こ のよ う な こ とか ら、 本実施例では、 空間面で極性に関係な く 誤 差の分布をバラ ンス させる方法を取るよ う に して い る。
すなわち 、 本実施例では、 空間面の出力丸めに偶数値を基準に誤 差の分布をバラ ンス させる方法、 すなわち偶数方向丸め を採用する よ う に して いる。 なお、 空間面で極性に関係な く 誤差の分布をバラ ンス させる方法と し ては、 空間面の出力丸めに奇数値を基準に誤差 の分布をバラ ン ス させる方法、 すなわち奇数方向丸めを採用する こ と も可能である
この対策によ っ てディ ジタルダイ レ ク 卜 ダビングは、 完全再構成 が実現さ れる こ と になる 。
次に、 図 1 0 には、 D C Tの演算精度が 1 2 ビッ ト , 1 3 ビッ ト 1 4 ビヅ 卜 の時に、 ス ト レ ー 卜 にダ ビン グを繰 り 返 した場合の ビデ ォ信号の Y信号と P R, P B信号の S Z Nの劣化の様子 (複数回ダ ビン グ後の シ ミ ュ レ ーシ ョ ン結果) を示 して いる 。 評価画像は圧縮 率 1 / 2である。 このシ ミ ュ レー シ ョ ンでは、 D C T , I D C Tの 演算は浮動少数点で行い 、 整数への丸めは I E E E 7 5 4規格によ る関数を使用 して い るが、 量子化では単純丸め (正方向丸め) の方 法で丸めを行っ て い る。
ま た、 図 1 1 には無限大方向丸め を用 いた場合の シ ミ ュ レ一シ ョ ン結果を 、 図 1 2 に は本実施例にお ける偶数方向丸めを用 いた場合 の シ ミ ュ レ 一 シ ョ ン結果を示す。
図 1 1 に示す無限大方向丸めの場合では、 図 1 0 に示 した正方向 丸めの場合に比べて僅かではあるが、 S / N劣化のカーブが緩やか にな る。
こ れに対 して 、 図 1 2 に示す本実施例 にお いて採用する偶数方向 丸めでは、 上記正方向丸めや無限大方向丸めに比べて S Z Nの劣化 は緩やかにな り 、 早い世代で収束 し てい る。 特に、 図 1 3 に示すよ う に、 D C Tの演算精度が少な い 1 2 ビ ヅ 卜 の場合にその効果が大 き い こ とが判る。 なお、 図 1 3 には、 同時に無限大方向丸め と正方 向丸めの シ ミ ュ レー シ ョ ン結果も示 して い る。
ま た、 上述 した実施例では D C T を用 いて いるが、 いわゆる ゥ ェ —ブレ ッ ト ( Wavelet ) 変換によ る ビッ ト レー ト リ ダク シ ョ ンは、 D C Tに比べて画質が優れて いる 。 特に 、 いわゆるハール ( Haar) 基底によ る Wave let 変換は整数の加減算で実現でき 、 D C T と同 じ よ う にブロ ッ ク分割 した後に周波数分解でき るので、 ハ ー ド ウ エア 規模も小さ く でき る 。
そ こで、 上述 し た実施例での D C Tの部分を上記 Haar基底によ る 8 X 8の 1 0分割離散ウ エ ーブレ ヅ 卜変換 ( Discrete Wavelet Tra nsform: D W T ) に置き換えて シ ミ ュ レー シ ョ ンする。 ただ し、 D C T とは異な り 、 D W Tの演算精度は最大 1 4 ビヅ 卜であ り 、 この 場合は丸め誤差は発生 しない。 D W T , I D W Tも量子化と同 じ丸 め方式と したので、 3力 所の丸めがある 。 これらの丸めを上記正方 向丸め、 無限大方向丸め、 偶数方向丸めで比較する 。
その結果は、 図 1 4〜図 1 7 に示すよ う になる。 なお 、 評価画像 と圧縮率は上述の D C T同様である 。 ま た 、 図 1 4 〜図 1 7 には、 ビデオ信号の Y , P B, P R信号を用い 、 D W Tの演算精度が 1 2 ビヅ 卜 , 1 3 ビヅ 卜 , 1 4 ビヅ ト の時に 、 ス ト レ ー ト にダビングを 繰 り 返した場合の シ ミ ュ レ ー シ ョ ン結果を示す。 図 1 4 には正方向 丸めを用 いた場合を 、 図 1 5 には無限大方向丸めを用いた場合を、 図 1 6 には偶数方向丸め を用 いた場合の シ ミ ュ レ ー シ ョ ン結果を示 す。 図 1 7 には、 D W Tの演算精度が少な い 1 2 ビ ヅ 卜 の場合の、 偶数方向丸め と無限大方向丸め及び正方向丸めのシ ミ ュ レー シ ョ ン 結果を示 している 。
こ の図 1 4〜図 1 7か ら 、 正方向丸め と無限大方向丸めは同 じ傾 向を示 し 、 D W T演算精度 1 2 ビ ッ トでは S / Nの劣化が大き く 世 代が重な る に従っ て演算誤差が累積 して い く こ とが判る 。 こ れに対 して 、 偶数方向丸めでは 1 2 ビ ヅ ト でも S / Nの劣化は 少な く 3 ~ 4世代目で収束 して い る 。
こ のよ う な こ とか ら 、 偶数丸め ( 奇数丸め ) は、 演算精度が不足 して いる場合でも 、 従来の丸め (正方向丸めや無限大方向丸め) に 比べて、 誤差の累積が少な い こ とが判る 。 こ れは、 切 り 捨て る ビ ヅ 卜 が少な いほど、 0 . 5 の発生する確率が高いので、 偶数方向丸め の効果が現れるか ら であ る 。
高画質を要求さ れ る ビ 、、 ' 卜 レー 卜 リ ダ ク シ ョ ン を行う V T Rでは 十分な演算精度を確保 し 、 さ ら に偶数方向丸めをする こ とが望ま し い。 マ ト リ クス変換のよ う に変換 Z逆変換を繰 り 返す場合にも 、 上 記偶数方向丸めは累積誤差を少な いする ために有効であ る。
上述したよ う に、 本発明実施例においては、 演算精度を十分に持 つた変換符号化で、 偶数方向丸め ( 又は奇数方向丸め) を用 いる こ と に よ り 、 丸めにお ける特異点であ る境界点の処理によ る影響の蓄 積を阻止する こ とができ . 固定画像パタ ー ンの振幅の単調増加, 減 少が防げる。 この効果に よ り 、 完全再構成すなわち マルチジエネ レ — シ ヨ ン を何度行っ ても画像劣化のない変換系が実現で き る。 言い 換えれば、 高能率符号化復号化装置の複数回の継続接続にお いて 、 画質劣化を最小限に抑えてデ一夕 を一定の値に収束させる こ とがで き る 。
なお、 本発明は、 ディ ジタル V T Rだ けでな く 、 変換符号化に よ る ビ ヅ 卜 リ ダク シ ョ を採用 した シ ステム に共通するので、 例えば オーディ オシステムなど他の同様の シス テム にも適用で き る。
最後に 、 蓄積サイ クルで画像が強調さ れるパター ン ( 画像強調パ ター ン ) を例に挙げる。 係数面の境界点の条件を逆変換行列に入力 して 、 蓄積サイ クルが 起き た場合のパタ ー ンの バ リ エー シ ョ ン は、 例えば、 図 1 8 に示す ( 1 ) のグループ〜図 2 1 に示す ( 4 ) のグループが考え られる 。 なお、 上記 ( 1 ) ~ ( 4 ) のグル ープにお いて は、
再量子化ステツフ。 入力係数値
( 1 ) q dc = q ac= 1 D C = - 1 1 , A C x = - l + l
( 2 ) q dc= q ac = 2 D C = - 1 + 1 A C x = - l + l ( 3 ) q dc= 4 D C = - 2 +2
(4 ) q dc= 2 , q ac= 1 D C = -2 ~ +2 , A C x = _ l 1 と してい る。
ま た、 入力係数値フ ォ ーマ ッ ト は、
DC , AC1 AC2 AC3
AC4 , AC5 ,AC6 , AC7 ,
AC8 , AC9 AC10 AC11
AC12,AC13,AC14,AC15
とする。
蓄積サイ クルは、 整数演算 (非線形処理) であ る丸め と 、 係数面 の各係数、 空間面サ ンプルが、 う ま く 関係 しあっ た特殊なケース で ある 。 例えば、 図 1 8〜図 2 1で示 した ( 1 ) , ( 2 ) , ( 3 ) の グループのパター ン を挙げる こ とができ る。

Claims

請 求 の 範 囲
1 . アナ ロ グ信号がアナ ロ グ /ディ ジタ ル変換さ れたデ ィ ジ夕ル信 号を直交変換する直交変換手段と 、
上記直交変換さ れたデ ィ ジタル信号を逆直交変換する逆直交変換 手段 と を有 し、
上記直交変換と逆直交変換の少な く と も一方に偶数方向の丸め又 は奇数方向の丸めを用 い る こ と を特徴と する高能率符号化復号化装 置。
2 . 上記直交変換後のディ ジ夕ル信号を量子化する量子化手段と 、 上記逆直交変換前のディ ジタル信号を逆量子化する逆量子化手段 と
を設ける こ と を特徴 とする請求の範囲第 1項記載の高能率符号化復 号化装置。
3 . 上記直交変換は離散コサイ ン変換であ り 、 上記逆直交変換は逆 離散コサイ ン変換である こ と を特徴 とする請求の範囲第 1項記載の 高能率符号化復号化装置。
4 . 上記直交変換は離散ウ ェーブレ ッ ト 変換 ( Discrete Wavelet Transform: D W T ') であ り . 上記逆直交変換は逆離散コサイ ン変換 ( Inverse Discrete Wavelet Transform: I D W T ) である こ と を 特徴とする請求の範囲第 1項記載の高能率符号化復号化装置。
PCT/JP1994/001546 1993-09-28 1994-09-20 High efficiency encoding/decoding device WO1995009487A1 (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP94927072A EP0671816B1 (en) 1993-09-28 1994-09-20 Encoding/decoding device with all odd or all even value rounding
US08/428,255 US5604494A (en) 1993-09-28 1994-09-20 Efficient encoding/decoding apparatus
DE69423734T DE69423734T2 (de) 1993-09-28 1994-09-20 Gerät zur kodierung/dekodierung mit rundung auf ausschliesslich gerade oder ungerade werte

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP24118893 1993-09-28
JP5/241188 1993-09-28

Publications (1)

Publication Number Publication Date
WO1995009487A1 true WO1995009487A1 (en) 1995-04-06

Family

ID=17070536

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP1994/001546 WO1995009487A1 (en) 1993-09-28 1994-09-20 High efficiency encoding/decoding device

Country Status (4)

Country Link
US (1) US5604494A (ja)
EP (1) EP0671816B1 (ja)
DE (1) DE69423734T2 (ja)
WO (1) WO1995009487A1 (ja)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002077916A (ja) * 2000-07-13 2002-03-15 Lucent Technol Inc ブロック変換ベースの復号化器で使用される方法と装置

Families Citing this family (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6256349B1 (en) * 1995-12-28 2001-07-03 Sony Corporation Picture signal encoding method and apparatus, picture signal transmitting method, picture signal decoding method and apparatus and recording medium
US5822005A (en) * 1996-01-11 1998-10-13 Tektronix, Inc. Pre-oddification
US6058215A (en) * 1997-04-30 2000-05-02 Ricoh Company, Ltd. Reversible DCT for lossless-lossy compression
US6574371B2 (en) * 1997-06-09 2003-06-03 Hitachi, Ltd. Image decoding method
US7801380B2 (en) * 1997-06-09 2010-09-21 Hitachi, Ltd. Recording medium having recorded thereon coded information using plus and/or minus rounding of images
ES2545066T3 (es) * 1997-06-09 2015-09-08 Hitachi, Ltd. Medio de grabación de información de imágenes
US6456745B1 (en) * 1998-09-16 2002-09-24 Push Entertaiment Inc. Method and apparatus for re-sizing and zooming images by operating directly on their digital transforms
US6393155B1 (en) * 1998-11-04 2002-05-21 International Business Machines Corporation Error reduction in transformed digital data
US6961472B1 (en) * 2000-02-18 2005-11-01 Intel Corporation Method of inverse quantized signal samples of an image during image decompression
JP3763276B2 (ja) * 2001-11-27 2006-04-05 ソニー株式会社 信号処理装置および方法、記録装置および方法、ならびに、記録再生装置および方法
US7197077B2 (en) * 2001-12-20 2007-03-27 Sbc Technologies Resources, Inc. Multilevel data compression using a single compression engine
WO2004075556A1 (ja) * 2003-02-19 2004-09-02 Ishikawajima-Harima Heavy Industries Co., Ltd. 画像圧縮装置、画像圧縮方法、画像圧縮プログラム、及び圧縮符号化方法、圧縮符号化装置、圧縮符号化プログラム、並びに復合化方法、復合化装置、復合化プログラム
JP4569840B2 (ja) * 2007-09-12 2010-10-27 ソニー株式会社 画像符号化装置、画像符号化方法
CN101637026B (zh) * 2007-09-12 2012-03-28 索尼株式会社 图像处理装置和方法
TWI392372B (zh) * 2007-09-12 2013-04-01 Sony Corp Image processing apparatus and image processing method
JP5212373B2 (ja) 2007-09-12 2013-06-19 ソニー株式会社 画像処理装置及び画像処理方法
CN101779467B (zh) * 2008-06-27 2012-06-27 索尼公司 图像处理装置和图像处理方法
EP2192785B1 (en) * 2008-06-27 2014-12-31 Sony Corporation Image processing device and method for computing a quantization factor
BRPI0904320A2 (pt) * 2008-06-27 2015-06-30 Sony Corp Dispositivo e método de processamento de imagem
WO2009157581A1 (ja) * 2008-06-27 2009-12-30 ソニー株式会社 画像処理装置及び画像処理方法
WO2013190346A1 (en) * 2012-06-18 2013-12-27 Abellaoui Lahcen Method and system for quantifying binary words symmetry
US10045024B2 (en) 2013-11-27 2018-08-07 Thomson Licensing Method and device for quantising the floating value of a pixel in an image

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH0485621A (ja) * 1990-07-30 1992-03-18 Matsushita Electric Ind Co Ltd 丸め装置
JPH05167456A (ja) * 1991-12-19 1993-07-02 Ricoh Co Ltd 演算回路

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5014134A (en) * 1989-09-11 1991-05-07 Aware, Inc. Image compression method and apparatus
CA2062200A1 (en) * 1991-03-15 1992-09-16 Stephen C. Purcell Decompression processor for video applications
US5379122A (en) * 1992-10-02 1995-01-03 Xerox Corporation Decompression of standard ADCT-compressed images
TW224553B (en) * 1993-03-01 1994-06-01 Sony Co Ltd Method and apparatus for inverse discrete consine transform and coding/decoding of moving picture

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH0485621A (ja) * 1990-07-30 1992-03-18 Matsushita Electric Ind Co Ltd 丸め装置
JPH05167456A (ja) * 1991-12-19 1993-07-02 Ricoh Co Ltd 演算回路

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP0671816A4 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002077916A (ja) * 2000-07-13 2002-03-15 Lucent Technol Inc ブロック変換ベースの復号化器で使用される方法と装置

Also Published As

Publication number Publication date
US5604494A (en) 1997-02-18
EP0671816B1 (en) 2000-03-29
EP0671816A1 (en) 1995-09-13
DE69423734D1 (de) 2000-05-04
EP0671816A4 (en) 1996-04-24
DE69423734T2 (de) 2000-07-20

Similar Documents

Publication Publication Date Title
WO1995009487A1 (en) High efficiency encoding/decoding device
KR101247011B1 (ko) 넓은 범위 계수의 적응형 코딩 및 디코딩
JP4906855B2 (ja) 変換ブロックの効率的なコーディングおよびデコーディング
US7729423B2 (en) Fixed bit rate, intraframe compression and decompression of video
JPH0485621A (ja) 丸め装置
US5719961A (en) Adaptive technique for encoder and decoder signal transformation
KR20090122390A (ko) 데이터 소스에 의해 생성된 비트 심볼들을 처리하기 위한 방법 및 장치, 컴퓨터 판독가능 매체, 컴퓨터 프로그램 요소
US7190884B2 (en) Signal recording apparatus and method, signal reproducing apparatus and method, medium, and information assembly
JP2002077916A (ja) ブロック変換ベースの復号化器で使用される方法と装置
JPH07153195A (ja) ディジタル記録装置
KR100480785B1 (ko) 그레이코드를 이용한 비트플레인 부호화 및 복호화 방법및 장치
JP3252567B2 (ja) 高能率符号化復号化装置
JPH1188183A (ja) ウェーブレット変換装置およびその方法並びにウェーブレット逆変換装置およびその方法並びに画像符号化装置およびその方法並びに画像復号化装置およびその方法
JPH02131671A (ja) 画像データ圧縮方法
JP3235761B2 (ja) 高能率符号化装置
JP2669278B2 (ja) 映像信号処理装置
JP3591663B2 (ja) 画像信号伝送方法及び画像信号伝送装置
JP2001128176A (ja) 画像圧縮および伸張装置
JPS62193382A (ja) 画像信号符号化復号化方式とその装置
JPH02272889A (ja) 画像の直交変換符号化装置
JPH0292162A (ja) 高能率符号化装置
JPH0412077B2 (ja)
Wang et al. High-fidelity image compression with nearly lossless arithmetic coding
JPH0530363A (ja) 画像符号化復号化装置
JPS61135283A (ja) 医用画像デ−タ伸張装置

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): JP US

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): DE FR GB

WWE Wipo information: entry into national phase

Ref document number: 1994927072

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 08428255

Country of ref document: US

121 Ep: the epo has been informed by wipo that ep was designated in this application
WWP Wipo information: published in national office

Ref document number: 1994927072

Country of ref document: EP

WWG Wipo information: grant in national office

Ref document number: 1994927072

Country of ref document: EP