JPH07203430A - Image coding device - Google Patents

Abstract

PURPOSE:To controls data to have a prescribed bit number and to decide an optimum quantization width by modeling a relation among a quantization width, a coding bit number and distortion and using the model so as to obtain the quantization width minimizing picture coding distortion. CONSTITUTION:A model decision section 21 uses plural quantization sections arranged in parallel to apply quantization to data with different quantization width, and calculates a bit number produced for each quantization. Then the generated bit number is expressed as a function of the quantization width and a parameter included in the function is decided. Furthermore, picture data to be quantized are subject to inverse quantization at a corresponding inverse quantization section and distortion with respect to the quantization width is calculated based on the data before quantization and data subject to inverse quantization. Then the distortion is expressed as a function of the quantization width and a parameter included in the function is decided. A quantization width decision section 22 uses two mathematical models obtained in the decision section 21 earlier to control bit number after compression into an allocated bit number and decides the quantization width minimizing the distortion.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image coding apparatus for digital images, and more particularly to high efficiency coding of image data.

[0002]

2. Description of the Related Art Conventionally, there has been proposed a method for compressing a digital moving image by using techniques such as interframe prediction, orthogonal transform, quantization, variable length coding and the like. In addition, ISO / IECJTC1 International Conference on Storage Media Video Coding
/ SC29 / WG11 (MPEG, Moving Picture Cod
ing Experts Group) is aiming to create international standards. When applying these methods to image transmission and image storage, design efforts are being made on how to control the amount of data after compression.

For example, the document "MPEG2 Quantization and Coding Control" (Technical Report Vol.
61. pp. The method described in (43-48) controls the average number of bits in a given section (usually a dozen frames) to be constant. FIG. 12 is a block diagram of an image coding apparatus using this conventional method.

In FIG. 12, a block division unit (10)
Divides a frame to be encoded into blocks called macroblocks (hereinafter abbreviated as MB). The motion compensation prediction unit (11) predicts the MB from other frames,
The prediction error is calculated. However, prediction is not performed for a MB of a frame (referred to as an I picture) for which only intra-frame coding is performed or for an MB for which intra-frame coding is more advantageous. For the prediction method, refer to "MPEG2 Interframe Prediction Method" (Technical Report of the Television Society of Japan, Vol. 16, No. 6).
1. pp. 37-42).

The orthogonal transform unit (13) converts the MB data into 2
The data is converted into data suitable for encoding by a dimensional orthogonal transform (discrete Fourier transform, Hadamard transform, discrete cosine transform, etc.). The quantization unit (14) includes a quantization width determination unit (3
The transform coefficient is quantized according to the quantization width obtained in 2). A variable length coding unit (15) performs variable length coding on the quantized transform coefficient. The inverse quantization unit (16) inversely quantizes the quantized transform coefficient, and the inverse orthogonal transform unit (17) inversely orthogonally transforms the inverse quantized transform coefficient to obtain a restored value. The frame memory (18) stores the restored image obtained from the restored MB. This restored image is used as a reference image for inter-frame prediction by the motion compensation prediction unit (11) through the block inverse division unit (20). However, if the restored value is a prediction error, the restored image is obtained by taking the sum with the predicted value from the frame memory (18).

The bit allocation unit (23), the reference quantization width determination unit (31), and the quantization width determination unit (32) are units for performing coding control by the conventional method. The bit allocation unit (23) stores the coding results up to the previous frame (the number of generated bits obtained from the variable length coding unit (15) and the average value of the quantization width obtained from the quantization width determination unit (32)). , Based on the given data rate, determine the number of bits to allocate to the frame to be encoded.

The reference quantization width determination unit (31), based on the number of generated bits up to the previous MB obtained from the variable length encoding unit (15) and the number of allocated bits, the reference quantization of the orthogonal transform coefficient. Determine the conversion width. The quantization width determination unit (32)
The activity indicating the fineness of the MB pattern obtained by the block dividing unit (10) is calculated, the reference quantization width is changed by this, and the quantization width used in the quantization unit (14) is determined. Here, it is assumed that the quantization width used in the quantization unit (14) is M types of q _{1 to} q _M.

Next, the reference quantization width determining unit (31) will be described. The reference quantization width is calculated by the following equation (1).

[Equation 1] Here, i (i = 1, 2, ..., MBcnt) is the number of MB in one frame (MBcnt is the total number of MBs), r is a constant called a reaction parameter, and d (i) is the following equation (2) ) Is the virtual buffer fullness.

[Equation 2]

In the equation (2), d (0) is a virtual buffer fullness at the start of encoding of the frame, and takes over the final value at the previous encoding. However, the initial value is a constant. b (i) is the sum of the number of bits generated in the 1st to i-th MBs, and T is the number of bits allocated to one frame by the bit allocation unit (23). That is, regarding the code amount generated in the coding of the frame up to that time (MB number 1 to (i-1)) at the time of starting the coding of the MB number i, the second term b on the right side of the equation (2) (I-
1) indicates the actually generated code amount, and the third term (i-1) T / MBcnt on the right side of the equation (2) indicates the target code amount that should occur.

Therefore, the equation (2) works to reflect the difference between the number of allocated bits up to the previous MB and the actual number of generated bits in the virtual buffer fullness, assuming that T bits are allocated to each MB evenly. It is thought that there is. That is, when the number of generated bits up to the previous MB is large, the virtual buffer fullness is large and the reference quantization width is also large, so that it works to reduce the bit generation amount. Conversely, if the number of generated bits up to the previous MB is small, the virtual buffer fullness becomes small and the reference quantization width also becomes small, so that it works to increase the amount of generated bits.

In practice, a frame for performing only intra-frame coding (called an I picture), a frame for performing predictive coding from a temporally previous frame (called a P picture),
There is a frame (referred to as a B picture) for which predictive coding is performed from frames preceding and succeeding in time, the virtual buffer fullness is calculated for each, and the reference quantization width is determined.

[0012]

However, according to the above-mentioned conventional image coding method, the reference quantization width is changed depending on the number of generated bits and the number of allocated bits to control the data amount. An appropriate quantization width is not always selected. That is, the reference quantization width determination unit (31) and the quantization width determination unit (32) in FIG. 12 do not always perform the quantization control that can obtain the minimum distortion with respect to the allocated number of bits.

That is, for example, in the case of an image in which a simple pattern appears after a complicated pattern in one frame, the number of bits generated in the complicated pattern is large, so that the virtual buffer fullness becomes large at the boundary portion and the simple pattern is simple. However, there is a problem that the image quality is deteriorated because the quantization width is set to a large value even though the pattern is changed. On the contrary, when changing from a simple picture to a complicated picture in one frame, the virtual buffer fullness becomes small and the quantization width is set to be small regardless of the complicated picture, and the subsequent image quality is deteriorated. There was a problem that it would end up.

In view of the above problems, it is an object of the present invention to solve the above problems and to provide an image coding apparatus which controls to a given number of bits and determines an optimum quantization width.

[0015]

In order to solve the above problems, the present invention has the following constitution. That is, the present invention is a block division unit that divides an image into a plurality of image blocks,
An image code that includes a quantizer that sets a quantization width for each of the divided image blocks to perform quantization, and an encoder that encodes the quantized image data, and performs image code amount control In the coding device, a modeling unit that models the relationship between the quantization width and the number of code bits and the relationship between the quantization width and the distortion, and the image coding distortion within a predetermined number of code bits using the model. And a quantization width determining unit that determines a quantization width that minimizes

Further, in the present invention, a plurality of quantizers arranged in parallel for quantizing the image block by using different quantizing widths, and a plurality of quantizers respectively provided corresponding to the quantizers are provided. It is possible to include a plurality of code length calculation units that calculate the obtained code lengths, and a modeling unit that models the relationship between the quantization width and the number of code bits based on the code lengths.

Further, according to the present invention, a plurality of dequantizing units for dequantizing the transform coefficients quantized by the plurality of quantizing units, an input of the quantizing unit and an output of the dequantizing unit. And a plurality of distortion calculation units that calculate the distortions according to a predetermined evaluation criterion, and a modeling unit that models the relationship between the quantization width and the distortion based on the distortion calculation results of the distortion calculation units. You can

[0018]

In the modeling of the relationship between the number of generated bits and the quantization width, the plurality of quantizers arranged in parallel perform quantization with different quantization widths, and the generated bit is generated for each quantization. The number is calculated. And
The number of generated bits is expressed as a function of the quantization width, and the parameters included in that function are determined. In modeling the relationship between the distortion caused by encoding and the quantization width, the image data quantized with different quantization widths are inversely quantized by the corresponding inverse quantization units, respectively, and inversely quantized with the data before quantization. The distortion with respect to the quantization width is calculated from the quantized data. Then, the distortion is expressed as a function of the quantization width, and the parameters included in the function are determined.
The quantization width determination unit controls the number of bits after compression to the allocated number of bits using the model of the relationship between the number of generated bits and the quantization width and the model of the relationship between distortion due to encoding and the quantization width. , And a quantization width that minimizes distortion is determined.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, an embodiment of an image coding apparatus according to the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the first embodiment. In the figure, a block division unit (10), a motion compensation prediction unit (11), an orthogonal transformation unit (13), a quantization unit (14), a variable length coding unit (1
5), the inverse quantization unit (16), the inverse orthogonal transform unit (17), the frame memory (18), and the bit allocation unit (23),
The components are the same as the corresponding components of the conventional image encoding device shown in FIG. 12, and the same reference numerals as those in FIG. 12 are given. The model determination unit (21) and the optimum quantization width determination unit (22) are constituent elements unique to the present invention, and control the number of bits after compression to the assigned number of bits and minimize distortion. This is the part that determines the conversion width.

Next, FIG. 2 is a detailed block diagram showing the details of the model determining unit (21) of the first embodiment. The model determining unit (21) uses the method of the present invention to model the relationship between the quantization width and the number of generated bits and the relationship between the quantization width and the coding distortion with mathematical expressions. The model determining unit (21) will be described below with reference to FIG.

Quantizer 1 (101) to quantizer M (10)
M) quantizes the transform coefficient using the respective quantization widths q _{1 to} q _M (the quantization width used in the quantization unit (14) is q
It is assumed that there are M types of _{1 to} q _M ). Quantizer 1 (101)
-The code length calculation unit 1 (201) to the code length calculation unit M (20M) connected to the quantization unit M (10M) calculate the code length when the respective quantization results are variable length coded. . The b (q) determination unit (501) is a unit that determines a function b (q) that expresses the relationship between the number of generated bits and the quantization width, which will be described later.

Inverse quantizer 1 (301) to inverse quantizer M
(30M), each quantized result, respectively obtains the restored value of the transform coefficients performs inverse quantization using the quantization width q ₁ to q _M. Inverse quantization unit 1 (301) to inverse quantization unit M (3
0M) respectively connected to the strain calculation unit 1 (401)
The distortion calculation unit M (40M) has quantization widths q _{1 to} q _M , respectively.
The distortion of the orthogonal transform coefficient when is used is calculated. D
The (q) determining unit (502) is a unit that determines a function D (q) representing the relationship between the distortion due to encoding and the quantization width described later.

Next, the functions b (q) and D (q) will be described. b (q) is a function for determining the number of bits generated when variable length coding is performed by quantizing the transform coefficient of MB with the quantization width q. The shape of the function changes depending on the MB pattern and variable-length coding method, but when q becomes large, b
It can be said that (q) becomes small. D (q) is the quantization width q
This is a function for obtaining the distortion that occurs when the transform coefficient is quantized by. The shape of the function changes depending on the MB pattern and the evaluation method of distortion, but it can be said that D (q) increases as q increases.

When the MB to be encoded and the quantization width q are determined,
The number of generated bits and distortion can be actually measured. These are represented as b ° (q) and D ° (q), respectively. Although b ° (q) can be uniquely determined if the encoding method is determined, there are various calculation methods for D ° (q) depending on how the distortion evaluation criterion is used. For example, when MSE (mean squared error) is used as the distortion evaluation criterion, the measured value D ° (q) of D (q) can be calculated by the following equation (3).

[Equation 3]

Here, g (n) is the n-th value of the transform coefficient obtained from the orthogonal transform unit (13), and g '(n,
q) is a value obtained by quantizing and dequantizing g (n) with a quantization width q. However, in the case of two-dimensional conversion, it is assumed that one-dimensional conversion is performed by raster scanning or zigzag scanning. Also, when two-dimensional conversion is performed with a small block obtained by further dividing the MB, the serial number that continues scanning across blocks is set to n. For example, the size of MB is 16 pixels in the vertical direction and 16 pixels in the horizontal direction.
When the block is divided into two blocks and each block is subjected to the 8 × 8 two-dimensional transform, the total number of transform coefficients N = 256.

3 and 4 show examples of the functions b ° (q) and D ° (q). Here, an integer from 1 to 31 is used as q. In addition, MSE (average 2
Multiply error) is used. The distortion evaluation criteria is MS
Besides E, a weighted one as shown in the following equation (4) can be used.

[Equation 4]

Here, w (n) is a weight for the n-th transform coefficient, and is used to reflect a reduction in human visual response to high frequency components. Also, A is M
This is an amount that reflects the fineness of the B pattern and the presence or absence of edges. A is MB obtained from the block division unit (10)
Of pixel values, dynamic range, difference in normalized activity described in Japanese Patent Application No. 5-30074, document "MPEG2 Quantization and Coding Control" (Technical Report of the Television Society, Vol. 16, No. 61. p
p. 43-48), the document "One Method of Quantization Control in Image Coding" (1992 Television Society Annual Conference, 20-1, pp.375-37).
It is calculated by the frequency h described in 6).

Alternatively, it can be obtained as follows. First

[Equation 5] H = (number of pixels where S _x is th _x or more) + (S _y
Is the number of pixels that is more than th _y ) (5) and A is obtained by the function shown in FIG. here,
S _x and S _y are changes in adjacent pixel values in the horizontal direction and the vertical direction, respectively, and th _x and th _y are constants. A _max and H _max in the figure represent the maximum values of A and H, respectively.
The shape of the curve in the figure uses a typical image and is determined in advance by visual experiments. The above is the case where the model determining unit (21) of FIG. 1 is configured as shown in FIG. 2. In this case, the block dividing unit (10) and the motion compensation prediction unit (11) in FIG. 1 are transferred to the model determining unit (21). The signal of is not used.

Next, a second embodiment of the present invention will be described. As the evaluation criterion of distortion, D ° (q) as shown in the following equation (6) can also be used.

[Equation 6]

Here, h (n) is the nth pixel value in the MB, and h '(n, q) is the value of the restored MB when the quantization width q is used. W (n) is a weight for the n-th pixel value, and is calculated from the current pixel value and the neighboring pixel value (for example, the reciprocal of the difference between the current pixel value and the neighboring pixel value is used).

FIG. 6 shows a second embodiment of the model determining unit (21) when using D ° (q) of the equation (6). 2 except the inverse orthogonal transform unit 1 (601) to the inverse orthogonal transform unit M (60M) and the distortion calculation unit 1 (801) to the distortion calculation unit M (80M). h ′ (n, q) is obtained in the inverse orthogonal transform unit 1 (601) to the inverse orthogonal transform unit M (60M). However, when the restored value of the difference value is obtained by the inverse orthogonal transform, the restored value is obtained by adding the prediction values obtained from the motion compensation prediction unit (11). Distortion calculator 1 (80
1) to the strain calculation unit M (80M) calculate the measured strain value according to the equation (6).

Alternatively, D ° (q) as shown in the following equation can be used.

## EQU00007 ## D.degree. (Q) = max [(h (n) -h '(n, q)) ² W (n)] / A (7) This is the squared error of the pixel value in the MB. The maximum value of is the error. Although the evaluation criteria described so far are based on the square error, the absolute value error may be used instead.

Next, a method of determining the functions b (q) and D (q) will be described. The functions b (q) and D (q) are polynomials of q as in the following equation,

[Equation 8]

[Equation 9] The coefficients b _k and D _l are _obtained by the method of least squares by using the data of the actually measured value b ° (q) of the number of bits and the actually measured value D ° (q) of the distortion for various q. Where K ₋ is b (q)
Is the lowest order, K ₊ is the highest order of b (q), L _- is D (q)
, L ₊ is the highest order of D (q). Figure 7,
8 shows an example when K ₋ = −1, K ₊ = 0, L ₋ = L ₊ = 1. These are examples of modeling FIGS. 7 and 8, respectively. The above is the description of the model determination unit (21) of the first and second embodiments.

Next, the optimum quantization width determining unit (22) common to the first and second embodiments will be described. For simplicity, consider a model such as the following equation.

[Equation 10]

D _i (q _i ) = β _i q _i (11) Here, the subscript i (i = 1, 2, ..., MBcnt) represents the number of the MB in one frame. Therefore, b _i (q _i ),
D _i (q _i ), q _i are the generated bit number function, the distortion function, and the quantization width in the i-th block, respectively. These are the same models used in FIGS.

When the total number of bits assigned to one frame is B, the optimum quantization width is

[Equation 12] Satisfies the total strain amount

[Equation 13] Q _i that minimizes. Such q _i can be solved by Lagrange's undetermined multiplier method for various models. In the case of formulas (10) and (11),

[Equation 14] Becomes The optimum quantization width determination unit (22) finds q _i according to this equation and then rounds it to an integer of 1 to M.
When using a quantization width that changes non-linearly, the quantization width may be rounded to a non-linear discontinuous value.

FIG. 9 shows the configuration of the optimum quantization width determining unit (22) of the first embodiment. The counting unit 1 (901) and the counting unit 2 (902) are units that respectively obtain Σ _j (α _j β _j ) ^1/2 and Σ _j c _j from the parameters obtained by the model determination unit. The storage unit (903) is a unit that stores α _i / β _i for one frame. As described above, in the first embodiment, since one frame of data is accumulated and the quantization width is determined, at least one frame of delay is required. However,
If the unit of data amount control is smaller than one frame, the delay can be reduced. q _i calculation unit (904)
Then, the quantization width is obtained using the equation (14). The q _i rounding unit (905) rounds the quantization width to an integer of 1 to M.

In the model of equations (8) to (11), b
Although a polynomial of q has been considered as (q) and D (q), it is not necessarily limited to the polynomial of q. For example

[Equation 15]

A model such as D _i (q _i ) = γ _i lnδ (q _i −1) (16) can also be used. in this case,

[Equation 17]

[Equation 18]

[Formula 19] Becomes Where α _i and γ _i are constants unique to the i-th MB,
β and δ are common constants between MBs. The optimum quantization width determination unit (22) finds q _i according to this equation and then rounds it to an integer of 1 to M. When using a non-linearly varying quantization width, the quantization width is rounded to a non-linear discontinuous value. When the quantization width that changes nonlinearly is used, the quantization width may be rounded to a non-linear discontinuous value. The above is the optimum quantization width determination unit (2) common to the first and second embodiments.
It is the explanation of 2).

Next, a third embodiment of the present invention will be described. The block diagram is the same as in FIG.
1) Except for the optimum quantization width determination unit (22), the same operation as in the first embodiment is performed. FIG. 10 shows a block diagram of the model determining unit (21) of the third embodiment. Here, a method based on the expression (14) will be described. Distributed calculator (911)
Then, the variance σ _i ² of the pixel values in the MB of the original image is calculated.
In the parameter determining unit (912), the parameter c
(Σ _i ² ), α (σ _i ² ) and β (σ _i ² ) are determined. These parameters are all the parameters of Expression (14) expressed as a function of the variance of MB. The shape of this function can be obtained in advance using a typical image.

Here, the parameter is expressed as a function only of the variance, but the dynamic range of the pixel value, the difference between the normalized activities described in Japanese Patent Application No. Heisei 5-30074, and the document "MPEG2 Quantization and Coding". Control ”(Technical Report of the Television Society of Japan, Vol.
61. pp. 43-48), a document "A Method of Quantization Control in Image Coding" (199).
2nd Annual Conference of Television Society, 20-1, pp. 37
The parameters may be expressed as a multivariable function such as the frequency h described in 5-376).

FIG. 11 shows an example of the optimum quantization width determining unit (22) of the third embodiment. The counting unit 1 (921) and the counting unit 2 (922) are the counting units (901, 9) of FIG.
Works the same as 02). The storage unit 1 (923) and the storage unit 2 (924) respectively store Σ _j (α _j β _j ) of the previous frame.
^This is a part for storing ^1/2 , Σ _j c _j . The α _i / β _i calculation unit (925) calculates α _i / β _i for each block. q
_{The i} calculation unit (926) obtains the quantization width using Expression (14) from the contents of the storage unit 1 (923) and storage unit 2 (924) and the result of the α _i / β _i calculation unit (925). . The q _i rounding unit (927) rounds the quantization width to an integer of 1 to M.

Although the embodiments have been described above, these embodiments do not limit the present invention. That is, the present invention can be applied to any encoding method in which an image is divided into blocks, a plurality of quantization widths are used, and a given data amount is controlled. For example, an encoding method that encodes the maximum value, the minimum value, and the quantized value of pixel values in a block without using orthogonal transformation, and also applies when transform coding is applied after subband division. can do.

[0042]

As described above, according to the present invention, in an image coding apparatus that divides an image into blocks and sets a quantization width for each block to control the code amount, the quantization width and the number of bits are set. And a modeling unit that models the relationship between the quantization width and the distortion, and the amount of data is controlled by determining the quantization width that minimizes the distortion by controlling to a given number of bits, so the optimum encoding is performed. The effect is that it can be done.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a first embodiment and a second embodiment of the present invention.

FIG. 2 is a detailed block diagram of a model determination unit in the first embodiment.

FIG. 3 is an example of a relationship between the number of generated bits and a quantization width.

FIG. 4 is an example of a relation between distortion due to encoding and quantization width.

FIG. 5 is an example of activity determination in an embodiment.

FIG. 6 is a detailed block diagram of a model determination unit in the second embodiment.

FIG. 7 is an example of modeling the relationship between the number of generated bits and the quantization width.

FIG. 8 is an example of modeling the relationship between distortion due to encoding and quantization width.

FIG. 9 is a block diagram of an optimum quantization width determination unit in the first embodiment.

FIG. 10 is a detailed block diagram of a model determination unit in the third embodiment of the present invention.

FIG. 11 is a detailed block diagram of an optimum quantization width determination unit in the third embodiment.

FIG. 12 is a block diagram showing a configuration of a conventional example.

[Explanation of symbols]

 10 block division unit 11 motion compensation prediction unit 12, 19 addition circuit 13 orthogonal transformation unit 14 quantization unit 15 variable length coding unit 16 inverse quantization unit 17 inverse orthogonal transformation unit 18 frame memory 21 model determination unit 22 optimal quantization width Decision unit 23-bit allocation unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical indication H04N 5/92 H04N 5/92 H

Claims (3)

[Claims] 1. A block division unit for dividing an image into a plurality of image blocks, and a quantization unit for performing quantization by setting a quantization width for each of the divided image blocks,
In an image coding apparatus that includes a coding unit that codes the quantized image data and that controls the image code amount, the relationship between the quantization width and the number of code bits and the relationship between the quantization width and the distortion are described. An image characterized by including a modeling unit for modeling each and a quantization width determining unit for obtaining a quantization width that minimizes image coding distortion within a range of a predetermined number of code bits using the model Encoding device. 2. The plurality of quantizers arranged in parallel for quantizing the image block using different quantizing widths, and the plurality of quantizers, respectively. Image coding, comprising: a plurality of code length calculation units that are provided to calculate the code length; and a modeling unit that models the relationship between the quantization width and the number of code bits based on the code length. apparatus. 3. The plurality of dequantization units according to claim 1 or 2, which dequantize the transform coefficients quantized by the plurality of quantization units, the input of the quantization unit, and the inverse unit. A plurality of distortion calculation units that respectively input the output of the quantization unit and calculate the distortion based on a predetermined evaluation criterion, and a model that models the relationship between the quantization width and the distortion based on the distortion calculation result of the distortion calculation unit. An image encoding device comprising an encoding unit.