JP3289862B2 - Image coding control 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 coding control method in high efficiency image coding.

[0002]

2. Description of the Related Art In conventional high-efficiency image coding in which an image is divided into blocks and orthogonal transformation is performed, a difference between a generated information amount with respect to a target information amount and an initial value of a virtual buffer occupancy amount is set for each of several blocks. Then, the quantization step of the orthogonal transform coefficient is calculated from the above, and the amount of encoded information is controlled by changing the quantization step. The quantization step has been set so as to be proportional to a value obtained by adding the initial value of the buffer occupancy to the difference value. Hereinafter, this prior art will be described in more detail.

[0003] A unit of an image for which a quantization step is determined is called a macroblock. A macroblock is composed of several orthogonal transform blocks, and encoding of the macroblock is performed via orthogonal transform for each orthogonal transform block belonging to the macroblock. In the transform coding of a macroblock, a sample value sequence of an image of each orthogonal transform block is orthogonally transformed, and each transform coefficient is converted into a quantization step q (j) (j is a macroblock number) determined for each macroblock. Is a process of quantizing the transform coefficient value as a unit and encoding the quantized transform coefficient value. As is well known, a typical orthogonal transform is DCT (Discrete Cosin).
e Transform, discrete cosine transform). In the following description, the target information amount per image is T bits,
The number of macroblocks per image is represented by M, and the order of the macroblocks is represented by i. Therefore, i = 1, 2, 3,.
j ... M, and the target information amount per macroblock is T
/ M. Further, when the i-th macroblock is transform-coded, the sum of the number of bits of the generated code is referred to as a generated information amount b (i). In addition, the quantity Q belonging to an arbitrary i-th macroblock is described as Q (i).

The quantization step of an arbitrary macro block j + 1 is
Step q (j + 1) occurs in the macroblock j + 1
Then, the predicted occurrence information amount b (j + 1)^* Is the goal
It is determined so as not to deviate from the information amount T / M. That
Therefore, the predicted occurrence information amount b (j +
1)^* Is larger than the target information amount T / M (data
Is expected to occur too much), the quantization step
Quantize the orthogonal transform coefficient roughly by setting q (j + 1) large
And reduce the amount of generated information by performing variable-length coding.
Conversely, if the amount of generated information is predicted to decrease (pre-
Measurement information b (j + 1)^* Is smaller than T / M
) To increase the amount of generated information by fine quantization
You. In other words, the predicted occurrence information for the target information amount T / M
Quantity b (j + 1) ^* Difference ε (j + 1)^* = B (j + 1)
^* -Quantize step so that (T / M) approaches 0
Change. Predicted occurrence information amount b (j + 1)^* as,
The image of the macroblock j just before the encoding has already been completed.
The information amount of the image information, that is, the previous value b (j) of the generated information amount is
Used. Therefore, the prediction difference ε (j + 1)^* Is ε
(J) = b (j)-(T / M), and ε (j)
Is positive, the quantization step q (j + 1) is
(J) is set larger than q, and when ε (j) is negative, q
To determine (j + 1) smaller than the previous value q (j)
Therefore, the difference ε (j) (the predicted difference ε (j + 1)^* ) To 0
You can get closer.

Conventionally, in order to change q (j + 1) with respect to ε (j) as described above, Δq (j + 1) = q (j +
1) -q (j) was placed so as to be proportional to ε (j). That is, q (j + 1) = q (j) + C [b (j)-(T / M)] (1) where C is a proportional constant. Also, for j = 0, b
Since (j) is not defined, macroblock 1 (j + 1
= 1) is equal to the initial value q
(0) is set. Q (j + 1) based on equation (1)
Is represented by an initial quantization step q (0) and an amount of generated information b (i), the following equation is obtained.

Q (j + 1) = q (0) + C (Σ _j b (i) −jT / M) (2) where Σ _j is the sum b (1) + b from i = 1 to i = j
(2) represents + ‥‥ + b (j). B (i) in equation (2)
Is the number of bits of the image information of the macroblock i for which the transform coding has already been completed. By dividing the equation (2) by the proportionality constant C, the following equation is obtained.

D (j) = d (0) + (Σ _j b (i) −jT / M) (3) where d (j) = q (j + 1) / C (4) d (0) = Q (0) / C (5) In the expression (3), d (j) is conventionally called a buffer occupancy, and d (0) is an initial buffer occupancy and is empirically determined so that the buffer occupancy d (j) does not overflow or underflow. Can be

As described above, q (0) in the equation (5) is a quantization step of the macroblock 1, and is set to an appropriate value. Further, d (0) is empirically determined as described above. Therefore, the proportionality constant C is determined from equation (5). Usually, the quantization step q (i) is very small compared to the buffer occupancy d (i), so C is very small compared to 1. For example, d (0) = 10000 bits,
If q (0) = 20 bits, C becomes 1/500.

Since q (j + 1) is a quantization step, it itself takes discrete values. Therefore, if the value on the right side of Expression (2) is an intermediate value between the discrete values, the value is rounded off. That is, q (j + 1)
Will also be quantized. Now, the second on the right side of equation (2)
Assuming that the term is B (j), Δq (j + 1) = CΔB (j) (6) As is clear from C, C is very small compared with 1, so that ΔB (j) does not become a certain size. Δq (j + 1)
Cannot be greater than or equal to the interval m of discrete values of q (j + 1), and q (j + 1) takes an intermediate value between the discrete values, and the quantization step does not change. In such a case, several macroblocks are coded without changing the quantization step, and the quantization step is performed only when the difference between the generated information amount and the target information amount becomes larger than a certain size. Change, and the amount of generated information is controlled. Therefore, the quantization step does not change for each macroblock of an image but for every several macroblocks, and the amount of information generated in one screen as a whole approaches the target information amount.

[0010]

In the above-mentioned conventional image coding control method, the difference between the target information amount calculated for each macroblock and the generated information amount is smaller than the initial buffer occupation amount for each image. Therefore, there is a disadvantage that the characteristic of following the target information amount is not good. This is because when an image is divided into blocks, orthogonal transformation is performed in block order, coefficients are quantized, and variable-length coding is performed, even if the amount of generated information suddenly increases in a certain block, the quantization step is performed immediately. The problem is that the information amount does not increase and does not suddenly approach the target information amount. Therefore, when all blocks in an image are processed, the amount of generated information may greatly deviate from the target amount of information per image. An object of the present invention is to provide an image coding control method capable of reducing a difference between a target information amount and a generated information amount, and to improve coding efficiency.

[0011]

In order to solve the above-mentioned problems, an image encoding control method according to the present invention divides an image into spatial blocks and transforms and encodes the image by filter bank processing for each spatial block. For each macroblock consisting of a predetermined number of spatial blocks, the amount of encoded image information is defined as the amount of information generated for each macroblock, and the amount of information allocated to each macroblock is targeted. when the amount of information in the macroblock to be transform coding, in order to reduce the absolute value of the difference of the generated information quantity to the target information amount, to be coded
Encoding already completed before the macroblock
As proportional to the difference in the macroblock,
Determine the rate of change of the quantization step with respect to the difference, and
The quantization step of the first macroblock of the
Set as the initial value of the
And then accumulate the rate of change of the quantization step sequentially
Calculate the quantization step in the macroblock ,
An image coding control method for determining a quantization step for each macroblock, wherein the processing for determining the quantization step is performed by using the difference in a macroblock that has already been encoded before the macroblock to be encoded. When,
The quantization step occurs by taking discrete values,
In order to eliminate the degradation of the characteristic of following the target information amount, a product of the amplification factor and a variable positive number that adjusts the rate of change of the quantization step with respect to the difference is calculated, and the matrix is proportional to the product .
Determine the rate of change of the quantization step of the black block and quantize
Starting from the initial value of the step, successive quantization steps
Is calculated by accumulating the rate of change of
A step of calculating the step ; wherein the amplification coefficient is adaptively changed according to the magnitude of the difference.

[0012] The amplification coefficient is determined by the absolute value of the magnitude of the difference.
It can be changed proportionally .

[0013] The filter banking may include orthogonal transform coding on a spatial block basis of the image.

[0014] Filter banking can include encoding wavelet output coefficients for each spatial block.

[0015]

The macroblock number to be coded is j
+1, the macroblock number which has already been encoded is j, the quantization steps of macroblocks j and j + 1 are q (j) and q (j + 1), respectively, and the generated information amounts are b (j) and b, respectively. (J + 1). Assuming that the information amount of an image is T, the number of macroblocks included in one image is M, the amplification coefficient is a, and the proportionality constant is C, q (j + 1) is determined by the following equation. q (j + 1) = q (j) + Ca (b (j) −T / M) (7) Therefore, the rate of change of q (j + 1) with respect to the difference ε (j) = b (j) −T / M is It is given by the following equation.

Δq (j + 1) / Δε (j) = Ca (8) Therefore, if a is sufficiently larger than 1, the difference ε
Even with a slight change in (j), q (j + 1) has a discrete value interval m
And the generated information amount of the macro block j + 1 is controlled with high sensitivity.

By adaptively changing the amplification coefficient a in proportion to the difference ε (j), it is possible to efficiently follow a rapid change in the difference ε (j).

The present invention relates to a method for quantizing transform coefficients after conversion by a filter bank. Therefore, the transform method may be an orthogonal transform or a transform using a filter bank such as a wavelet. However, for high-efficiency coding, the quantization step q (j + 1) of the macroblock j + 1 is predicted using the difference of the previous macroblock j, that is, the previous value ε (j) of the spatial block difference information. The method is taken.
Therefore, any conversion means can be used as long as a conversion coefficient can be obtained for each unit (spatial block) of the image.

[0019]

Next, embodiments of the present invention will be described with reference to the drawings. Also in this embodiment, the image is divided into spatial blocks, and the spatial transform is orthogonally transformed by a filter bank for each spatial block. The output coefficients (orthogonal transform coefficients in this embodiment) of the filter bank are quantized and coded for each macroblock, as in the above-described conventional method.

This embodiment is different from the conventional method in that the amount of generated information b with respect to the target amount of information T / M of macroblock j is determined.
The product of the difference ε (j) of (j) and the amplification coefficient a is proportional to the change Δq (j + 1) of the quantization step q (j + 1). That is, q (j + 1) = q (j) + Ca (b (j) -T / M) (9) holds. Here, q (j + 1) = q (1) + [q (2) −q (1)] + [q (3) −q (2)] + ... + [Q (j + 1) − q (j)] = q ( 1) + Ca (Σ j b (i) -jT / M) (9a) where, q (j + 1) / C = e (j) (10) q (1) / If C = e (0) (11), the following equation is obtained from equation (9a) . When Ce (j + 1) = Ce (0) + Ca (Σ j b (i) -jT / M) (9b) equation both sides of (9b) is divided by C, and the following expression is obtained
You .

[0021] e (j) = e (0 ) + a (Σ j b (i) -jT / M) (12) where, a is a variable positive, e (j) buffer of the conventional method This is a virtual buffer state amount corresponding to the occupancy amount. Further, e (0) is the initial value of e (j), which is the initial buffer occupancy corresponding to d (0) in the conventional method.

FIG. 1 is a block diagram of an image coding control method according to the present embodiment. When encoding is completed up to the j-th macroblock, the target information amount 1 is jT /
M, and the generated information amount 2 is Σ _j b (i). These difference information amounts 3 are used to calculate the quantization step of the (j + 1) th macroblock. The difference information amount 3 is multiplied by the amplification coefficient 4 to obtain the amplified difference information amount 5. An initial buffer occupation amount 6 is added to the amplified difference information amount 5 to obtain a virtual buffer state amount 7. The quantization step 9 is calculated by multiplying the virtual buffer state quantity 7 by a proportional constant 8. This value is used for quantization of the (j + 1) th macroblock. The quantization step 9 is input to the encoding processing unit 10 for the (j + 1) th macroblock, and as a result, the generated information amount 11 of the macroblock is measured. The generated information amount up to the j-th macroblock stored in the memory 13 is added to the generated information amount 11, and is output as the generated information amount up to the (j + 1) th macroblock.
This generated information amount is stored in the memory 15 for encoding control of the next macroblock.

When the amplification coefficient a is large, the quantization step q (j + 1) changes sensitively to the buffer occupancy. Conversely, when the amplification coefficient a is small, the quantization step q (j + 1) is performed even if the buffer occupancy greatly changes.
The change is insensitive. In the most extreme case, the amplification factor is a =
In this case, the quantization step is determined only by the initial buffer occupancy e (0). In the present invention, since the object is to improve the ability to follow the target value, a
> 1. The amplification coefficient a is calculated by calculating the second
The difference information amount _j b (i) −jT / M of the term can be adaptively changed in proportion to the absolute value of the change. When the absolute value of the difference between the target information amount and the generated information amount is small, a = a (j) is a small value, and when the absolute value of the difference is large, the value of a (j) is set large. In this way, when the amount of generated information is close to the target information amount, the adjusting action for approaching the target value is weakened. Conversely, when the generated information amount greatly exceeds the target information amount or is much smaller, the adjusting action is reduced. Enhance.

FIG. 2 shows an experimental result of the present invention by computer simulation. In the experiment, the target bit rate per frame is set to 187 kbit for the moving image. When the amplification coefficient a is a = 1 (corresponding to the conventional method) and a =
When the amount of generated information is set to 16 (the present invention), it can be seen that when a = 16, the effect of approaching the target number of bits can be obtained in the first few frames.

As the filter bank processing, orthogonal transform in block units by DCT or the like can be performed. In addition, the present invention is applicable not only to orthogonal transform but also to orthogonal transform as long as transform coefficients to be quantized are grouped in arbitrary units of an image. Therefore, the present invention can also be applied to a case where arbitrary numbers of filter bank output data such as wavelets are collectively subjected to quantization. Also, in the present embodiment, as shown in Expression (9), the quantization step q (j +
1) is replaced by the quantization step q of the immediately preceding macroblock j.
Although the example is determined based on (j), it may be determined based on the quantization step of an arbitrary macroblock for which encoding has been completed.

[0026]

The present invention has the following effects.

(1) The difference value of the amount of generated information with respect to the target information amount is enlarged to a value larger than the actual value, and the quantization step is calculated. If the generated information amount is larger than the target value, the large quantization step is immediately performed. To reduce the amount of information generated,
Conversely, if the amount of generated information falls below the target value, a small quantization step is immediately provided, and the target value is quickly approached. As a result, when the encoding of one screen is completed, the amount of generated information per image does not greatly deviate from the target amount of information, and therefore, efficient encoding can be performed when the upper limit of the amount of generated information is defined. . That is, when encoding a still image in which the upper limit of the number of bits is determined, the target number of bits can be set to a value slightly smaller than the upper limit, and the amount of information can be sufficiently used. Further, when a moving image is encoded with a fixed number of bits, the fluctuation of the amount of generated information for each screen is reduced, and it can be realized with a smaller buffer.

(2) The effect of (1) can be stably and surely realized by adaptively changing the amplification coefficient in proportion to the absolute value of the difference information amount.

(3) According to the present invention, not only the orthogonal transform in block units but also the output of a filter bank such as a wavelet, as long as the transform coefficients to be quantized are grouped for each predetermined unit of the image. Since the present invention can be applied to quantization of coefficients, it can be applied to a wide range of applications, and the effects (1) and (2) can be realized in the wide range of applications.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of an image encoding control method according to the present invention.

FIG. 2 is a diagram showing experimental results of the present invention by computer simulation.

[Explanation of symbols]

1 Target information amount jT / M 2 Generated information amount Σ _j b (i) 3 Difference information amount 4 Amplification coefficient a 5 Amplified difference information amount 6 Initial buffer occupation amount e (0) 7 Virtual buffer state amount e (j) 8 Proportional constant C 9 Quantization step q (j + 1) 10 Encoding processing unit F (q) 11 Macroblock generation information amount b (j + 1) 12 Generation information amount up to jth macroblock 13, 15 Memory 14 j + 1th The amount of information generated up to the macro block

Continuation of the front page (56) References JP-A-4-137985 (JP, A) JP-A-3-286691 (JP, A) JP-A-3-16491 (JP, A) JP-A-7-107473 (JP) , A) JP-A-7-107479 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) H04N ^7/ 24-7/68

Claims (4)

(57) [Claims] 1. An image is divided into spatial blocks,
Converts the image by filter bank processing for each block
And a fixed number (M) of spatial blocks
The amount of encoded image information for each block (j).
The amount of information generated for each macroblock (b (j))
Target the amount of information allocated to each macroblock
When the amount of information (T / M) is
Generated information for the target information amount in block (j + 1)
Decrease the absolute value of the difference (b (j + 1)-(T / M))For
ToBefore the macroblock (j + 1) to be coded
Macro block (j) that has already been encoded.
In proportion to the difference (b (j)-(T / M))
The rate of change of the quantization step (q (j + 1) -q (j)),
Quantize the quantization step of the first macroblock in the image
Set the initial value of the quantization step (q (1))
Starting from the initial value of the
By accumulating the quantization rate, the quantization step
To calculateQuantization step for each macroblock (j + 1)
In the image coding control method that determines the q (j + 1)
The process for determining the steps is performed before the macroblock (j + 1) to be encoded.
For a macroblock (j) that has already been encoded
The difference (b (j)-(T / M)),The quantization step takes discrete values
Of follow-up to target information volume
To eliminateChange of quantization step for the difference
A variable positive amplification factor a that adjusts the rate (q (j + 1) -q (j))
With the product (a (b (j)-(T / M)))The quantization step of the macroblock in proportion to the product.
Rate of change (q (j + 1) -q (j)), Starting from the initial value of the quantization step (q (1)),
Next, the change rate of the quantization step is accumulated and the macro block is calculated.
Compute the quantization step in block (j + 1) processing
And adaptively changing the amplification coefficient according to the magnitude of the difference.
An image encoding control method, characterized in that: 2. The image encoding control method according to claim 1, wherein the amplification coefficient is changed in proportion to an absolute value of the magnitude of the difference. 3. The image encoding control method according to claim 1, wherein the filter bank processing includes orthogonal transform encoding in units of spatial blocks of the image. 4. The image encoding control method according to claim 1, wherein the filter bank processing includes encoding a wavelet output coefficient for each spatial block.