JP3271098B2 - Digital image signal decoding apparatus and method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a digital image signal receiving / reproducing apparatus applied to record / reproduce a digital image signal by, for example, a digital VTR, and more particularly to a method for converting quantized data into a restored value. Di for
The present invention relates to a digital image signal decoding apparatus and method .

[0002]

2. Description of the Related Art When a digital video signal is recorded on a recording medium such as a magnetic tape or the like, the amount of information to be recorded is large. Usually achieved. ADRC, DC, which divides a digital video signal into a number of small blocks and processes each block as high-efficiency coding for compressing a digital video signal,
T (Discrete Cosine Transform) and the like are known.

[0003] DCT performs cosine transform on pixels in a block, and requantizes coefficient data obtained by the cosine transform. Further, variable-length coding is performed on the requantized coefficient data. In this variable-length coding, entropy coding such as Huffman coding is often used. Therefore, the image data is divided into a large number of frequency data from low frequency to high frequency by orthogonal transformation.

When requantization is performed on the divided frequency data, low-frequency data, which is important in consideration of human visual characteristics, is finely quantized, and in consideration of human visual characteristics. For high-frequency data of low importance, there is a feature that by performing coarse quantization, high image quality can be maintained and efficient compression can be realized. Such adaptive quantization is performed in a coefficient data requantization circuit. Further, by varying the quantization step width at the time of requantization, the amount of encoded data can be controlled, and a buffering process that keeps the output data rate constant is possible.

Conventional decoding using DCT converts quantized data for each frequency component into a representative value of the code,
Inverse DCT (IDCT: Inverse DCT)
By performing T), reproduced data is obtained. When converting to this representative value, the quantization step width at the time of encoding is used. When adaptive quantization is performed, the quantization step width at the time of encoding needs to be known on the decoding side, and therefore, this quantization step width is transmitted. This transmission information is called a threshold value TH. 6 to 10 show DCT as an example.
2 shows how to encode and decode the coefficients of each frequency component. FIG. 6 shows the image data. DCT for this image data
And the coefficient data for each frequency component can be obtained. This coefficient data is shown in FIG. After the coefficient data is obtained, quantization is performed.

Here, for simplicity, the coefficient data of each frequency component other than the DC component is divided into two parts, a low-frequency area and a high-frequency area, as shown in FIG. And the coefficient data in the high frequency region is divided by `8 '. It is assumed that quantization is performed by truncating each decimal data portion of the data calculated by this division. '4' and '8' of these quantization steps or information indicating this is transmitted as the threshold value TH. FIG.
FIG. 8 shows an example of quantization performed on the coefficient data shown in FIG. Generally, quantized data is subjected to entropy coding such as Huffman coding, but is omitted here for simplicity.

Next, the decoding operation will be described. Quantized data decoded from an entropy code to a normal code, for example, quantized data shown in FIG. 8, is decoded into coefficient data. On the encoding side, since it is known from the threshold value TH that the coefficient data is divided by `4 'and the quantization process of rounding down the decimal data portion is performed, the quantization data in the low frequency region is multiplied by four, The coefficient data is obtained by multiplying the quantized data in the high frequency region by eight. FIG. 9 shows the result of converting the quantized data shown in FIG. 8 into coefficient data. A decoded image is obtained by performing IDCT on coefficient data as shown in FIG. FIG. 10 shows the result of performing IDCT on the coefficient data shown in FIG. This ID
The CT processing ends, and the decoding ends when the reproduction code is obtained.

[0008]

As described above, the DCT
In compression using orthogonal transform such as that described above, encoding is performed in consideration of human visual characteristics, thereby maintaining high image quality and achieving high-efficiency compression. However, if the quantization of each coefficient data is made coarse in order to increase the compression ratio, the error of the restored image with respect to the original image increases, and the restored image deteriorates. The distortion is blurred in the image,
Appearing in the form of so-called mosquito noise and block distortion at the edge portion, it is a major problem.

Incidentally, image data generally has a strong correlation. In a method in which the data is divided into small regions and quantized with the same step width, and the quantized data is treated as encoded data, the encoded data has a correlation similarly to the image data. Similarly, the orthogonally transformed image data and peripheral coefficient data also have a correlation.

In the conventional method, decoding of coefficient data of a component of interest is performed without using coefficient data of peripheral components. For this reason, when the compression ratio is high, there is a problem that image deterioration is conspicuous. By utilizing the local correlation between coefficient data, it is possible to improve the decoding accuracy of the coefficient data and create a decoded image with reduced image degradation.

Accordingly, it is an object of the present invention to form finer appropriate decoded coefficient data without increasing the amount of information in encoding using orthogonal transform, thereby reducing quantization errors. It is to provide a possible digital image signal decoding device and method .

[0012]

SUMMARY OF THE INVENTION The first aspect of the present invention, the components after orthogonal transformation is a digital image signal decoding apparatus for decoding transmission data including the coefficient data quantized, the quantized coefficients Inverse quantum that inversely quantizes data
Quantization means, the inversely quantized coefficient data of interest and the inversely quantized
For each pattern defined by a plurality of coefficient data, a memory for storing prediction coefficients data for use in correcting the present coefficient data, inverse quantized present coefficient data and inverse
A plurality of coefficient data adjacent to the attention data quantized
Is multiplied by the prediction coefficient data, and the multiplication result is added
By doing so, the dequantized attention coefficient data is corrected,
A correction coefficient data generation means for generating a correction coefficient data, on the basis of the correction coefficient data, and means for generating a decoded data by performing an inverse transform, prediction coefficient
Numerical data is a digital image signal for training
Training coefficients
Data is generated and training coefficient data is quantized
And the quantized training coefficient data
By being quantized, inverse quantized coefficient data is generated.
Multiplies the quantization coefficient data by the prediction coefficient data and multiplies the result
Focus on the coefficient data generated by adding
The error from the training coefficient data is statistically minimized.
As in the value determined by the learning using the least squares method.
There is provided a digital image signal decoding device. According to a fifth aspect of the present invention , each component after the orthogonal transformation is an amount
The coefficient data is corrected for the child coefficient data.
Digital image to generate the filter coefficients used in
In the image signal decoding device, the input training
Tray that performs orthogonal transformation on digital image signals
Orthogonal transformation means for generating the coefficient data for
Quantization means for quantizing the coefficient data for
Dequantizing the modified training coefficient data
A dequantizing means for generating dequantized coefficient data,
Based on the dequantized coefficient data
Classifying means for classifying into one of a plurality of patterns
And the inverse quantization coefficient data and filter coefficient for each pattern
And the coefficient generated by multiplying by
Data and the training corresponding to the noted coefficient data
As error between grayed coefficient data is minimized in statistical,
Generate filter coefficients by learning using the least squares method
Digital having coefficient generation means.
An image signal decoding device. According to a sixth aspect of the present invention, in the digital image signal decoding method for decoding transmission data including coefficient data in which each component after the orthogonal transformation is quantized, the step of dequantizing the quantized coefficient data is performed.
For each pattern determined by the dequantized target coefficient data and the dequantized coefficient data, the prediction coefficient data used when correcting the target coefficient data is recorded.
And outputting the re, the inverse quantized present coefficient data, with respect to a plurality of coefficient data adjacent to the inverse quantized data of interest, multiplying the prediction coefficient data, a multiplication result
By adding, corrects the inverse quantized present coefficient data, and generating a correction coefficient data, on the basis of the correction coefficient data, a step of generating a decoded data by performing an inverse transform Have, prediction coefficient data
Is applied to the digital image signal for training.
By performing the orthogonal transformation, the training coefficient data is
Generated, the training coefficient data is quantized and updated.
The training coefficient data quantized to
The inverse quantization coefficient data is generated by the
Multiplies numeric data and prediction coefficient data and adds the multiplication results
Data and the trainees
Error with the coefficient data for
A digital image signal decoding method characterized in that the value is determined by learning using the least squares method . According to a tenth aspect of the present invention , each component after the orthogonal transformation is quantized.
When correcting coefficient data for the set coefficient data
Digital image signal generating filter coefficients used for
In the decoding method, the input training
Training to perform orthogonal transformation on digital image signal
Generating coefficient data for training;
Quantizing the coefficient data for
Inverse quantization by inversely quantizing the training coefficient data
Generating the normalized coefficient data;
Data based on the inverse quantization coefficient data.
Classifying into one of the
Product data and filter coefficients, and add the multiplication result.
The coefficient data generated by calculation to attention coefficients
Error from the training coefficient data corresponding to the data
For learning using the least squares method so that it is statistically minimized
Generating filter coefficients.
Characteristic digital image signal decoding method.

[0013]

The correction data in which the filter coefficients for generating appropriate decoding coefficients are stored using the coded data of the coefficient of the component of interest and the coded data of the coefficients of a plurality of adjacent components.
Over data table is prepared. A present coefficient data to the correction data table <br/> le, a plurality of coefficient data, pre
The measurement coefficient data is input, an appropriate decoding coefficient is output, and IDCT is performed on the coefficient to obtain more appropriate decoding data than before.

[0014]

An embodiment of the present invention will be described below. FIG. 1 illustrates this embodiment, namely, the digital VT.
1 shows a schematic configuration of R signal processing. A video signal is supplied from an input terminal 1 and one sample is subjected to, for example, 8-bit digital processing by the A / D converter 2. The output data of the A / D converter 2 is supplied to the blocking circuit 3. In this embodiment, in the blocking circuit 3, the effective area of one frame is (4 × 4) pixels, (8 × 4) pixels.
8) The image is divided into blocks each having a size such as a pixel.

A digital video signal scan-converted in the order of blocks from the blocking circuit 3 is supplied to a shuffling circuit 4. In the shuffling circuit 4, for example, shuffling is performed in block units. The output of the shuffling circuit 4 is supplied to the block encoding circuit 5.
The block encoding circuit 5 compresses the pixel data by requantizing the pixel data for each block. Here, the shuffling circuit 4 may be provided after the block encoding circuit 5.

In this embodiment, the block coding circuit 5 performs DCT processing on an input signal divided into, for example, (8 × 8) blocks. As a result, coefficient data of 63 AC components and coefficient data of one DC component can be obtained from the input signal divided into (8 × 8) blocks. The coefficient data of the AC component is requantized. As a method of requantization, in this embodiment, as an example, a method of dividing by a quantization step width of each of a low band and a high band indicated by a threshold value TH and performing truncation after a decimal point is used. . The requantized coefficient data of the plurality of AC components is subjected to variable-length coding having different bit lengths, that is, entropy coding, according to the appearance probability, and is output from the block coding circuit 5. The DC component coefficient data is output from the block encoding circuit 5 without being subjected to requantization and entropy coding.

The output data of the block encoding circuit 5 is supplied to a framing circuit 6. Recording data is generated from the framing circuit 6. The framing circuit 6 generates parity of an error correction code and generates recording data having a structure in which sync blocks are continuous. As the error correction code, for example, a product code for performing error correction coding in each of a horizontal direction and a vertical direction of a matrix arrangement of data can be adopted. In the sync block, a sync block synchronization signal and an ID signal are added to encoded data and parity. The recording data in which the sync blocks are continuous is transmitted to the channel encoding circuit 7.
The recording data is supplied to the channel encoding circuit 7 and subjected to a channel encoding process for reducing the DC component of the supplied recording data.

The output data of the channel encoding circuit 7 is converted into a bit stream, and is further supplied to a rotary head H via a recording amplifier 8 so that the recording data is transferred to a magnetic tape T.
Recorded as diagonal tracks on top. Typically, multiple rotating heads are used, but for simplicity, only one head is shown.

The reproduction data taken out of the magnetic tape T by the rotary head H is supplied to a channel decoding circuit 12 via a reproduction amplifier 11 and is subjected to channel coding decoding. The output data of the channel decoding circuit 12 is supplied to a frame decomposing circuit 13, where various kinds of data are separated from the recording data and subjected to error correction processing. The output data generated from the frame decomposition circuit 13 includes an error flag indicating the presence or absence of an error after error correction in addition to the reproduction data.

The output data of the frame decomposition circuit 13 is supplied to a block decoding circuit 15. Further, the block decoding circuit 15, as described below, the decoded coefficients generated with reference to correction data Te <br/> Buru for correction, it by performing IDCT, adapted to generate a decoded value ing.

The decoded data of the block decoding circuit 15, that is, restored data corresponding to each pixel is supplied to the deshuffling circuit 16. This deshuffling circuit 1
Numeral 6 is complementary to the shuffling circuit 4 on the recording side.
A process of returning the spatial position of the block to the original position is performed.
Output data of the deshuffling circuit 16 is supplied to a block decomposition circuit 17. In the block decomposition circuit 17, the order of the data is returned to the order of the raster scanning. Output data of the block decomposition circuit 17 is supplied to an error interpolation circuit 18. The error interpolation circuit 18 performs error detection on a pixel-by-pixel basis. The pixel data detected as an error is interpolated with peripheral pixel data.

As the interpolation processing, for example, an interpolation processing in which an interpolation circuit in a spatial direction, that is, a two-dimensional direction and an interpolation circuit in a time direction are sequentially connected can be used. The output data of the error interpolation circuit 18 is supplied to the D / A converter 19, and the output terminal 20
, Restored data in the raster scanning order corresponding to each pixel is obtained.

The present invention is applied to the block decoding circuit 15 described above. FIG. 2 shows an example of the block decoding circuit 15 according to the present invention. Reproduction data is supplied from an input terminal denoted by reference numeral 31, and the reproduction data is supplied to the frame decomposition circuit 13. In the frame decomposing circuit 13, the coding coefficient data DT and the threshold value TH are separated and taken out from the supplied reproduction data. The data extracted from the frame decomposition circuit 13 is
2, the encoded coefficient data DT is subjected to a representative value conversion by multiplying each of the low-pass and high-pass quantization step widths indicated by the threshold value TH after the entropy code is decoded. The coefficient data CD of each component is decoded. Of the coefficient data CD of each component, the coefficient data of the component with higher importance is supplied to the quantization circuit 33, and the coefficient data of the other components are supplied to the memory. In this example, the data of the component with high importance is assumed to be five components (coefficient data CD1 to CD5) starting from the low-order component as shown in FIG. Here, the memory 36 holds the supplied coefficient data, and outputs the coefficient data from the memory 36 to the IDCT circuit 38 after a delay for a predetermined time.

The quantization circuit 33 has a low-order 5
Quantizes the coefficient data of the component. This quantization is for reducing the number of patterns of the coefficient data of the component of high importance. Since the coefficient obtained by performing the DCT processing has a characteristic of a strong concentration tendency centering on `0`, for example, nonlinear quantization as shown in FIG. 4 is performed. Thereby, when the DCT processing is performed on the coefficient data in the existing range of (−255 to +255), the data after quantization is compressed to the range of (−8 to +8).
Data obtained by performing such quantization on the coefficient data CD1 to CD5 is referred to as quantized data QD1 to QD5.
The coefficient data CD1 to CD5 and the quantized data QD1 to QD5 are supplied from the quantization circuit 33 to the correction data table 34, respectively.

The correction data table 34 stores the coefficient data C
The correction values of D1 to CD5 are generated. Further, the correction data table 34 is formed of a memory, and stores filter coefficients formed in advance by training as described later. When the patterns of the quantized data QD1 to QD5 are supplied from the quantization circuit 33 to the correction data table 34, the quantization data QD1
Coefficient data CD1 to CD corresponding to the patterns QD5 to QD5
5 are respectively read, and the read filter coefficients are multiplied by the coefficient data CD1 to CD5 to calculate correction coefficient data ND1 to ND5. In the pattern of the quantized data QD1 to QD5,
The filter coefficient for obtaining the correction coefficient data ND1 is (a
1, a2, a3, a4, a5), the correction coefficient data ND1 is obtained by the following equation (1).

ND1 = a1 × CD1 + a2 × CD2 + a3 × CD3 + a4 × CD4 + a5 × CD5 (1)

Similarly, assuming that the filter coefficients for obtaining the correction coefficient data ND2 are (b1, b2, b3, b4, b5), the correction coefficient data ND2 is obtained by the following equation (2).

ND2 = b1 × CD1 + b2 × CD2 + b3 × CD3 + b4 × CD4 + b5 × CD5 (2)

When calculating the correction data for the five components of the correction coefficient data ND1 to ND5, the quantization data QD1
5 sets of filter coefficients are stored for each pattern of .about.QD5. From the correction data table 34, correction coefficient data ND1 to ND5 are calculated. The calculated correction coefficient data ND1 to ND5 are supplied from the correction data table 34 to the data conversion circuit 35.

The data conversion circuit 35 clips the correction coefficient data ND1 to ND5. In accordance with the encoded coefficient data DT and the threshold value TH supplied from the frame decomposition circuit 13 via the coefficient decoding circuit 32, the data conversion circuit 35 calculates the original existence range of the correction coefficient data of each component. For example, the encoding coefficient data DT is
Assuming that the threshold value TH is “4” as 3 ′, 3 × 4 =
It becomes 12. Here, since the threshold value TH is equivalent to the quantization step, when the coding coefficient data DT is “3”, the original value of the coding coefficient data DT is 12 or more and 1
It is present in a range of less than 6. This range is called an existence range.

The existence range and the correction coefficient data ND1 to ND1
By comparing the values of ND5, the correction coefficient data ND1 to ND5
If each value of is not within its original range,
Clipping is performed by replacing the correction coefficient data with the closest value from the coefficient data included in the existence range. For example, when the existing range of the correction coefficient data ND1 is 12 or more and less than 16, the correction coefficient data ND1 becomes 1
When the value is 1.6 and is not included in the existing range, the correction coefficient data ND1 is clipped from 11.6 to 12. The output of the data conversion circuit 35 is supplied to the IDCT circuit 38. The output of the data conversion circuit 35 and the output of the memory 36 are supplied to the IDCT circuit 38, respectively, and the DC component is supplied from the terminal 37 to the IDCT circuit 38.

The IDCT circuit 38 performs IDCT processing on the supplied coefficient data of each component and decodes the data into image data. The decoded image data is supplied to the output terminal 39.

FIG. 5 shows an example of a block diagram during training for creating the correction data table 34. In FIG. 5, a digital video signal is input to an input terminal 41, and the input digital video signal is supplied to a blocking circuit 42. The digital video signal subjected to the blocking in the blocking circuit 42 is supplied from the blocking circuit 42 to the DCT coding circuit 43, and the DCT coding circuit 43 performs DCT coding. This input data is preferably a standard digital video signal for training.

In the DCT encoding circuit 43, the encoded code DT and the threshold value TH after the supplied digital video signal is quantized, and the coefficient data RD of each component before the quantization are respectively performed. Output. In the coefficient decoding circuit 44, the encoded code D
T and threshold TH are supplied respectively. The supplied encoded code DT is decoded according to the threshold value TH to generate coefficient data CD, and the coefficient decoding circuit 44 is a block that performs the same processing as the coefficient decoding circuit 32. The coefficient data CD generated by the coefficient decoding circuit 44 is supplied to a quantization circuit 45. Quantization circuit 45
Is a block that performs the same processing as the quantization circuit 33, and performs quantization of coefficient data of five components having high importance.
This quantization circuit 45 reduces the number of patterns of the coefficient data of the component of high importance.

The simplest method of class classification is a method in which the pattern of the decoded coefficient data of each component is used as a class. However, in this method, the number of patterns becomes enormous, and a very large capacity ROM is required. Therefore, by selecting a component having high importance and performing quantization, the number of classes is effectively reduced.

The quantized data QD1 to QD5 output from the quantization circuit 45, the coefficient data RD1 to RD5 before the quantization output from the DCT encoding circuit 43, and the coefficient output from the coefficient decoding circuit 44 are output. Data CD1 to CD5
Are supplied to the normal equation addition circuit 46, respectively.

Here, the normal equation used in the normal equation adding circuit 46 will be described. Using the above-described coefficient data CD1 to CD5 and the coefficient data RD1 to RD5 before quantization, the quantized data QD1 to QD5
Coefficient w1 _, ... For each class determined by the pattern
_, w _n , the linear estimation equation of n taps is shown as the following equation (3).

RD1 = w ₁ CD1 + w ₂ CD2 + w ₃ CD3 + w ₄ CD4 + w ₅ CD5 (3)

Before training, w _i is an undetermined coefficient. Further, in this method, since each of the coefficient data RD1 to RD5 before the quantization is corrected, an equation must be actually set for each of the coefficient data RD1 to RD5.

The training is performed on a plurality of signal data for each class. When the number of data is m, according to equation (3),

RD1 _j = w ₁ CD1 _j + w ₂ CD2 _j + w ₃ CD3 _j + w ₄ CD4 _j + w ₅ CD5 _j (j = 1, 2,..., M) (4)

When m> n, w _1, ... _, W _n are not uniquely determined.

E_j= RD1_j− ｛W₁CD1_j+ W_TwoCD2_j+ W_ThreeCD3_j+ W_FourCD4 _j + W_FiveCD5_j｝ (J = 1, 2,..., M) (5)

Then, a coefficient that minimizes the following equation (6) is obtained.

[0045]

(Equation 1)

This is a so-called least squares solution. Here, the partial differential coefficient based on w _{i in} equation (6) is obtained.

[0047]

(Equation 2)

It is sufficient to find each w _i so that equation (7) is set to `0 '.

[0049]

(Equation 3)

Using a matrix as

[0051]

(Equation 4)

Is as follows. This equation is generally called a normal equation. The normal equation adding circuit 46 adds the normal equations. After the input of all the training data is completed, the normal equation adding circuit 46 outputs the normal equation data to the prediction coefficient determining circuit 47. The prediction coefficient determination circuit 47 solves w _i using a general matrix solution such as a sweeping method of a normal equation, and calculates a prediction coefficient. The prediction coefficient determination circuit 47 writes the calculated prediction coefficient to the memory 48.

As a result of training as described above,
The memory 48 stores, for each pattern defined by the quantized coefficient data QD1 to QD5, a prediction coefficient statistically closest to the true value for estimating the target coefficient data RD1 (and RD2 to RD5). . The table stored in the memory 48 is the correction data table 34 used in the block decoding circuit 15 described above. Here, each of the obtained w _i is a filter coefficient (a1,
a2, a3, a4, a5) and the like are used at the time of decoding.

In this embodiment, the method of estimating the data of the five components from the data of the five components of high importance has been described. However, the present invention is not limited to this combination.

Further, in this embodiment, the prediction coefficients are stored in the memory 48. However, it is also possible to use a case where the representative values obtained by the centroid method are stored in the memory 48.

[0056]

According to the present invention, since the amount of data to be transmitted can take at least a large number of decoding levels in the coefficient of a component having a high importance, quantization errors and block distortion can be reduced, and restoration can be performed. Images can be good. Further, the present invention has an advantage that the transmission information is not increased at all and is efficient.

[Brief description of the drawings]

FIG. 1 shows a digital V to which the present invention can be applied.
It is a block diagram of a recording / reproducing circuit of TR.

FIG. 2 is a block diagram illustrating an example of a block decoding circuit to which the present invention has been applied.

FIG. 3 is a schematic diagram for explaining a case where a correction data table is created.

FIG. 4 is a block diagram showing a configuration at the time of training for creating a correction data table in one embodiment of the present invention.

FIG. 5 is a schematic diagram for explaining a case where a correction data table is created.

FIG. 6 is a schematic diagram for explaining encoding by orthogonal transformation.

FIG. 7 is a schematic diagram for describing encoding by orthogonal transformation.

FIG. 8 is a schematic diagram for explaining encoding by orthogonal transformation.

FIG. 9 is a schematic diagram for describing encoding by orthogonal transformation.

FIG. 10 is a schematic diagram for explaining encoding by orthogonal transformation.

FIG. 11 is a schematic diagram for explaining a frequency domain after performing orthogonal transformation.

[Explanation of symbols]

 32 coefficient encoding circuit 33 quantization circuit 34 correction data table 35 data conversion circuit 38 IDCT circuit

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) H04N ^7/ 24-7/68

Claims (10)

(57) [Claims] 1. A digital image signal decoding apparatus for decoding transmission data including coefficient data in which each component after orthogonal transformation has been quantized, an inverse quantization method for inversely quantizing the quantized coefficient data.
And the step, for each pattern determined by the inverse quantized present coefficient data and the inverse quantized plurality of said coefficient data, a memory for storing prediction coefficients data for use in correcting the present coefficient data, The inversely quantized coefficient data of interest and the inversely quantized
A plurality of pairs and the coefficient data adjacent to the serial data of interest
And multiplies the prediction coefficient data and adds the multiplication result
The correction coefficient data generating means for correcting the inversely quantized coefficient data of interest and generating the correction coefficient data, and performing the inverse transform on the basis of the correction coefficient data to decode the decoded data. Means for generating, the prediction coefficient data includes a digital data for training.
The orthogonal transformation is performed on the image signal to train
The training coefficient data is generated, and the training coefficient
The data is quantized and further trained
The inversely quantized coefficient data is
Data is generated, and the inverse quantization coefficient data and the prediction
Generated by multiplying numeric data and adding the above multiplication result
Coefficient data and the training coefficient data
The least squares method so that the error from the
A digital image signal decoding device characterized in that the value is a value determined by learning using a digital image signal. 2. The coefficient data of interest is a low-order coefficient data.
2. The digital image signal decoding device according to claim 1, wherein 3. The prediction coefficient data is characterized in that the target coefficient data and coefficient data of a plurality of components are classified based on data obtained as a result of nonlinear quantization. The digital image signal decoding device according to claim 1. 4. A clipping means for clipping the correction coefficient data when the correction coefficient data from the correction coefficient data generation means does not fall within its existing range. Item 2. The digital image signal decoding device according to Item 1. 5. A coefficient obtained by quantizing each component after orthogonal transformation.
Used to correct the above coefficient data for data
Image signal decoding to generate filtered filter coefficients
In the device, the input digital image signal for training
Generates training coefficient data that performs orthogonal transformation on the data
Orthogonal transforming means, and quantizing means for quantizing the training coefficient data.
And inverse quantization of the quantized training coefficient data
In this way, the inverse quantization
The step and the noted coefficient data are calculated based on the dequantized coefficient data.
Pattern classification based on
Means, and for each of the patterns, the dequantized coefficient data and the file
The result is obtained by multiplying
The coefficient data generated and the noted coefficient data
The error from the corresponding training coefficient data is statistically
By learning using the least squares method,
A digital image signal decoding device , comprising: coefficient generation means for generating a filter coefficient . 6. In a digital image signal decoding method for decoding transmission data including coefficient data in which each component after orthogonal transformation is quantized, a step of dequantizing the quantized coefficient data.
If, for each pattern determined by the inverse quantized present coefficient data and the inverse quantized plurality of the coefficient data, prediction coefficients data used again to correct the present coefficient data
Outputting from the memory, the inversely quantized coefficient data of interest, and the inversely quantized
A pair of the coefficient data adjacent to the target data
And multiplies the prediction coefficient data and adds the multiplication result
Doing, it corrects the present coefficient data subjected to inverse quantization and generating a correction coefficient data, based on the correction coefficient data, and a step that occur decoded data by performing an inverse transform And the prediction coefficient data has a digital value for training.
The orthogonal transformation is performed on the image signal to train
The training coefficient data is generated, and the training coefficient
The data is quantized and further trained
The inversely quantized coefficient data is
Data is generated, and the inverse quantization coefficient data and the prediction
Generated by multiplying numeric data and adding the above multiplication result
Coefficient data and the training coefficient data
The least squares method so that the error from the
A digital image signal decoding method, characterized in that the value is a value determined by learning using . 7. The digital image signal decoding method according to claim 6, wherein the coefficient data of interest is low-order coefficient data. 8. The method according to claim 6, wherein the prediction coefficient data is classified based on coefficient data of interest, coefficient data of a plurality of components, and data resulting from nonlinear quantization. Digital image signal decoding method. 9. A clipping step for clipping the correction coefficient data when the correction coefficient data from the correction coefficient data generation means does not fall within its existing range. A method for decoding a digital image signal according to claim 6. 10. A system in which each component after orthogonal transformation is quantized.
Used to correct the above coefficient data for several data.
Digital image signal decoding to generate used filter coefficients
In Patent method, the digital image signal for the input training
Generates training coefficient data that performs orthogonal transformation on the data
And quantizing the training coefficient data.
And inverse quantization of the quantized training coefficient data
And generating the inverse quantized coefficient data by performing the
Classifying the data into one of a plurality of patterns, and for each of the patterns,
The result is obtained by multiplying
The coefficient data generated and the noted coefficient data
The error from the corresponding training coefficient data is statistically
By learning using the least squares method,
Generating a filter coefficient.
Digital image signal decoding method.