JPH10224791A - Adaptive quantization system for orthogonal transformation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To encode an image with high image quality by a quantization characteristic with a small dynamic range and to send the image with a small quantization number. SOLUTION: An orthogonal transformation device 1 applies orthogonal transformation to an input signal to generate a transformation signal. A separator 2 separates the transformation signal into a low frequency transformation signal and a high frequency transformation signal. An adaptive quantizer 6 applies prediction coding processing to the low frequency transformation signal through adaptive quantization with a dynamic range of the transformation signal. The amplitude of the high frequency transformation signal is limited depending on the magnitude of maximum quantization noise and a quantizer 7 obtains a prediction error signal with a dynamic range of the transformation signal, for quantization.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a predictive coding system for digitizing a television signal, quantizing a prediction error signal of the television signal, and compressing and transmitting the data, and an orthogonal transform system for orthogonally transforming and encoding an image signal. In particular, the present invention relates to a quantization method used in an apparatus that performs orthogonal transformation on a prediction error signal, performs quantization, and performs encoding transmission.

[0002]

2. Description of the Related Art Conventionally, as an encoding system (prediction error orthogonal transformation encoding system) combining differential encoding (prediction encoding) and orthogonal transformation encoding, for example, the one shown in FIG. 3 is known. As an example of a typical method of orthogonally transforming a prediction error signal, quantizing a transformed signal, and encoding and transmitting the signal,
For example, there is TTC standard JT-H261 (issued by the Telegraph and Telephone Technical Committee).

Referring to FIG. 3, in the illustrated system, a subtracter 31 subtracts an 8-bit image signal and an 8-bit motion-compensated inter-frame prediction signal to obtain a 9-bit difference signal (prediction error). Signal). Then, the orthogonal transformer 32 converts the difference signal into a DC of 8 × 8 blocks.
Orthogonal transform by T (discrete cosine transform) is performed to obtain 64 types of 12-bit transform coefficients. In the quantizer 33, a predetermined quantization characteristic (a dynamic range of a maximum of 12 bits and a range of -20
48 to 2047), the quantized transform coefficient is code-converted by the coding converter 34 and transmitted.

[0004] The quantized transform coefficients are also supplied to an inverse orthogonal transformer 35, where they are subjected to inverse orthogonal transform. The output of the inverse orthogonal transformer 35 is added to the prediction signal by the adder 36, and is added to the predictor 37. Then, the predictor 37 outputs the above-mentioned 8-bit motion compensation inter-frame prediction signal.

On the other hand, as a conventional adaptive quantization method for predictive coding, for example, the one shown in FIG. 4 is known.

Referring to FIG. 4, an 8-bit input signal is supplied to a subtractor 41 and an adaptive quantizer 42,
Then, the prediction signal is subtracted from the input signal by an 8-bit modulo operation to obtain an 8-bit prediction error signal. The adaptive quantizer 42 has an 8-bit dynamic range quantization characteristic that is 1 bit smaller than 9 bits, quantizes the prediction error signal to obtain a provisional quantization output, and quantizes the quantization noise and the input signal. The provisional local decoded signal is obtained, and it is determined whether or not this value exceeds or falls below the 8-bit dynamic range. The above quantization level is selected and output as a quantization output. This quantized output is output from the coding converter 4
3 and transmitted.

[0007] The quantized output is also provided to an adder 44, where it is added to the prediction signal. Then, the output of the adder 44 is given to the predictor 45, and the predictor 45 outputs a prediction signal.

As another quantization technique, for example, the upper and lower limits of the amplitude of the input signal are limited in advance by the limiter by the magnitude of the maximum quantization noise, and the quantization is performed with a quantization characteristic of a dynamic range smaller by one bit. Techniques for performing the conversion are known. In other words, this method is known as aliasing quantization, and although there is a point that the input signal is amplitude-limited by the limiter, it is not necessary to make a special determination in the quantizer. That is, it is only necessary to have a function similar to that of a normal quantizer, and a limiter for limiting the amplitude of an input signal is required. However, a simpler configuration than the adaptive quantization system shown in FIG.

[0009]

In order to reduce the coding transmission bit rate by using a quantizer in which the dynamic range of the quantization characteristic is reduced by one bit while maintaining the same image quality, the prediction error is reduced. Even if an adaptive quantizer for predictive coding is applied to a quantizer used in the orthogonal transform coding scheme, as described above, in the prediction error orthogonal transform coding scheme, the prediction error is calculated every 8 × 8 block. Is performed, and orthogonal transform / inverse orthogonal transform is performed before / after the quantizer. For this reason, in the provisional local decoded signal of 64 pixels decoded for each block, each pixel is affected by the quantization noise of 64 transform coefficients, and 64 quantization noises are weighted and added by the transform coefficients. Will be affected by the resulting quantization noise. Therefore, the locally decoded signal exceeds the dynamic range.

It is not easy to determine which of the transform coefficients exceeds the dynamic range due to the influence of the quantization noise. For example, as a simple brute force method, 64 transform coefficients are temporarily set as However, it is not limited to the effect of one coefficient, and the brute force calculation becomes enormous, making it difficult to perform the processing in real time.

In the adaptive quantization method for predictive coding,
Since the adaptive local quantization signal is determined by determining whether the dynamic range is exceeded or not and adaptive quantization is performed, the circuit scale becomes large, and the quantization of the 64 types of transform coefficients in the orthogonal transform coding is all performed by the adaptive quantization. If so, there is a problem that the scale of the apparatus becomes large.

An object of the present invention is to apply an adaptive quantizer to a quantizer for predictive error orthogonal transform coding so as not to increase the scale of the apparatus, and to achieve a quantization characteristic having a dynamic range smaller by one bit than the conventional one. An image having the same image quality as conventional quantization can be encoded without deterioration of load noise, and transmission can be performed with a smaller number of quantization levels (number of allocated bits: that is, a smaller number of encoded transmission bits) as compared with the related art. It is to provide a quantization method.

[0013]

According to the present invention, determination is made not by using a temporary local decoded signal in the time domain but by using a temporary local decoded signal in the domain of transform coefficients. Is converted to In other words, when the positions of the orthogonal transform and inverse orthogonal transform of the prediction error orthogonal transform coding method are moved by equivalent transform, the input signal is first orthogonally transformed to obtain transform coefficients, and then the predictive coding for each transform coefficient is performed. Is performed. With this configuration, an adaptive quantizer that can reduce the dynamic range by one bit is used as a quantizer for predictive coding of a transform coefficient.

An input signal having a size of 8 bits is an 8 × 8 D signal.
The orthogonal transform is performed by CT, thereby obtaining 64 kinds of transform coefficients having a size of 11 bits. Each transform coefficient is subtracted from a prediction signal of the transform coefficient by a subtractor of a modulo operation,
It becomes an 11-bit difference signal. Adaptive quantization is performed by using a quantization characteristic having a dynamic range of 11 bits to determine whether a temporary local decoded signal of a transform coefficient exceeds a dynamic range of 11 bits.

If adaptive quantization is applied to all 64 types of transform coefficients, the circuit scale becomes large. To improve this, the statistical property of the magnitude of the transform coefficients is used. That is, since the TV signal has a strong spatial and temporal correlation, the frequency component of the image signal has a large low band and a small frequency component in a high band. When encoding a TV signal, this correlation is used. However, even if encoding such as prediction encoding is performed, this correlation still remains in the prediction error signal, and the amplitude of the low-frequency component is statistically calculated. Is large, and the amplitude of the high frequency component is small.

For this reason, the transform coefficient having a large amplitude of the low frequency component is quantized by the adaptive quantization technique. On the other hand, even if the upper limit of the transform coefficient of the high-frequency component is small, the dynamic range of the transform coefficient is hardly affected even if the upper and lower amplitudes are limited by the limiter. The circuit size is reduced by using aliasing quantization.

According to the configuration of the present invention, the conversion signal of the prediction error signal can be quantized and coded and transmitted with the quantization characteristic of the dynamic range smaller by one bit than the conventional one, and the reproduced image is the same as the conventional one without overloading. The image can be decoded.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings.

Referring to FIG. 1, an 8-bit digital input signal X (-128 to 127) is supplied to orthogonal transformer 1, and is composed of 64 pixels (X1) of "8 samples" .times. "8 lines".
1 to X88) as one block, and performs orthogonal transformation for each block to output 64 types of conversion coefficients (Xt11 to Xt88). These transform coefficients have a dynamic range of 11 bits. In the separator 2, the transform coefficients are divided into two groups, a transform coefficient of a low-frequency component and a transform coefficient of a high-frequency component, according to the frequency component of the transform vector. The band transform coefficient is supplied to a subtractor 4 and an adaptive quantizer 6, and the high band transform coefficient is supplied to a limiter 3.

The limiter 3 limits the upper and lower amplitudes of the converted signal in accordance with predetermined characteristics and supplies the converted signal to the subtracter 5.

The orthogonal transformer 13 has the same function as the orthogonal transformer 1 and converts the prediction signal P output from the predictor 12 into 64.
The orthogonal transform is performed for each block of pixels to output 64 types of conversion prediction signals Pt having a dynamic range of 11 bits. Then, these transform prediction signals Pt are supplied to the adder 10 and are divided into two groups of low-pass transform coefficients and high-pass transform coefficients by the separator 14. The low-frequency transformed prediction signal is supplied to a subtractor 4, and the high-frequency transformed prediction signal is supplied to a subtractor 5.

The subtractor 4 subtracts each low-frequency transform prediction signal from each input low-frequency transform coefficient by an 11-bit modulo operation for each block, outputs an 11-bit low-frequency transform difference signal Et, and performs adaptive quantization. To the vessel 6.

The subtractor 5 subtracts each high-frequency transform prediction signal from each input high-frequency transform coefficient by an 11-bit modulo operation for each block and outputs each 11-bit high-frequency transform difference signal. Supply to

The dynamic range of the difference conversion coefficient signal Et output from each of the subtracters 4 and 5 has the same 11-bit dynamic range as the input conversion coefficient signal Xt.
Similarly, the dynamic range of the adaptive quantizer 6 and the quantizer 7 for quantizing the difference transform coefficient signal may be 11 bits, which is one bit less than the conventional 12 bits.

The difference transform coefficient signals Et output from the subtracters 4 and 5 are supplied to adaptive quantizers 6 and 7 respectively, and the difference transform coefficient signals are converted into respective quantized signals corresponding to respective transform coefficients for each block. Quantization is performed according to the quantization characteristics, and a quantization output Qt is output.

The adaptive quantizer 6 has an 11-bit dynamic range quantization characteristic, quantizes the low-frequency difference conversion signal Et in accordance with a predetermined quantization characteristic, and sets a corresponding quantization level to a provisional quantization. output Qt quantization level up to output as a Q ⁺ and down one quantization level Q ^- is a combined output. Then, the adaptive quantizer 6 obtains a quantization error signal Nt (= Qt-Et) by subtracting the low-frequency difference conversion signal Et from the temporary quantization output Qt, adds the input error signal Nt to the input conversion signal Xt, Local decoding conversion signal Yt
(= Pt + Qt = Xt + Nt). Further, the adaptive quantizer 6 determines whether the provisional local decoded transform signal is within the dynamic range of 11 bits, and if it is within the range, the provisional quantization output Qt is used as a true quantization output. When the output and overflow occur, the next lower quantization level is selected to be a true quantization output. On the other hand, when an underflow occurs, the next higher quantization level is selected and true quantized output is supplied to the combiner 8.

The quantizer 7 has an 11-bit dynamic range quantization characteristic, quantizes the high-frequency difference transform signal in accordance with a predetermined quantization characteristic, and outputs a high-frequency group quantized output. It is supplied to the synthesizer 8.

The synthesizer 8 combines the low band and the high band to obtain one block of the transformed quantized output, and obtains the code converter 9 and the adder 1
0 and supply. The code converter 9 code-converts each level of the quantized output together with other information into a code for transmission on a block-by-block basis and sends the code to the transmission path.

The adder 10 modulo-adds the quantized output Qt and the predictive transform coefficient Pt to obtain an 11-bit local decode transform coefficient Yt. The dynamic range of the local decoding transform coefficient signal is 11 bits, like the dynamic range of the input transform coefficient signal Xt.

The inverse orthogonal transformer 11 has an inverse transform characteristic of the orthogonal transform characteristic of the orthogonal transformer 1, and inversely transforms the locally decoded signal for each block to output a locally decoded signal Y. The local decoded signal Y is clipped to suppress an overflow due to an operation error of the orthogonal transformation, subjected to amplitude limitation to the same 8-bit dynamic range as the input signal, and supplied to the predictor 12.

The predictor 12 obtains the next prediction signal for each block from the locally decoded signal in accordance with the prediction characteristics and outputs it. The predictor 12 has a function of performing motion compensation prediction, finds an optimal motion vector for the next input block by a matching method for each block, and outputs a motion-corrected prediction signal. The motion vector is encoded together with the quantized output signal and sent to the receiving side.

Note that it is conceivable to perform the prediction without performing the inverse orthogonal transform on the locally decoded transform signal. In this case, however, efficient prediction cannot be performed because the signal is transformed in the domain of the transform coefficient.

The orthogonal transformer 13 has the same function as that of the orthogonal transformer 1, and has 8 samples × discrete cosine transform DCT.
The cosine transform is performed with the pixels of eight lines as one block, and an 11-bit predictive transform coefficient Pt is output and supplied to the separator 14 and the adder 10.

Next, the conversion characteristics of the orthogonal transformers 1 and 13 will be described.

In the 8-row, 8-column DCT transform, a two-dimensional discrete cosine transform that can be separated into an 8 × 8 one-dimensional transform is performed.
The signal of one block of eight rows and eight columns is represented by f (x, y) (X11 to X8
8), and the conversion output coefficient of 8 rows and 8 columns is F (u, v) (Xt
11 to Xt 88), the conversion output F (u, v) uses the one given by Equation 1 shown in ITU-T Rec. H.261.

[0036]

(Equation 1) Note that for the block to be converted, x = 0 corresponds to the left end of the block, and y = 0 corresponds to the upper end of the block.

The inverse transform characteristic of the inverse orthogonal transformer 6 is as shown in Equation 2.

[0038]

(Equation 2) When this conversion is performed, the conversion coefficient F becomes a signal having a widened 3-bit dynamic range. That is, when the signal f has 8 bits, the conversion coefficient F has a dynamic range of 11 bits. In the conventional difference encoding, the difference signal E obtained by calculating the 8-bit prediction signal P from the 8-bit input signal X is 9 bits, and the transform coefficient obtained by orthogonally transforming the 9-bit difference signal is 12 bits. Of the dynamic range. In other words, in the conventional example, the difference transform coefficient has a 12-bit dynamic range, and the quantization characteristic requires a 12-bit dynamic range to quantize the transform coefficient.

On the other hand, in the present invention, an 11-bit transform coefficient obtained by orthogonally transforming an 8-bit signal is modulo-subtracted by 1
What is necessary is just to quantize the 1-bit differential transform coefficient signal Et,
The dynamic range of the quantization characteristic is only a half (1 bit less) 11-bit range as compared with the related art.

Next, a specific example of the adaptive quantizer is shown in FIG.
This will be described with reference to FIG.

The 11-bit (-1024 to 1023) difference conversion signal E input to the input terminal 26 of the adaptive quantizer 6
t is supplied to the quantizer 21 and the subtractor 22. The quantizer 21 has an 11-bit dynamic range, quantizes an input signal according to a predetermined quantization characteristic, and outputs an 11-bit quantized output quantized to a predetermined level. Further, the quantizer 11 also outputs a signal Q ⁺ of a quantization bell one level higher than the quantization output Q and a signal Q ⁻ of a quantization level one level lower than the quantization output Q, and supplies them to the switch 23. I do.

The subtractor 22 calculates the input signal E from the quantized output Q.
By subtracting t, the quantization noise N added by quantization is obtained and supplied to the adder 24. The adder 24 adds the input signal Xt input from the input terminal 27 and the quantization noise N to obtain 1
A 2-bit temporary local decoded conversion signal Yt is obtained.

The decision circuit 25 determines that the provisional local decoded conversion signal is 1
It is determined whether an overflow or an underflow has occurred within or beyond the 1-bit dynamic range, and a switching control signal is supplied to the switch 23 according to the determination result. The determination can be made by looking at the state of the upper two bits. Upper 2 bits are “01”
If it is "10", it is understood that it is an underflow, and if it is "00" or "11", it is understood that it is a dynamic range (-1024 to 1023) of 11 bits.

The selector 23, when the temporary local decoded converted signal Yt overflows in response to the switching control signal, Q ^- of selecting a quantization level, in the case of underflow, Q ⁺ quantization Select a level. If it is within the range, the quantization level of Q is selected and output to the output terminal 28.

The quantization characteristic is represented by coordinates (E, Q), and E
When the quantization level of Qi is output in response to the quantization input in the range of iB to EiT, the quantization characteristic of the quantizer 21 is such that a point satisfying (Qi, Qi) always exists. Having. In other words, the left end EiB of a certain quantization level is above the straight line of Q = E, and the right end EiT is below the straight line of Q = E.

When the quantizer 11 has such a quantization characteristic, the input / output conversion characteristic is such that the input signal Ei is ｛EiBi
If the signal is in the range of (EiT), the quantization level of Qi is output. At this time, the quantization noise (Qi-Ei) is a positive or negative value. When the input is in the range of EiB to Qi, the positive quantization noise is obtained.
Negative quantization noise occurs in the range of Qi to EiT.

However, when Ei is in the range of (｛EiB to EiT), forcibly outputting the next higher quantization level,
From the characteristics of the above-described quantization characteristics, the quantization noise always has a positive value. Similarly, if the next lower level is forcibly output, the quantization noise always becomes a negative value.

From this, the local decoded signal Y (= X +
When N) overflows, if the next lower quantization level is output, the quantization noise N is a negative value, so that the local decoded signal becomes Y = X + N ≦ X and overflow does not occur. On the other hand, when an underflow occurs, if the quantization level one level higher is output, the quantization noise is a positive value, so that the local decoded signal becomes Y = X + N ≧ X and no underflow occurs.

As described above, since the modulo operation is used for addition and subtraction in predictive coding, the next higher quantization level when the quantization level is the highest is the lowest quantization level. When the quantization level is the lowest, the next lower quantization level is the highest quantization level.

The quantization characteristics of the quantizer 11 can be determined for each signal of 64 transform coefficients. Since the amplitude of the inner low-frequency component of the converted signal is expected to increase statistically, a large number of levels are assigned, and a quantization characteristic covering the entire range of 11 bits is used. Since the high-frequency component is expected to have a statistically small amplitude and a low frequency of occurrence, the number of levels is allocated smaller than that of the low-frequency component, and the amplitude of the maximum quantization level of the quantization characteristic is small and the minimum quantization level is also small. Roughen.

Each of the quantization characteristics has such a characteristic that since the frequency distribution of the prediction error conversion signal is concentrated on 0, the portion having a small amplitude is finely quantized and the portion having a large amplitude is coarsely quantized. By using the above-described quantization characteristics, the average quantization noise can be statistically reduced.

When variable-length coding (entropy coding) is performed on each quantization level, efficient coding is possible even if the number of levels increases, if variable-length codes are used according to the frequency of occurrence of each level. Where the limit on the number of quantization levels is
This is necessary when performing efficient encoding when a lot of information is generated due to a scene change and information is generated too much.

Referring to FIG. 1 again, the quantization of aliasing will be described.

The quantizer 7 performs aliasing quantization. The quantizer 7 has an 11-bit dynamic range (−102).
4 to 1023), and the magnitude of the maximum quantization noise is Qma
It has a quantization characteristic of x. In this case, since the maximum quantization noise is Qmax, the limiter 3 converts the converted signal At into-
The amplitude is limited to the range of 1024 + Qmax to 1023-Qmax and output. Since the quantization noise added when the high-frequency transform difference signal is quantized by the quantizer 7 is at most Qmax, the local decoded signal Y has a value obtained by adding the quantization noise N to the transform signal At. The decoded signal always falls within the 11-bit dynamic range (−1024 to 1023). That is, the encoding process can be performed in the 11-bit dynamic range.

[0055]

As described above, according to the present invention, the dynamic range of the quantizer can be quantized using half the quantization characteristic as compared with the conventional one, and efficient coded transmission is performed. be able to.

Further, according to the present invention, in an orthogonal transform encoding apparatus for orthogonally transforming a prediction difference signal of a predictive encoding system, quantizing transform coefficients, and encoding and transmitting the same, the dynamic range of quantization characteristics is improved with a small number of bits. As a result, the transient response characteristics can be improved.

In addition, according to the present invention, when quantizing a signal obtained by predictively encoding a transform coefficient subjected to orthogonal transform, a difference signal of the transform coefficient is adaptively quantized with respect to a low-frequency component, and a high-frequency component is quantized. Since the amplitude is limited in advance by the magnitude of the maximum quantization noise, the dynamic range of the quantization characteristics is reduced to 11 bits, which is half the conventional 12 bits. Can be.
As a result, the number of bits for encoding the quantized output can be reduced, and the encoded transmission can be performed efficiently.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an example of an adaptive quantization scheme for orthogonal transform encoding according to the present invention.

FIG. 2 is a block diagram for specifically explaining the adaptive quantizer shown in FIG.

FIG. 3 is a block diagram showing a configuration of a conventional interframe predictive orthogonal transform coding scheme.

FIG. 4 is a block diagram showing a configuration of a conventional DPCM (prediction coding) system using adaptive quantization.

[Explanation of symbols]

 1,13 orthogonal transformer 2,14 separator 3 limiter 4,5,22 subtractor 6 adaptive quantizer 7,21 quantizer 8 synthesizer 9 code converter 10,24 adder 11 inverse orthogonal transformer 12 prediction Unit 23 Switching unit 25 Judgment circuit

Claims (2)

[Claims] 1. An orthogonal transform coding system for orthogonally transforming and encoding a prediction error signal, wherein first means for orthogonally transforming an input signal to generate a transformed signal, Predictive coding is performed by adaptive quantization in the dynamic range of the transformed signal, and amplitude limitation is performed on high-frequency components of the transformed signal in accordance with the magnitude of maximum quantization noise, and prediction is performed in the dynamic range of the transformed signal. And a second means for obtaining an error signal, quantizing the error signal, and coding the error signal. 2. An orthogonal transform coding system for orthogonally transforming and encoding a prediction error signal, comprising: first means for orthogonally transforming an input signal for each block and outputting a converted signal for each block; Second means for dividing the converted signal into two groups of a low band and a high band to obtain a low band converted signal and a high band converted signal and limiting the amplitude of the high band converted signal to obtain an amplitude-limited high band converted signal. Third means for subtracting a low-frequency conversion prediction signal from the low-frequency conversion signal by a modulo operation to obtain a low-frequency conversion difference signal, and a high-frequency conversion prediction signal from the amplitude-limited high-frequency conversion signal by a modulo operation. A fourth means for subtracting the high-frequency transform difference signal into a high-frequency transform differential signal, and outputting a temporary low-frequency quantized output by quantizing the low-frequency transform differential signal in accordance with a predetermined quantization characteristic and generating the quantized signal at the time of quantization. Quantization noise and A temporary converted local decoded signal obtained by adding the low-band converted signal is obtained, and it is determined whether the temporary converted local converted signal exceeds the dynamic range. Output as a quantized output, and an adaptive quantizing means for outputting, as a high-frequency quantized output, a quantization level one level higher than the provisional high-frequency quantized output when the dynamic range falls below the dynamic range, Quantizing means for quantizing the high-frequency transform difference signal according to the quantization characteristic and outputting a high-frequency quantized output signal; and a transform quantization output for each block obtained by combining the low-frequency quantized output and the high-frequency quantized output. And a code conversion means for converting the transformed quantized output into a channel code and sending out the transformed quantized output, and adding the transformed quantized output and the transform prediction signal by a modulo operation to obtain a transformed local decoded signal. Profit And the means of the sixth,
Seventh means for inversely orthogonally transforming the transformed local decoded signal for each block to obtain a local decoded signal, eighth means for obtaining a prediction signal for a next block from the local decoded signal, and A ninth step of orthogonally transforming to obtain a conversion prediction signal, dividing the conversion prediction signal for each block into two groups of a low band and a high band, and outputting the low band conversion prediction signal and the high band conversion prediction signal Means for orthogonal transform coding.