JP2003189309A - Re-coding method and apparatus for compressed moving picture - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide re-encoding technologies capable of effectively suppressing accumulation of rounding errors in a motion compensation processing process of error signals so as to prevent deterioration in image quality. <P>SOLUTION: A compressed moving picture stream of a prescribed encoding system is received and decoded. After performing motion compensation predictive processing including rounding processing to a finite length in the unit less than pixel accuracy to the decoded stream, the resulting decoded stream is encoded again. The decoded image signal is quantized again and an error predictive signal synthesis section 102 performs a motion compensation predictive process for a resulting re-quantized error signal E2(n-1). A rounding system decision section 101 at that time revises a rounding system for the rounding processing in each processing unit such as a frame/field or a macro block independent of the encoding processing when the input compressed moving picture stream is generated. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for re-encoding a compressed moving image, and more particularly to a compressed moving image re-encoding method and apparatus for reducing the code amount of an input compressed moving image stream for output.

[0002]

2. Description of the Related Art When transmitting and storing moving image signals in digital broadcasting systems and services, many moving image signals are compression-coded and transmitted and stored. In recent years, ISO / ICE IS13818-2 (M
PEG-2 VIDEO) has been standardized and is used in digital broadcasting systems and services.

Broadcast stations and homes are expected to have devices and applications for re-encoding a compressed moving picture stream compressed at a predetermined bit rate into a compressed moving picture stream having a different bit rate and transmitting or accumulating it. For example, there is a long-time recording function in a digital moving image signal recording device. The compressed moving image stream distributed from the broadcasting station to the home is encoded at a predetermined bit rate. When a viewer records a compressed moving image stream distributed in a limited storage capacity for the purpose of long-time recording, it is necessary to re-encode the compressed moving image stream at a bit rate lower than the bit rate at the time of distribution.

Hereinafter, a conventional re-encoding method and apparatus for performing such bit rate conversion will be described with reference to MPEG-2 VIDEO.
A case of a compressed moving image stream (hereinafter referred to as an MPEG-2 bit stream) that is compression-encoded according to the above will be described as an example.

As shown in FIG. 17, the basic re-encoding circuit has a configuration in which an MPEG-2 bit stream decoder and an encoder are connected in series. The decoder is provided with a variable length decoder 201, an inverse quantizer 202, and an inverse discrete cosine transformer 203, and obtains a prediction error signal R1 (n) from the input MPEG-2 bit stream. Here, the subscript “1” indicates that the signal is obtained by the decoder,
“N” represents the number in the time direction of the image signal in the image signal sequence (the same applies hereinafter). The adder 204 adds the prediction error signal R1 (n) and the prediction image signal MC [I1 (n-1), V] to generate a decoded image signal I1 (n). Therefore, the image signal I1 (n) is expressed by the following equation.

[0006]

[Equation 1]   I1 (n) = R1 (n) + MC [I1 (n-1), V] (1).

The decoded image signal I1 (n) is supplied to the encoder and is delayed by the memory 206, and the motion compensator 2
It is output to 05. The motion compensator 205 performs motion compensation MC using the already-decoded image signal I1 (n-1) and the motion vector field V supplied from the variable length decoder 201 to obtain the predicted image signal MC [I1 (n-1 ), V] and adder 20
Output to 4.

The encoder is the image signal I1 decoded by the decoder.
Re-encode (n). First, the subtractor 207 outputs the second predicted image signal MC [I2 (n-1), from the decoded image signal I1 (n).
V] is subtracted to obtain the prediction error signal R2 (n). Here, the subscript “2” represents a signal obtained by the encoder, and “n” represents the number in the time direction of the image signal in the image signal sequence (the same applies hereinafter). That is, the prediction error signal R2 (n) is expressed by the following equation.

[0009]

[Equation 2]   R2 (n) = I1 (n)-MC [I2 (n-1), V] (2).

The predicted image signal MC [I2 (n-1), V] is motion-compensated by the motion compensator 208 using the already decoded image signal I2 (n-1) and the motion vector field V. Desired.

The encoder includes a discrete cosine converter 20.
9, a quantizer 210, and a variable-length encoder 211 are provided to generate and output an MPEG-2 bitstream that is bit rate converted from the obtained prediction error signal R2 (n).

Further, the inverse quantizer 212 performs inverse quantization on the discrete cosine transform coefficient obtained by the quantizer 210, and the inverse discrete cosine transformer 213 performs inverse discrete cosine transform, thereby predicting An error signal R2 (n) + E2 (n) is obtained. However, since the quantization and the inverse quantization are not reversible transforms, the obtained prediction error signal includes the error signal E2 (n).

The adder 214 calculates the prediction image signal MC [I2 (n-1), V] and the prediction error signal obtained by the motion compensator 208.
The image signal I2 (n) is obtained by adding R2 (n) + E2 (n). Therefore, the image signal I2 (n) can be expressed as follows using the equation (2).

[0014]

[Equation 3]   I2 (n) = MC [I2 (n-1), V] + R2 (n) + E2 (n)         = I1 (n) + E2 (n) (3).

This image signal I2 (n) is delayed by the memory 215 and output to the motion compensator 208. Motion compensator 208
Performs the motion compensation MC using the already decoded image signal I2 (n-1) and the motion vector field V to obtain the predicted image signal MC
[I2 (n-1), V] is generated and output to the subtractor 207.

By the way, the motion compensation process is realized by selecting an already decoded image signal block having a high correlation with the image signal block to be encoded, and shifting the selected block using a motion vector. . Therefore, linearity is established for the motion compensation processing. A conventional technique in which the device configuration is simplified by utilizing the linearity of the motion compensation processing is disclosed in Japanese Patent Laid-Open No. 8-51631 and Japanese Patent No. 31.
66501, or US Pat. No. 5,623,312. These devices are provided with means for adding the result of motion compensation processing to the error signal generated by requantization to the prediction error signal.

Here, the compressed moving picture re-encoding device described in Japanese Patent Laid-Open No. 8-51631 will be described as an example. First, the process of simplification by applying the linearity of the motion compensation process will be described.

Since I2 (n) = I1 (n) + E2 (n) from the above equation (3), the predicted image signal MC [I2 (n-1), V] can be expressed by the following equation. .

[0019]

[Equation 4]   MC [I2 (n-1), V] = MC [(I1 (n-1) + E2 (n-1)), V] (4).

Here, the following expression holds from the linearity of the motion compensation processing.

[0021]

[Equation 5]   MC [(I1 (n-1) + E2 (n-1)), V] = MC [I1 (n-1), V] + MC [E2 (n-1), V]                                               ... (5).

For the prediction error signal R2 (n), equation (2)
By applying the equation (1), the equation (4), and the equation (5) to the above, the following equation is established.

[0023]

[Equation 6]   R2 (n) = R1 (n) + MC [I1 (n-1), V]-MC [I1 (n-1) + E2 (n-1), V]         = R1 (n) + MC [I1 (n-1), V]-MC [I1 (n-1), V]-MC [E2 (n-1), V]         = R1 (n)-MC [E2 (n-1), V]                                                 ... (6).

From this equation (6), the prediction error signal R2 (n) necessary for re-encoding is the motion compensation processing result (hereinafter, error) of the decoded prediction error signal R1 (n) to the error signal E2 (n). It is obtained by subtracting the prediction signal. Therefore,
As compared with the re-encoding device shown in FIG. 15, the number of memories required for re-encoding can be reduced and the device configuration can be simplified.

FIG. 18 is a circuit diagram showing an example (Japanese Patent Laid-Open No. 8-51631) of a re-encoding device simplified in this way. Variable length decoder 301, inverse quantizer 30
2 and the inverse discrete cosine converter 303 obtain the prediction error signal R1 (n) from the input MPEG-2 bit stream, and the discrete cosine converter 306, the quantizer 308 and the variable length encoder 309 predict the prediction error signal R2. The structure in which the output MPEG-2 bit stream is obtained from (n), and the prediction error signal R2 (n) + E2 (n) is further obtained by the inverse quantizer 310 and the inverse discrete cosine transformer 311 is the same as in FIG. Is.

In this simplified re-encoding device, the subtractor 304 calculates the error signal E2 from the prediction error signal R1 (n).
The prediction error signal R2 (n) is obtained by subtracting the error prediction signal MC [E2 (n-1), V] of (n-1). Further, the subtractor 307 subtracts the prediction error signal R2 (n) from the prediction error signal R2 (n) + E2 (n) output from the inverse discrete cosine converter 311 to obtain the error signal E2 (n). It is obtained, delayed in the memory 312, and output to the motion compensator 305. Motion compensator 305
Performs the motion compensation MC of the error signal using the error signal E2 (n) and the motion vector field V to obtain the error prediction signal MC [E
2 (n-1), V] is generated and output to the subtractor 304. With this configuration, motion compensation processing of the error signal generated by requantization can be basically performed.

[0027]

However, when the motion compensation of the error signal is performed and the rounding process to the finite length is performed in the calculation process, there is a problem that the image quality deteriorates. The cause of image quality deterioration in the conventional re-encoding device will be described below. Again, the compressed video stream is MP
It shall be compressed and encoded according to EG-2 VIDEO.
As is well known, in MPEG-2 VIDEO, half-pixel precision motion compensation processing is used in which the minimum unit of a motion vector is half the length of an adjacent pixel. When the motion vector is not an integer value in the motion compensation process, the pixel value of a pixel position that does not exist is obtained by interpolation.

As shown in FIG. 19, integer pixel position 40
If the pixel values of the pixel signals at 1, 402, 403, and 404 are La, Lb, Lc, and Ld, respectively, half-pixel positions 405, 406, 407, and 408 to be interpolated in MPEG-2 VIDEO.
The pixel values Ia, Ib, Ic, and Id of the pixel signal of
It is calculated by the equation 10.

[0029]

[Equation 7] Ia = La (7).

[0030]

(8) Ib = (La + Lb + 1) / 2 (8).

[0031]

## EQU9 ## Ic = (La + Lc + 1) / 2 (9).

[0032]

[Equation 10] Id = (La + Lb + Lc + Ld + 2) / 4 (10).

Here, '/' is an operation for rounding the result of division to an integer not exceeding that value. In addition, formula 7 to formula 10
The reason that 1 or 2 is added to the pixel value in is that the value after the decimal point is rounded off. When motion compensation processing with half-pixel precision is performed in this way, an error occurs due to rounding pixel values into integer values. Hereinafter, an error caused by rounding the pixel value and the error value of the error signal into an integer value is referred to as “rounding error”.

As described above, in the re-encoding device shown in FIG. 18, the device configuration is simplified by utilizing the linearity shown in equation 5. In this case, the left side of Equation 5 is MC [(I1 (n-1) +
E2 (n-1)), V] has a rounding error, and the right side has the first term MC [I1 (n-1), V] and the second term MC [E2 (n-1), V]. Rounding error occurs in each term of.

However, MC [(I1 (n-1) + E2 (n-1)), V]
Since the rounding error of is not equal to the sum of the rounding error of the first term MC [I1 (n-1), V] and the second term MC [E2 (n-1), V] on the right-hand side, Equation 5 is Generally does not hold. as a result,
In such a simplified configuration, a result different from that of the re-encoding device shown in FIG. 17 is generally obtained due to a rounding error when performing motion compensation processing.

If the rounding error does not accumulate, the effect of the image quality deterioration due to the rounding error is considered to be small in practical use.
It is important that the expected value of the rounding error is stochastically zero. According to the above equation 6, the rounding error is caused by MC [E2 (n-
1), V], which is the motion compensation process, so let us consider the expected rounding error in MC [E2 (n-1), V].

The integer pixel positions 401 and 40 shown in FIG.
When the error values of the error signals at 2, 403, and 404 are La, Lb, Lc, and Ld, respectively, the half pixel positions 405 and 40
6, the error values Ia, Ib of the error signal at 407, 408,
Ic and Id can be calculated by Equations 11 to 14, respectively.

[0038]

[Equation 11]   Ia = La (11).

[0039]

[Equation 12]   If (La + Lb> = 0), then Ib = (La + Lb + 1) / 2                    else Ib = (La + Lb) / 2 (12).

[0040]

[Equation 13]   if (La + Lc> = 0), then Ic = (La + Lc + 1) / 2                    else Ic = (La + Lc) / 2 (13).

[0041]

[Equation 14]   if (La + Lb + Lc + Ld> = 0), then Id = (La + Lb + Lc + Ld + 2) / 4                          else Id = (La + Lb + Lc + Ld + 1) / 4 (14).

Equations 7 to 1 in the re-encoding apparatus shown in FIG.
The difference from 0 is that the simplified re-encoding device shown in FIG. 18 performs motion compensation processing on error signals having positive and negative values. As a rounding method, positive / negative symmetrical rounding with 0 at the center is generally used, and therefore Equations 11 to 14 are obtained by positively and negatively symmetric Equations 7 to 10.

If the half-pixel positions for which the error signal is to be obtained are ¼, 405, 406, 407, and 408 shown in FIG. 19, respectively, the rounding error at the pixel position 405 is obviously 0.

The rounding error at the pixel position 406 is La + L
It differs depending on whether the sign of b is odd or even. No rounding error occurs when La + Lb is 0, so the probability of 0 is not considered. Therefore, the probability that La + Lb is positive is p, and the probability that it is negative is (1-p). Also, the odd-even probability of La + Lb is set to 1/2. The rounding error at the pixel position 406 is 0 when La + Lb is even, 1/2 when La + Lb is positive and odd, and La + Lb
Is negative and odd, it is -1/2, and its expected value is
0 x (p + 1-p) x 1/2 + 1/2 x p x 1/2 + (-1/2) x
(1-p) × 1/2 = (2p-1) / 4. The expected value of the rounding error at the pixel position 407 is (2p-1) / 4 as in the pixel position 406.
Becomes

The rounding error at the pixel position 408 is La + L
The difference is positive / negative of b + Lc + Ld and the remainder when La + Lb + Lc + Ld is divided by 4. Therefore, the probability that La + Lb + Lc + Ld is positive is q, and the probability that it is negative is (1-q). When La + Lb + Lc + Ld is positive, the rounding error when the remainder divided by 4 is 0, 1, 2, 3 is 0, -1/4, 1/2, 1/4, respectively. . At this time,
The probability that the remainder becomes 0 to 3 is equal. Also, La + L
When b + Lc + Ld is negative, the rounding error when the remainder of dividing the absolute value of La + Lb + Lc + Ld by 4 is 0, 1, 2, 3 is 0, 1/4, −, respectively. It becomes 1/2 and -1/4. At this time, the probability that the remainder becomes 0 to 3 is equal. From this, the expected value of the rounding error at the pixel position 408 is q × (0 × 1 /
4-1 / 4 × 1/4 + 1/2 × 1/4 × 1/4) + (1-q) × (0 ×
1/4 + 1/4 × 1 / 4−1 / 2 × 1 / 4−1 / 4 × 1/4) = (2q
-1) / 8.

Therefore, the expected value of the final rounding error is 0.
× 1/4 + (2p-1) / 4 × 1/4 + (2p-1) / 4 × 1/4 + (2q-
1) / 8 × 1/4 = (2p-1) / 8 + (2q-1) / 32. This is (2p-1) / 8 + (2q-
1) It means that a rounding error of / 32 occurs.

Here, if the positive and negative probabilities are equal probabilities, that is, if p = q = 1/2, the expected rounding error value is 0,
No rounding error occurs. However, here, in order to simplify the explanation, the odds, the remainder, and the probability of the pixel position of the error value are assumed to be equal probabilities.

However, in general, the positive / negative of the error value in the error signal differs depending on the image characteristics, the quantization characteristics, etc. and does not become equal probability. Fig. 20 shows three types of MPEG-2 VIDEO
4 shows a distribution of error values when bit streams 1, 2, and 3 are actually re-encoded.

Here, in MPEG-2 VIDEO, compression is performed using intraframe prediction or interframe prediction. Since a frame encoded using intra-frame prediction is not motion-compensated, a rounding error in motion compensation does not occur.
In MPEG-2 VIDEO, I picture corresponds to it. Further, among frames coded using inter-frame prediction, a frame that is not used as a reference image does not accumulate rounding errors. In MPEG-2 VIDEO, this corresponds to B picture.

Therefore, it is important to prevent accumulation of rounding error in a P picture used as a reference image and compressed using motion compensation prediction. Here, re-encoding was performed using an MPEG-2 VIDEO bitstream having the IPPPP ... structure. The interval between I pictures is 30 frames.

In the graph of FIG. 20, the vertical axis represents the average value of error values per frame in the luminance signal, and the horizontal axis represents the frame number. It can be seen from FIG. 20 that the positive / negative of the error value in the error signal generally differs depending on the image characteristics, the quantization characteristics, and the like, and is not equal probability. as a result,
Depending on the distribution of error values, rounding errors may accumulate in a positive or negative fixed direction, and the image quality may be deteriorated by performing the interpolation processing of Expressions 11 to 14.

Various methods for suppressing the accumulation of rounding errors have been proposed as a system for encoding and decoding compressed moving images.

For example, Japanese Laid-Open Patent Publication No. 11-317957 discloses a method of suppressing accumulation of rounding error by switching between a positive rounding (rounding up) method and a negative rounding (rounding down) method at the time of encoding. . Similarly, in Japanese Patent Laid-Open No. 11-69345, a method of avoiding accumulation of rounding errors by selecting a method in which the rounding direction is a positive infinity direction and a method in which a rounding direction is a negative infinity direction with approximately the same probability. Is disclosed. In other words, the constant direction as in the above formulas 7 to 10 (here, the positive direction)
If only the rounding method is used, the rounding error at the time of encoding is always accumulated in a fixed direction, and the image quality deteriorates. In order to avoid the accumulation of this rounding error, the rounding method is switched at the time of encoding.

However, what is disclosed in these publications is a method for suppressing the accumulation of rounding error as an encoding and decoding system. Therefore, in the techniques disclosed in these publications, the rounding method used at the time of encoding must be given at the time of decoding in order to match the reference image at the time of encoding and the reference image at the time of decoding. for that reason,
Rounding information must be included in the compressed video stream. That is, it is necessary to notify the decoding device of the rounding system switching information in the encoding device, so that the decoding device also performs the rounding process in the same rounding system. In other words, the method described in the above publication is intended to suppress the rounding error that occurs due to the calculation accuracy of the motion compensation processing defined by the encoding method.

On the other hand, the re-encoding technique to which the present invention belongs is different regardless of the rounding method at the time of encoding, by performing some conversion (such as reduction of the code amount) on the input compressed moving image stream. It is output as a compressed moving image stream. Therefore, the decoding device inputs the converted stream and generates an image signal. In the re-encoding technology to which the present invention belongs, a rounding error that occurs when a rounding process to a finite length is performed in the calculation process of the motion compensation process of an error signal, in other words, a rounding error that occurs due to the calculation accuracy of the motion compensation process is canceled. Is intended. In other words, it is intended to suppress a rounding error that occurs in the calculation accuracy (reduced accuracy) of the motion compensation processing that is not specified by the encoding method.

Therefore, even if the encoding methods described in Japanese Patent Laid-Open Nos. 11-317957 and 11-69345 are used, a simplified re-encoding device as shown in FIG. 18 is used. If the compressed moving image stream data is re-encoded, a rounding error of the error signal E2 (n) will occur, and the image quality will deteriorate. That is, the rounding error that is the subject of the present invention is different from the rounding error described in the above publications, and the methods of these publications cannot solve the problems of the present invention.

An object of the present invention is to prevent re-encoding which has a simple structure and which can prevent the deterioration of image quality by suppressing the accumulation of rounding error when rounding processing to a finite length in the motion compensation processing process of an error signal. A method and apparatus are provided.

[0058]

A re-encoding method according to the present invention is a motion-compensated prediction in units of less than pixel precision including rounding processing to a finite length by inputting and decoding a compressed moving image stream of a predetermined encoding method. In the method of re-encoding after performing processing, an image signal obtained by decoding at least a part of the compressed moving image stream is requantized, and a signal before requantization and a signal after requantization In the process of calculating a requantization error signal from the above and performing the motion compensation prediction process on the requantization error signal, a process that does not depend on the encoding process when the input compressed moving image stream is generated. The rounding method of the rounding process is changed for each unit.

The predetermined coding method is preferably a coding method in which conversion processing into frequency space is combined.
The image signal obtained by decoding at least a part of the compressed moving image stream is a frequency coefficient obtained by decoding at least a part of the compressed moving image stream.

The processing unit whose rounding method is changed is
It is preferably at least one frame / field or at least one small block in a frame / field, and may be at least one pixel.

According to an embodiment of the present invention, a plurality of rounding schemes are preset for different frame / field coding types, and one of the plurality of rounding schemes is selected for each processing unit. It is characterized by Further, at least one of frame / field coding types is characterized in that the rounding method is fixed.

It is preferable that the motion compensation prediction process has a plurality of interpolation processes, and at least one rounding method is set corresponding to each of the plurality of interpolation processes.

According to another embodiment of the present invention, a compressed moving image stream of a predetermined encoding method is input and decoded, and motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length is executed. Then, in the method of re-encoding, the image signal obtained by decoding at least a part of the compressed moving image stream is requantized, and requantized from the signal before requantization and the signal after requantization. In the process of calculating a quantization error signal, generating statistical information of the requantization error signal, and performing the motion compensation prediction process on the requantization error signal, the input compressed moving image stream is generated. The rounding method of the rounding process is changed based on the statistical information for each processing unit that does not depend on the encoding process.

The statistical information is generated based on an error value of the requantization error signal for each pixel.
Further, it is preferable that the statistical information is generated based on an average value of error values of the requantization error signal in a plurality of pixels. The statistical information is preferably generated based on an average value of error values of the requantization error signal in a small block in a frame / field.

According to another aspect of the present invention, after inputting and decoding a compressed moving image stream of a predetermined encoding method and executing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length. , A re-encoding device for re-encoding a compressed image includes re-quantizing means for re-quantizing an image signal obtained by decoding at least a part of the compressed moving image stream, and a signal before re-quantizing. Error calculation means for calculating a requantized error signal from the requantized signal, error prediction signal synthesizing means for executing the motion compensation prediction process on the requantized error signal, and the motion compensated prediction In the process of processing, there is provided rounding method determining means for changing the rounding method of the rounding processing for each processing unit that does not depend on the encoding processing when the input compressed moving image stream is generated.

According to one embodiment of the present invention, a compressed moving image stream of an encoding method in which a conversion process into a frequency space is combined is input, decoded, and rounded to a finite length. After performing the motion compensation prediction process,
A re-encoding device for re-encoding a compressed image includes converting means for converting a first image signal obtained by decoding at least a part of the compressed moving image stream into a frequency space to generate frequency coefficients, and the frequency. Requantizing means for requantizing coefficients, dequantizing means for dequantizing the requantized signal, inverse transforming means for transforming the dequantized signal into a second image signal, and Error calculating means for calculating a requantized error signal from the first image signal and the second image signal; error predicted signal synthesizing means for executing the motion compensation prediction process on the requantized error signal; In the process of the motion compensation prediction process, a rounding method determining unit that changes the rounding method of the rounding process is provided for each processing unit that does not depend on the encoding process when the input compressed moving image stream is generated. Characteristic To.

According to still another aspect of the present invention, a compressed moving image stream of a predetermined encoding method is input and decoded,
A re-encoding device for a compressed image that performs re-encoding after performing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length is obtained by decoding at least a part of the compressed moving image stream. Requantization means for requantizing the image signal, error calculation means for calculating a requantization error signal from the signal before requantization and the signal after requantization, and the requantization error signal On the other hand, an error prediction signal synthesizing unit that executes the motion compensation prediction process to generate an error value by the rounding process, a storage unit that accumulates statistical information of the error value by the rounding process, and a process of the motion compensation prediction process. In the above, there is provided a rounding method determining unit that changes the rounding method of the rounding processing based on the statistical information for each processing unit that does not depend on the encoding processing when the input compressed moving image stream is generated. It is characterized in.

As described above, according to the present invention, since the rounding method of motion compensation is changed for each predetermined processing unit such as a frame / field of image data or a small block, accumulation of rounding error is suppressed. Even if the distribution of the error values is biased, the expected value of the rounding error can be brought close to 0 on the average in the time direction, and the occurrence of the rounding error in the fixed direction can be suppressed.

Furthermore, by using the statistical information of the rounding error, the rounding method can be determined according to the actual occurrence of the error, and the accumulation of the rounding error can be suppressed more effectively.

[0070]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below, but the basic configuration of a re-encoding device is the same as that shown in FIG. Therefore, the description of the variable length decoder and the inverse quantizer is omitted, and the motion compensator according to the present invention will be mainly described in detail.

(First Embodiment) FIG. 1 is a block diagram showing the configuration of a motion compensation section in a first embodiment of a re-encoding device according to the present invention. Motion compensation unit 10 according to the present invention
Inputs input bitstream coding information, rounding scheme change information, and a reference error signal, and outputs an error prediction signal. The motion compensation unit 10 includes a rounding method determination unit 101 and an error prediction signal synthesis unit 102.

The rounding scheme determination unit 101 determines the next rounding scheme based on the current rounding scheme, the rounding scheme change information and the coding information, and outputs the rounding scheme information to the error prediction signal synthesis unit 102. Here, the rounding method change information is information about a period and a condition for changing the rounding method, as will be described later. For example, the rounding method change information that is set in advance in the error prediction signal synthesis unit 102 (see FIG. 2). ) Index or rounding formula information. The rounding method determination unit 101 sets the rounding method according to the rounding method change information, and the error prediction signal synthesizing unit 102 executes the motion compensation process according to the set rounding method.

The error prediction signal synthesizing unit 102 uses the error signal E2 (n-1) to be referred to, the coding information of the input stream such as the motion vector, and the rounding method information input from the rounding determining unit 101 to calculate the error. The prediction signal MC [E2 (n-1), V] is obtained and output to the subtractor 20.

Here, it is assumed that there are M types of interpolation processing depending on the pixel position of the error signal to be interpolated, the encoding mode, etc., and a plurality of rounding methods Z are set in each interpolation processing. For example, in the interpolation processing 1, there are K (1) rounding methods Z _{11 to} Z _{1K (1)} , and similarly, the interpolation processing M
Then, it is assumed that there are K (M) rounding methods Z _{M1 to} Z _{MK (M)} . For example, a plurality of rounding methods for each interpolation process are sequentially or randomly selected from the initial setting method.

FIG. 2 shows a rounding method used in the first embodiment.
It is a figure which shows the list of a formula. In each interpolation process, input
Condition C using encoding information included in the input stream
There may be a rounding method for each. for example
For example, K (1) condition C₁₁~ C _{1K (1)}Against each
Rounding method Z₁₁~ Z_{1K (1)}Exists, and similarly, the interpolation process M
Then, there are K (M) conditions C_M1~ C_{MK (M)}Against that
Re rounding method Z_M1~ Z_{MK (M)}May exist.

FIG. 3 shows the rounding determination unit 1 in this embodiment.
It is a flow chart which shows operation of 01. It is assumed that an initial value is given to the rounding method of the interpolation processing.
Further, in the rounding determination unit 101, the encoded information is used in the determination when changing the rounding method and the determination as to whether it is a reference image.

In step S1, the rounding method determining unit 101
It is determined whether to change the rounding method using the rounding method change information and the encoding information supplied to the. If the rounding method is changed, the process proceeds to step S2, and if not, the process proceeds to step S3. The change of the rounding method is performed for each certain processing unit defined by the rounding method change information. As will be described later, it is a frame / field unit or a block unit in a frame / field.

In step S2, it is judged using the coding information whether or not the error signal obtained by the interpolation processing is used as the reference image. If it is used as a reference image, the process proceeds to step S4, and if not, the process proceeds to step S5.

In step S3, the same rounding method as the previous interpolation process is used without changing the rounding method. That is, in this step S3, the rounding method determination process is terminated without doing anything.

In step S4, the rounding method is changed.
Since there are multiple types of interpolation processing depending on the pixel position of the error signal to be interpolated, the encoding mode, etc., the rounding method in each interpolation processing is changed.

In step S5, the rounding method of the error signal not used as the reference image is determined. Since the rounding error of the error signal that is not used as the reference image is not accumulated,
When the rounding method is used with the initial value as it is, nothing is done in this step, and the rounding method determination processing is ended.

In step S6, the rounding method determined in step S4 is output to the error prediction signal synthesizing unit 102 as rounding method information, and the rounding method determination processing ends.

Next, the operation of this embodiment will be described in detail using a specific example.

(1) First Example of First Embodiment In the first example, the input compressed moving image stream is MPEG.
-2 It shall be a VIDEO stream and the error signal shall be rounded to an integer value. In MPEG-2 VIDEO, it is only the P picture that uses forward prediction that interpolates the error signal and becomes the reference image signal. In the present embodiment, it is assumed that the rounding method is changed only for the P picture, and the rounding change information is supplied for each frame during the period.

The changing period and conditions of the rounding method may be changed for each luminance signal or each color signal, for each block size of an image signal, for each certain number of frames, or for each certain coding type. Various methods can be considered, such as changing for each prediction method. For example,
In MPEG-2 VIDEO, the prediction methods include forward prediction, backward prediction, and bidirectional prediction, and these prediction methods include field prediction and frame prediction.

The rounding method can be set for each of a plurality of interpolation processes and a plurality of conditions. For example, in MPEG-2 VIDEO, there are five types of interpolation processing: horizontal half-pixel rounding, vertical half-pixel rounding, horizontal / vertical half-pixel rounding, dual prime rounding, and bidirectional predictive rounding.

As the condition, for example, a condition based on a signal of a luminance signal or each color difference signal, a condition using encoded information such as frame motion compensation or field motion compensation, and a combination thereof can be considered. However, here, for simplification, the rounding method is not set according to the conditions, and the following rounding methods are set for the above-mentioned five types of interpolation processing.

First, as the initial values of the rounding method for P pictures and B pictures, the above equations 12 to 14 and the following equation 1 are used.
A rounding method of 5 is given.

[0089]

[Equation 15]   if (Le + Lf> = 0), then Ie = (Le + Lf + 1) / 2,                     else Ie = (Le + Lf) / 2 (15).

Expression 15 is a rounding method used in dual prime prediction of P pictures and bidirectional prediction of B pictures. In dual prime prediction and bidirectional prediction, the two pixel signals Le and Lf are averaged.

The above equations 12 to 15 are put together, and the initial values of the rounding method for the above five kinds of interpolation processing will be described as follows.

Rounding of half horizontal pixel: Rounding method Z ₁₁ when La + Lb> = 0: Rounding method Z _{11 when} Ib = (La + Lb + 1) / 2, La + Lb <0 ₁₂ : Ib = (La + Lb) / 2.

Rounding of vertical half pixels: Rounding method Z ₂₁ if La + Lc> = 0: Rounding method Z _{21 if} Ic = (La + Lc + 1) / 2, La + Lc <0 ₂₂ : Ic = (La + Lc) / 2.

Rounding of horizontal / vertical half pixels: When La + Lb + Lc + Ld> = 0, rounding method Z ₃₁ : Id = (La + Lb +
Lc + Ld + 2) / 4, La + Lb + Lc + Ld <0, rounding method Z ₃₂ : Id = (La + Lb +
Lc + Ld + 1) / 4.

Rounding of dual prime / bidirectional prediction: Rounding method Z ₄₁ when Le + Lf> = 0: Ie = (Le + Lf + 1) / 2, Le + Lf <0 Rounding method Z ₄₂ : Ie = (Le + Lf) / 2.

FIG. 4 is a flow chart showing the operation of the rounding method determining section in the first embodiment according to the present invention. First, the initial value is given to the rounding method of each interpolation process.

In step S11, it is determined using the coding information whether the position of the error signal obtained by the interpolation processing is the head of the frame. If it is the beginning of the frame, step S1
2; otherwise, to step S13.

In step S12, it is determined using the coding information whether the error signal obtained by the interpolation processing is a P picture. If the error signal is a P picture, step S
14, otherwise, to step S15.

In step S13, the rounding method is not changed. In this case, the rounding method similar to the previous processing is used to end the rounding method determination processing.

In step S14, the rounding method for four types of interpolation processing of horizontal half pixel, vertical half pixel, horizontal vertical half pixel, and dual prime is changed. In the first embodiment,
For each P picture frame, a rounding method is set according to whether the sum of error values inside the interpolation process is positive or negative.

Therefore, when the number of P pictures is an even number, the rounding method is the same as the initial value. That is, in the case of rounding the horizontal half pixel, when La + Lb is 0 or more, the rounding method Z
₁₁ : Ib = (La + Lb + 1) / 2 is set, and when La + Lb is negative, a rounding method Z ₁₂ : Ib = (La + Lb) / 2 is set, and in the case of rounding half a vertical pixel, Rounding method when La + Lc is 0 or more Z ₂₁ : Ic = (La + Lc
Rounding method when +1) / 2 is negative La + Lc Z ₂₂ : Ic = (La + Lc) / 2
Is set, and if horizontal / vertical half-pixel rounding is performed, La + L
Rounding method when b + Lc + Ld is 0 or more Z ₃₁ : Id = (La + Lb + Lc + Ld
Rounding method when +2) / 4 and La + Lb + Lc + Ld are negative Z ₃₂ : Id = (La
+ Lb + Lc + Ld + 1) / 4 is set, and in case of dual prime rounding, when Le + Lf is 0 or more, rounding method Z ₄₁ : Ie = (Le +
Rounding method when Lf + 1) / 2 is negative Le + Lf Z ₄₂ : Ie = (Le + Lf)
/ 2 is set.

When the number of P pictures is odd, the rounding method is changed as follows. That is, in the case of rounding half a horizontal pixel, the rounding method Z ₁₂ : Ib = (La + Lb) / 2 when La + Lb is 0 or more, and the rounding method Z ₁₁ : Ib = when La + Lb is negative. When (La + Lb + 1) / 2 is set and the rounding of the vertical half pixel is performed, when La + Lc is 0 or more, the rounding method Z ₂₂ : Ic = (La +
Rounding method when Lc) / 2 is La + Lc is negative Z ₂₁ : Ic = (La + Lc + 1)
When / 2 is set and the horizontal / vertical half pixel is rounded, La
Rounding method when + Lb + Lc + Ld is 0 or more Z ₃₂ : Id = (La + Lb + Lc +
Rounding method when Ld + 1) / 4 and La + Lb + Lc + Ld are negative Z ₃₁ : Id =
(La + Lb + Lc + Ld + 2) / 4 is set, and in the case of dual prime rounding, when Le + Lf is 0 or more, the rounding method Z ₄₂ : Ie =
Rounding method when (Le + Lf) / 2 is negative Le + Lf Z ₄₁ : Ie = (Le + L
f + 1) / 2 is set.

In step S15, the B picture rounding method is determined. In this embodiment, however, since the B picture rounding method uses an initial value, nothing is done in this step, and the rounding method determination processing ends. .

In step S16, the rounding method determined in step S14 is output to the error prediction signal synthesizing unit 102 as rounding method information, and the rounding method determination processing ends.

(2) Second Example of First Embodiment The second example is different in the rounding method changing steps (S14 and S15 in FIG. 4) in the first example. The second embodiment is different from the first embodiment in that one rounding method is set irrespective of the condition inside the complementary processing, that is, whether the sum of error values is positive or negative.

FIG. 5 is a flow chart showing the operation of the rounding method determining unit in the second embodiment according to the present invention. As described above, first, the initial value is given to the rounding method of each interpolation process. Sequential steps S21 to S23
Is the same as steps S11 to S13 in FIG.

In step S24, as in the first embodiment, the rounding method for the four types of interpolation processing of horizontal half pixel, vertical half pixel, horizontal vertical half pixel, and dual prime is changed. However, unlike the first embodiment, the rounding method is set for each P picture frame, regardless of whether the sum of error values in the interpolation process is positive or negative.

Specifically, when the P picture is an even number, the rounding method is the same as the initial value. That is, the rounding method Z ₁₁ : Ib = (La + Lb + 1) / 2 is set in the case of horizontal half-pixel rounding, and the rounding method Z ₂₁ : Ic = (is set in the case of vertical half-pixel rounding. When La + Lc + 1) / 2 is set and the horizontal / vertical half pixel is rounded, the rounding method Z ₃₁ : Id = (La + Lb + Lc + Ld + 2) / 4
Is set, and in the case of dual prime rounding, the rounding method Z ₄₁ : Ie = (Le + Lf + 1) / 2 is set.

When the P picture is an odd number, the rounding method is changed as follows. That is, the rounding method Z ₁₂ : Ib = (La + Lb) / 2 is set in the case of horizontal half pixel rounding, and the rounding method Z ₂₂ : Ic is set in the case of vertical half pixel rounding.
= (La + Lc) / 2 is set and horizontal / vertical half-pixel rounding is used, the rounding method Z ₃₂ : Id = (La + Lb + Lc + Ld + 1) / 4 is set and dual In the case of prime rounding, the rounding method Z ₄₂ : Ie = (Le + Lf) / 2 is set.

In step S25, the B picture rounding method is determined. However, since the B picture rounding method uses an initial value in this embodiment, nothing is done in this step, and the rounding method determination processing ends. . However, as in S24, one rounding method is used regardless of the internal conditions of the interpolation processing.

In step S26, the rounding method determined in step S24 is output to the error prediction signal synthesizing unit 102 as rounding method information, and the rounding method determining process is terminated.

In the second embodiment, since one rounding method is used inside each interpolation process, the conditional branch inside the interpolation process can be omitted. Therefore, the device scale can be reduced as compared with the first embodiment. Further, in the case of a program or the like, determination of conditional branching inside the interpolation processing can be omitted, so high-speed processing can be realized.

(Second Embodiment) The second embodiment of the present invention has the same structure as the first embodiment, but the first embodiment
The operation is different from that of the embodiment in that the rounding method is changed for each frame coding type in the rounding method determination unit 101.

FIG. 6 is a flow chart showing the operation of the rounding method determination unit in the second embodiment according to the present invention.
Here, in the rounding scheme determination unit, the coding information is used in the rounding scheme change determination and the coding type determination. Also, it is assumed that there are W types of frame encoding types, the rounding method is changed for K types of encoding types 1 to K, and the remaining encoding types K + 1 to W are set to initial values. The set rounding method shall be used.

In step S31, the rounding method determining unit 10
It is determined whether or not to change the rounding method by using the rounding method change information and the encoding information supplied to 1. The change of the rounding method is performed for each certain processing unit determined by the rounding method change information. If the rounding method is changed, the process proceeds to step S32, and if not, the process proceeds to step S33. In step S33, the rounding method is not changed. In this case, the rounding method similar to the previous processing is used to end the rounding method determination processing.

In step S32, it is judged using the coding information whether or not the error signal obtained by the interpolation processing is coding type 1. If it is the coding type 1, the process proceeds to step S36, and if not, the process proceeds to the coding type 2 conditional branching process. Similarly, each coding type is sequentially determined until the conditional branching process of the coding type K (step S34). If the coding type is K in step S34, the process proceeds to step S37. If not, the process proceeds to conditional branching processing (step S35) to determine whether the coding types are K + 1 to W.

In step S36, the coding type 1 rounding method is changed. Here, it is assumed that there are M (1) types of interpolation processing. Then, it proceeds to step S39. Similarly, if it is determined in each conditional branching step that the encoding type is the corresponding encoding type i, the rounding method of the encoding type i is changed. Here, it is assumed that there are M (i) types of interpolation processing.

In step S37, the coding type K rounding method is changed. Here, it is assumed that there are M (K) types of interpolation processing.

In step S38, the coding type K + 1
The rounding method of ~ W is determined, but in these coding types, since the rounding method uses the initial value, nothing is done in this step.

As described above, in the first embodiment,
As the rounding method, the same rounding method is set for the reference image signal, and the rounding method set as an initial value is set for the signal that is not the reference image signal.

On the other hand, in the second embodiment, it is possible to decide whether to set the rounding method or use the rounding method of the initial value for each coding type. As a result, it is possible to set the rounding method according to the characteristics of the coding type, and it is possible to more effectively suppress the image quality deterioration during re-coding.

(1) Example of Second Embodiment The example of the second embodiment is the same as the first example of the first embodiment and the step of changing the B picture rounding method (S15 in FIG. 4).
Behave differently. In the first example (FIG. 4) of the first embodiment, the rounding method is changed for each P-frame, whereas in the present example, not only P-pictures but also B-pictures are changed frame by frame. The rounding method is changed.

FIG. 7 is a flow chart showing the operation of the rounding method determination unit in an example of the second embodiment according to the present invention. The initial values of the rounding method for P pictures and B pictures are given in the same manner as in the first example of the first embodiment. Subsequent steps S41 to S44 are the same as steps S11 to S14 in FIG.

In step S45, the rounding method in each interpolation process of the B picture is changed. The rounding method is changed for four types of interpolation processing of horizontal half pixel, vertical half pixel, horizontal vertical half pixel, and bidirectional prediction. In this embodiment, the rounding method is set according to whether the sum of error values inside the interpolation process is positive or negative.

Therefore, when changing the rounding method for each B picture frame, if the number of B picture frames is even, the rounding method is the same as the initial value. That is, in the case of rounding of half horizontal pixels, the rounding method Z ₁₁ : Ib = (La + Lb + 1) / 2 when La + Lb is 0 or more, and the rounding method Z ₁₂ : when La + Lb is negative. When Ib = (La + Lb) / 2 is set and the vertical half pixel is rounded, when La + Lc is 0 or more, the rounding method Z ₂₁ : I
c = (La + Lc + 1) / 2 is a rounding method when La + Lc is negative Z ₂₂ : Ic =
When (La + Lc) / 2 is set and the horizontal / vertical half-pixel rounding is performed, when La + Lb + Lc + Ld is 0 or more, the rounding method Z ₃₁ : Id = (L
Rounding method Z when a + Lb + Lc + Ld + 2) / 4 is negative and La + Lb + Lc + Ld is negative
₃₂ : Id = (La + Lb + Lc + Ld + 1) / 4 is set, and in the case of bidirectional prediction rounding, when Le + Lf is 0 or more, the rounding method Z ₄₁ : Ie =
Rounding method when (Le + Lf + 1) / 2 is negative Le + Lf Z ₄₂ : Ie = (Le
+ Lf) / 2 is set.

When the number of B pictures is odd, the rounding method is changed as follows. That is, in the case of rounding half a horizontal pixel, the rounding method Z ₁₂ : Ib = (La + Lb) / 2 when La + Lb is 0 or more, and the rounding method Z ₁₁ : Ib = when La + Lb is negative. When (La + Lb + 1) / 2 is set and the rounding of the vertical half pixel is performed, when La + Lc is 0 or more, the rounding method Z ₂₂ : Ic = (La +
Rounding method when Lc) / 2 is La + Lc is negative Z ₂₁ : Ic = (La + Lc + 1)
When / 2 is set and the horizontal / vertical half pixel is rounded, La
Rounding method when + Lb + Lc + Ld is 0 or more Z ₃₂ : Id = (La + Lb + Lc +
Rounding method when Ld + 1) / 4 and La + Lb + Lc + Ld are negative Z ₃₁ : Id =
(La + Lb + Lc + Ld + 2) / 4 is set, and in the case of rounding for bidirectional prediction, when Le + Lf is 0 or more, the rounding method Z ₄₂ : Ie = (Le + L
Rounding method when f) / 2 is negative Le + Lf Z ₄₁ : Ie = (Le + Lf + 1) /
2 is set.

In step S46, the rounding method determined in step S44 or S45 is output to the error prediction signal synthesizing unit 102 as rounding method information, and the rounding method determination processing ends.

As described above, if the positive / negative of the error value for motion compensation occurs at 1/2, the expected value of the rounding error of the error value is 0, and the rounding error is not accumulated. However, if the distribution of the error values is positively biased (for example, stream 1 shown in FIG. 18), the motion compensation processing of the error signal causes a rounding error in the positive direction.

In the first and second embodiments, since the rounding method of motion compensation is changed for each frame of P / B pictures, accumulation of such rounding error can be suppressed.
Even if the distribution of error values is biased, it is possible to suppress the occurrence of rounding error in a certain direction.

For example, if the two rounding methods of rounding and truncating are switched for each P picture for the distribution as shown in FIG. 8A, the rounding error in each P picture is as shown in FIG. 8B. Although it occurs, the expected value of the rounding error is close to 0 in the average in the time direction. As a result, it becomes possible to suppress the accumulation of rounding error in a certain direction.

FIG. 9 shows the peak signal to color difference signal when the MPEG-2 VIDEO bit stream is actually re-encoded.
It is a graph which shows noise ratio (PSNR). Here, the vertical axis represents PSNR and the horizontal axis represents the frame number. Further, the conventional method represents the case where the rounding method of Expression 11 to Expression 14 is used, and Example 1 and Example 2 are the methods of the first example and the second example of the first embodiment 1. Indicates that it has been used. The input MPEG-2 VIDEO bitstream is IP
It has the structure of PPP ... and the interval between I pictures is 60 frames.

As is apparent from FIG. 9, in the conventional method, the image quality is deteriorated due to the error accumulation, whereas when the first embodiment of the present invention is used, the accumulation of the error accumulation is suppressed. Therefore, it can be seen that this has the effect of suppressing image quality deterioration.

(Third Embodiment) FIG. 10 is a block diagram showing the structure of a motion compensator in a third embodiment of the re-encoding device according to the present invention. The motion compensation unit 10 according to the third embodiment includes a rounding method determination unit 110, an error prediction signal synthesis unit 111, and a statistical information storage unit 112.

The rounding method determining section 110 determines the next rounding method using the rounding method change information, the input bit stream coding information, the current rounding method, and the rounding error value input from the statistical information accumulating section 112. , Rounding method information is supplied to the error prediction signal synthesis unit 111.

The error prediction signal synthesizing unit 111 obtains and outputs the error prediction signal using the error signal to be referred to, the coding information of the input stream such as the motion vector, and the rounding method information supplied from the rounding determining unit 110. . Further, the rounding error value generated in the interpolation process and its position information are supplied to the statistical information storage unit 112.

The statistical information accumulator 112 adds the rounding error value supplied from the error prediction signal synthesizer 111 to the already accumulated rounding error value, and supplies the addition result to the rounding method determiner 110. If the error signal of the pixel position to be accumulated is intra-frame prediction, the rounding error value of the accumulated pixel position is initialized.

FIG. 11 is a flow chart showing the operation of the rounding method determination unit in the third embodiment according to the present invention. Steps S51 to S53, S55 and S56 are
This is the same as steps S11 to S13, S15 and S16 in FIG. The operation is different in the rounding method changing step S54.

In step S54, the rounding method is changed for each interpolation process. Here, it is assumed that there are M types of interpolation processing. Further, in each interpolation process, N kinds of rounding error value conditions 1 determined using the rounding error addition value and the coding information input from the rounding error value accumulating unit 112.
~ The rounding method is changed for each N, and the process proceeds to step S56.

In the third embodiment, the rounding method is determined in consideration of the actually occurring rounding error. This is because the expected value of the rounding error may be a value apart from 0 depending on the distribution of the rounding error value. Here, the actual rounding error value is accumulated, and the rounding method is adjusted so that the expected rounding error value approaches 0 in the next rounding process. As a result, compared to the configuration of the first embodiment, the expected value of the rounding error can be set to a value close to 0, and the accumulation of the rounding error can be suppressed more effectively.

(1) Example of Third Embodiment First, the statistical information storage unit 112 will be described. For example, a rounding method for an error signal at a certain position Z ₁₁ : Ib = (La + Lb + 1)
Suppose you ask for it at / 2. If La = 1 and Lb = 6, then Ib = 4. If there is no rounding, Ib = 3.5, so a rounding error of +0.5 occurs in this error signal. In this case, the error prediction signal synthesis unit 111 outputs the rounding error value +0.5 generated in the interpolation process and the pixel position thereof to the statistical information storage unit 112.

Since a rounding error occurs for each minimum unit of error, it is sufficient to add the rounding error value for each minimum unit to accumulate the cumulative sum of rounding error values. However, in this case, a memory for storing the rounding error value for the number of pixels is required. Therefore, for example, as shown in FIG. 13, it is possible to reduce the memory capacity for storing the rounding error value by accumulating as an average value for each certain block size or a cumulative sum of the average values of the luminance signal and the color difference signals. is there.

In the present embodiment, the rounding method is changed for each macroblock, and the condition of the rounding error value is that the average value of the cumulative sum of rounding error values of the pixels referred to at the time of interpolation of the macroblock in the error prediction signal synthesis unit 111 is The case is divided into positive and negative. A specific example is shown below.

FIG. 12 is a flow chart showing the operation of the rounding method determination unit in an example of the third embodiment according to the present invention. First, an initial value is given to the rounding method of interpolation processing. As the rounding method for P pictures and B pictures, the rounding methods of Expressions 12 to 15 are used.

In step S61, it is determined using the coding information whether or not the position of the error signal obtained by the interpolation processing is at the head of the macroblock. If it is the head of the macroblock, go to step S62, and if not, go to step S6.
Move to 3.

In step S62, it is determined using the coding information whether or not the error signal obtained by the interpolation processing is a P picture. If the error signal is a P picture, step S
If not, the process proceeds to step S65.

In step S63, the rounding method is not changed. In this case, the rounding method similar to the previous processing is used to end the rounding method determination processing.

In step S64, the rounding method for four types of interpolation processing of horizontal half pixel, vertical half pixel, horizontal vertical half pixel, and dual prime is changed. The rounding method is determined depending on whether the average of rounding error value cumulative sums in a macroblock is positive or negative and the sum of error values is positive or negative.

In practice, when the cumulative sum average value of the rounding error is 0 or more, the rounding method that makes the rounding error negative is applied. As a result, in the case of horizontal half-pixel rounding, when La + Lb> = 0 and when La + Lb <0, the rounding method Z ₁₂ : Ib = (La +
Lb) / 2 is set. If the cumulative average of the rounding errors is negative, the rounding method that makes the rounding error positive is applied. As a result, in the case of horizontal half-pixel rounding, when La + Lb> = 0 and when La + Lb <0, the rounding method Z ₁₁ : Ib = (La + Lb +
1) / 2 is set. In this way, the rounding method is the same regardless of whether La + Lb is positive or negative.

Similarly, in the case of rounding of half a vertical pixel, the rounding method Z ₂₂ : Ic = (La + L) when the cumulative average value of rounding errors is 0 or more.
c) / 2 is the rounding method Z when the cumulative average of rounding errors is negative
₂₁ : Ic = (La + Lc + 1) / 2 is set. When rounding horizontal / vertical half pixels, the rounding error cumulative average value is 0 or more, and the rounding method Z ₃₂ : Id = (La + Lb + Lc + Ld + 1) / 4 is the rounding error cumulative average value. If N is negative, rounding method Z ₃₁ : Id = (La + Lb + Lc + Ld + 2) / 4
Is set. In the case of dual prime rounding, if the cumulative average value of rounding error is 0 or more, rounding method Z ₄₂ : Ie = (Le + L
f) / 2 is the rounding method Z when the cumulative average of rounding errors is negative
₄₁ : Ie = (Le + Lf + 1) / 2 is set.

When the rounding error value cumulative average value is used for each block size, as shown in FIG. 14, a method of taking a weighted average of the areas of the referred pixels can be considered. In FIG. 14, for the sake of simplicity, the motion compensation unit and the unit for accumulating the sum are set to every four pixels.

(Structure of Recoding Device) FIG. 15 is a block diagram showing the structure of the first embodiment of the recoding device according to the present invention. Here, except for the motion compensator 10, FIG.
Since it is the same as the conventional re-encoding device shown in FIG. Motion compensator 10 described above
Rounding operation suppresses accumulation of rounding error and achieves high quality re-encoding. Even if the motion compensator 10 is configured as shown in FIG. 10, the same effect as described above can be obtained.

Here, the discrete cosine transform (DC
Although a re-encoding device using T) has been illustrated, the present invention uses DC
Needless to say, it is applicable even when T is not used. Furthermore, it goes without saying that the present invention is applicable to general re-encoding devices using motion compensation of re-quantized error signals.

FIG. 16 is a schematic block diagram showing an example of an information processing system which implements the re-encoding device according to the present invention. The re-encoding device including the motion compensator 10 according to the present invention is, as apparent from the above description,
Although it can be configured by hardware, it can also be implemented by a computer program.

The information processing system shown in FIG. 16 includes a processor 121, a program memory 122, and a storage medium 123.
And 124. Storage media 123 and 124
May be separate storage media or different storage areas of the same storage medium. As a storage medium,
A magnetic recording medium such as a hard disk can be used.

The processor 121 uses the program memory 1
By executing the program stored in 22, it is possible to realize each function of the above-described re-encoding device according to the first, second, or third embodiment. For example,
In the program memory 122, the block 30 shown in FIG.
1 to 312 and a program that implements the motion compensator 10 may be stored and executed by the processor 121. Specifically, a high bit rate (for example, 22 Mbps) bit stream is input from the hard disk 123, and the processor 121 performs the above-mentioned re-encoding process.
A low bit rate (for example, 10 Mbps) bit stream is output to the hard disk 124.

[0156]

As described above in detail, according to the present invention, the rounding error accumulates and the image quality deteriorates depending on the distribution of the error value and the like, but according to the present invention, the expected value of the rounding error approaches 0. Therefore, the accumulation of rounding error is suppressed and the image quality is not deteriorated. Therefore, high quality re-encoding can be achieved.

Further, by using the statistical information of the rounding error, the rounding method can be determined according to the actual occurrence of the error, and the accumulation of the rounding error can be suppressed more effectively.

In the present invention, since rounding error is unlikely to be accumulated, there is an effect that the storage capacity for storing the error signal can be reduced. Further, by simplifying the condition of the rounding method, high-speed processing of the program and simplification of the apparatus can be realized while maintaining the image quality equivalent to that of the normal method.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a motion compensation unit in a first embodiment of a re-encoding device according to the present invention.

FIG. 2 is a diagram showing a list of rounding methods used in the first embodiment.

FIG. 3 is a flowchart showing an operation of a rounding determination unit 101 in the first embodiment.

FIG. 4 is a flowchart showing an operation of a rounding method determination unit in the first exemplary embodiment of the present invention.

FIG. 5 is a flowchart showing an operation of a rounding method determination unit in the second exemplary embodiment of the present invention.

FIG. 6 is a flowchart showing an operation of a rounding method determination unit according to the second embodiment of the present invention.

FIG. 7 is a flowchart showing an operation of a rounding method determination unit in an example of the second exemplary embodiment of the present invention.

FIG. 8A is a diagram showing a distribution of positively biased error values, and FIG. 8B shows a distribution of error values when two rounding methods of rounding and truncation are switched for each P picture. It is a figure.

FIG. 9 is a peak signal to noise ratio of a color difference signal when an MPEG-2 VIDEO bit stream is actually re-encoded.
It is a graph which shows (PSNR).

FIG. 10 is a block diagram showing a configuration of a motion compensation unit in the third embodiment of the re-encoding device according to the present invention.

FIG. 11 is a flowchart showing an operation of a rounding method determination unit in the third exemplary embodiment of the present invention.

FIG. 12 is a flowchart showing an operation of a rounding method determination unit in an example of the third exemplary embodiment of the present invention.

FIG. 13 is a schematic diagram for explaining the operation of an example of the third exemplary embodiment of the present invention.

FIG. 14 is a schematic diagram for explaining the operation of an example of the third exemplary embodiment of the present invention.

FIG. 15 is a block diagram showing a configuration of a first embodiment of a re-encoding device according to the present invention.

FIG. 16 is a schematic block configuration diagram showing an example of an information processing system implementing a re-encoding device according to the present invention.

FIG. 17 is a block diagram showing a circuit configuration of a basic re-encoding device.

FIG. 18 is a circuit diagram showing an example of a conventional re-encoding device.

FIG. 19 is a schematic diagram for explaining an interpolation process in a motion compensation process with half-pixel accuracy.

FIG. 20 is a graph showing an average value of error values per frame in a luminance signal.

[Explanation of symbols]

10 Motion compensation section 101 Rounding method determination unit 102 error prediction signal synthesis unit 110 Rounding method determination unit 111 error prediction signal synthesis unit 112 Statistical information storage 201 variable length decoder 202 Dequantizer 203 Inverse Discrete Cosine Transform 204 adder 205 motion compensator 206 memory 207 Subtractor 208 motion compensator 209 Discrete Cosine Transform 210 quantizer 211 variable length encoder 212 Dequantizer 213 Inverse Discrete Cosine Transform 214 adder 215 memory 301 variable length decoder 302 Inverse quantizer 303 Inverse Discrete Cosine Transform 304 subtractor 305 Motion Compensator 306 Discrete Cosine Transform 307 Subtractor 308 quantizer 309 Variable length encoder 310 Dequantizer 311 Inverse Discrete Cosine Transform 312 memory

   ─────────────────────────────────────────────────── ─── Continued front page    F-term (reference) 5C053 FA23 GA11 GB17 GB22 GB26                       GB30 GB32 GB38 KA03 LA14                 5C059 KK01 KK08 MA00 MA23 MC11                       MC38 ME01 NN15 NN21 PP06                       PP07 PP16 RC16 SS11 TA61                       TB04 TB05 TB08 TB10 TC03                       TC11 TC37 UA02 UA05 UA33                 5D044 AB07 BC01 CC05 GL01 GL22                       GM21                 5J064 AA00 BA01 BA16 BB14 BC00                       BC08 BC14 BC16 BC21 BC29                       BD02 BD03 BD04

Claims (27)

[Claims] 1. A method of inputting and decoding a compressed moving image stream of a predetermined encoding method, executing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length, and then encoding again. An image signal obtained by decoding at least a part of the compressed moving image stream is requantized, and a requantization error signal is calculated from the signal before requantization and the signal after requantization, In the process of performing the motion compensation prediction process on the quantized error signal, the rounding method of the rounding process is changed for each processing unit that does not depend on the encoding process when the input compressed moving image stream is generated. A method for re-encoding a compressed moving image, characterized by: 2. The predetermined encoding method is an encoding method in which conversion processing into a frequency space is combined, and an image signal obtained by decoding at least a part of the compressed moving image stream is the compressed moving image. It is a frequency coefficient obtained by decoding at least one part of a stream, The re-encoding method of the compressed moving image of Claim 1 characterized by the above-mentioned. 3. The processing unit in which the rounding method is changed is
The method of re-encoding a compressed moving image according to claim 1 or 2, wherein at least one frame / field is used. 4. The processing unit in which the rounding method is changed is
The method of re-encoding a compressed moving picture according to claim 1 or 2, wherein at least one small block in a frame / field. 5. The processing unit in which the rounding method is changed is
3. The re-encoding method for a compressed moving picture according to claim 1, wherein the re-encoding is at least one pixel within a frame / field. 6. A plurality of the rounding methods are preset for each coding type of different frames / fields,
3. The re-encoding method for a compressed moving image according to claim 1, wherein one of the plurality of rounding methods is selected for each processing unit. 7. The re-encoding method according to claim 6, wherein at least one of the frame / field coding types has a fixed rounding method. 8. The motion compensation prediction process includes a plurality of interpolation processes, and at least one rounding method is set corresponding to each of the plurality of interpolation processes. Alternatively, the compressed video re-encoding method according to the item 2. 9. The motion compensation prediction process includes a plurality of interpolation processes, and at least one rounding method is set corresponding to each of the plurality of interpolation processes. A method for re-encoding a compressed moving image as described. 10. A method of inputting and decoding a compressed moving image stream of a predetermined encoding method, executing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length, and then encoding again. An image signal obtained by decoding at least a part of the compressed moving image stream is requantized, and a requantization error signal is calculated from the signal before requantization and the signal after requantization, A process that does not depend on the encoding process when the input compressed moving image stream is generated in the process of generating the statistical information of the quantization error signal and executing the motion compensation prediction process on the requantization error signal. A re-encoding method for a compressed moving image, wherein the rounding method of the rounding process is changed for each unit based on the statistical information. 11. The predetermined encoding method is an encoding method in which conversion processing into a frequency space is combined, and an image signal obtained by decoding at least a part of the compressed moving image stream is the compressed moving image. It is a frequency coefficient obtained by decoding at least one part of a stream, The re-encoding method of the compressed moving image of Claim 10 characterized by the above-mentioned. 12. The re-encoding method for a compressed moving image according to claim 10, wherein the statistical information is generated based on an error value of the re-quantization error signal for each pixel. 13. The recoding of a compressed moving image according to claim 10, wherein the statistical information is generated based on an average value of error values of the requantization error signal in a plurality of pixels. Method. 14. The compressed moving image according to claim 13, wherein the statistical information is generated based on an average value of error values of the requantization error signal in a small block in a frame / field. Re-encoding method. 15. A compressed moving image stream of a predetermined encoding method is input and decoded, a motion compensated prediction process in units of less than pixel precision including rounding processing to a finite length is executed, and then a compressed image to be encoded again. In the re-encoding device, re-quantization means for re-quantizing an image signal obtained by decoding at least a part of the compressed moving image stream, a signal before re-quantization and a signal after re-quantization Error calculation means for calculating a requantization error signal from the error prediction signal combining means for executing the motion compensation prediction process on the requantization error signal, and the input compression in the process of the motion compensation prediction process. A re-compressed moving image re-composition comprising: a rounding method determining unit that changes the rounding method of the rounding process for each processing unit that does not depend on the encoding process when the moving image stream is generated. Goka apparatus. 16. A compressed moving image stream of an encoding method that combines conversion processing into frequency space is input and decoded, and after performing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length. A re-encoding device for re-encoding a compressed image, wherein the first image signal obtained by decoding at least a part of the compressed moving image stream is converted into a frequency space to generate a frequency coefficient; Requantizing means for requantizing frequency coefficients; dequantizing means for dequantizing the requantized signal; and inverse transforming means for transforming the dequantized signal into a second image signal, Error calculation means for calculating a requantization error signal from the first image signal and the second image signal; error prediction signal synthesizing means for executing the motion compensation prediction process on the requantization error signal; Rounding method determining means for changing the rounding method of the rounding processing for each processing unit that does not depend on the encoding processing when the input compressed moving image stream is generated in the process of the motion compensation prediction processing; A re-encoding device for compressed moving images, characterized by: 17. A compressed moving image stream of a predetermined encoding method is input and decoded, and after performing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length, a compressed image to be encoded again. In the re-encoding device, a re-quantization means for re-quantizing an image signal obtained by decoding at least a part of the compressed moving image stream, a signal before re-quantization and a signal after re-quantization An error calculation means for calculating a requantization error signal from the error prediction signal combining means for executing the motion compensation prediction process on the requantization error signal to generate an error value by the rounding process; Accumulating means for accumulating statistical information of error values due to processing, in the process of the motion compensation prediction processing, for each processing unit that does not depend on the encoding processing when the input compressed moving image stream is generated, Serial statistics recoding device of the compressed moving image and having a a method determination means rounding changes the rounding mode of the rounding process based on. 18. The predetermined encoding method is an encoding method combining a conversion process into a frequency space, and an image signal obtained by decoding at least a part of the compressed moving image stream is the compressed moving image. The re-encoding device according to claim 17, wherein the re-encoding device is a frequency coefficient obtained by decoding at least a part of the stream. 19. The re-encoding device for a compressed moving image according to claim 17, wherein the statistical information is generated based on an error value of the re-quantization error signal for each pixel. 20. The recoding of a compressed moving image according to claim 17, wherein the statistical information is generated based on an average value of error values of the requantization error signal in a plurality of pixels. Device. 21. The compressed moving image according to claim 20, wherein the statistical information is generated based on an average value of error values of the requantization error signal in a small block in a frame / field. Re-encoding device. 22. A compressed moving image stream of a predetermined encoding method is input and decoded, and after motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length is performed, requantization is performed and then again. In a program for causing a computer to execute a process of encoding, a step of determining whether or not it is a head of a process unit that does not depend on an encoding process when the input compressed moving image stream is generated, and the process unit The step of determining whether or not the image is a pre-designated image at the beginning of the, and if it is the pre-designated image, the motion compensation prediction process is performed on the requantized error signal. And a step of changing the rounding method of the rounding process in correspondence with the designated image in the process. 23. A compressed moving image stream of a predetermined encoding method is input and decoded, a motion-compensated prediction process in units of less than pixel precision including rounding processing to a finite length is executed, and then re-quantization is performed, and again. In a program for causing a computer to perform a process of encoding, a requantization error signal is calculated from a signal before requantization and a signal after requantization, and statistical information of the requantization error signal is calculated. A step of calculating, a step of determining whether or not it is the head of a processing unit that does not depend on the encoding processing when the input compressed moving image stream is generated, and when the image is at the head of the processing unit, the image is A step of determining whether or not the image is a pre-specified image, and if the image is a pre-specified image, in the process of performing the motion compensation prediction process on the requantized error signal,
Changing the rounding method of the rounding process based on the statistical information corresponding to the designated image. 24. After inputting and decoding a compressed moving image stream of an encoding method in which conversion processing into frequency space is combined, and performing motion compensation prediction processing in units of less than pixel precision including rounding processing to a finite length. In the compressed image re-encoded stream data generated by re-encoding, the first image signal obtained by decoding at least a part of the compressed moving image stream is converted into a frequency space to generate frequency coefficients. Requantizing the frequency coefficient, dequantizing the requantized signal, converting the dequantized signal into a second image signal, re-quantizing from the first image signal and the second image signal, A quantized error signal is calculated, the motion-compensated prediction process is performed on the re-quantized error signal, and the input compressed moving image stream is generated in the process of the motion-compensated prediction process. And for each processing unit that is independent of the encoding process time, changing the rounding mode of the rounding, compressed moving picture re-encoding stream data, characterized in that it is produced by. 25. Re-encoded stream data of a compressed moving image generated by the compressed moving image re-encoding method according to any one of claims 1 to 14. 26. A plurality of the rounding methods is preset for each different prediction method, and one of the plurality of rounding methods is set in advance.
3. The re-encoding method for a compressed moving image according to claim 1, wherein one is selected for each processing unit. 27. The rounding method is fixed for at least one of the prediction methods.
A method for re-encoding compressed moving pictures.