JPH0472870A - Picture encoding system - Google Patents

Abstract

PURPOSE:To reduce the code volume without degrading the picture quality by changing the adaptive method corresponding to a symbol for adaptive arithmetic encoding in accordance with local properties of a block. CONSTITUTION:A picture is divided to plural blocks, and suitable conversion is selected for each block, and a coefficient corresponding to the picture element value in the converted block is quantized, and a symbol corresponding to it is subjected to adaptive arithmetic encoding. The adaptive method corresponding to the symbol for this encoding is changed by local properties of the block. That is, the adaptive method is changed at the time of symbol adaptive arithmetic encoding by statistical properties of respective converted coefficients such as selected orthogonal conversion, the variance value of coefficients, the probability density distribution, and the appearance probability of each symbol. Thus, the picture is efficiently encoded, and the code volume is reduced without degrading the picture quality.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) This invention relates to an image coding method, and in particular, an image code that uses adaptive arithmetic coding in adaptive transform coding and performs variable length coding. Regarding the conversion method.

(Prior Art) As a high-efficiency image encoding method, an adaptive transform encoding method is generally used in which an image is divided into several blocks and an appropriate transformation is selected and performed for each block. Furthermore, variable length coding is known in which a transformed block is quantized and then coded using code words of different lengths according to symbols corresponding to the quantized values.

Variable-length coding includes Huffman coding, which encodes into a code that can be arbitrarily decoded when the probability of appearance of a symbol is known in advance, and further minimizes the amount of code. A method of constructing a Huffman code when the probability of appearance of a symbol is known in advance is described, for example, in Hideki Imai's ``Information Theory'' by Shokodo.

In Huffman encoding, there are two possible methods for determining code words. One is a method in which both the encoding means and the decoding means are provided with a codeword table in which one codeword is associated with one symbol, and this codeword table is used to perform encoding. . The other method is to check the appearance frequency of the symbol corresponding to the quantized value before encoding, and then send the code word to the decoding means using the Huffman code based on the appearance frequency or the appearance frequency. It is a method of transmission.

(Problem to be solved by the invention) However, in the method using a codeword table, the determined codeword may not be optimal for the image to be encoded, and in this case, the encoding efficiency deteriorates. There was a drawback to that. Furthermore, this method has a problem in that it is not possible to perform encoding suitable for local characteristics of the image, that is, characteristics that vary depending on the position in the image.

On the other hand, the method of transmitting appearance frequencies and code words has a problem in that encoding efficiency decreases due to the amount of information involved in this transmission. Furthermore, this method has problems such as an increase in processing time and a heavy burden on hardware due to the construction of code words.

The present invention was made in view of the above-mentioned conventional circumstances, and its purpose is to use adaptive arithmetic coding instead of Huffman coding to create a system that is suitable for the local characteristics of images. An object of the present invention is to provide an image encoding method that can perform encoding and reduce the amount of code without deteriorating the image quality.

[Configuration of the Invention (Means for Solving the Problem) In order to achieve the above object, the present invention divides an image into a plurality of blocks, and selects an appropriate transformation for each block. An image encoding method that quantizes coefficients corresponding to pixel values in the transformed block, and then adaptively arithmetic encodes symbols corresponding to the quantized values, the method comprising: The adaptive method corresponding to the symbol is configured to vary depending on the local properties of the block. Further, the local property is determined by the selected orthogonal transform, the coding order of the symbols, the variance value of the coefficient, the probability density distribution, and the quantization method used when quantizing the coefficient. The quantization value determines whether it is the probability distribution of the symbol, or the variance value of the coefficient, or the probability of occurrence of the symbol, and further configured to have a bit allocation amount corresponding to the coefficient. has been done.

(Operation) With the above configuration, the present invention divides an image into a plurality of blocks, and selects and performs an orthogonal transformation suitable for each block. The coefficients of the transformed block are quantized, and the symbols corresponding to the quantized values are encoded by adaptive arithmetic coding such as Qcog. In adaptive arithmetic coding of symbols, the adaptive method is changed depending on the statistical properties of each transformed coefficient, such as the selected orthogonal transformation, coefficient variance, probability density distribution, or probability of appearance of each symbol.

In addition, Q Korg is J, L, former tchell, etc.:
“Sortware implementation
s 01' the Q-Codcr", l13M J
Our own of Research and Dev
elopment, Vol. 32. No, 6. pp, 7
53-774. This Qcog is adaptive arithmetic coding for binary symbols, and by converting a multilevel symbol into a binary symbol, it brings about the encoding of a multilevel symbol.

(Example) Hereinafter, the present invention will be described in detail with reference to the drawings.

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

In the figure, a block division circuit 1 divides an input image into blocks of N pixels by N pixels and outputs the divided blocks. In addition, block division circuit 1
is a signal 201 indicating that image encoding has been completed.
This outputs the following.

The transformation selection circuit 3 selects a suitable transformation for each divided block, and uses an index k (k-
1, 2, ..., K).

The encoding circuit 5 encodes the index k into a uniquely decodable code and outputs the encoded code.

The transformation matrix memory 7 outputs the transformation matrix T corresponding to index 1 according to index k.

The orthogonal transformation circuit 9 orthogonally transforms the block output from the block division circuit 1 based on the transformation matrix T, and outputs the orthogonally transformed block.

[quantization circuit]] is each coefficient C of the orthogonally transformed block.
(i, j), (i, j=1, 2..., N) are quantized, and the symbols A(ij), (1, J
=1,2. ..., N) is output.

One-dimensional circuit] 3 converts symbols A(i, j) into symbols a(1) a(2), . . . according to a certain predetermined order.
・, a (L). Also,
One-dimensionalization circuit] 3 outputs (i, j) indicating the position of the symbol within the block.

The arithmetic encoding circuit 15 is the central part of the present invention, and performs adaptive arithmetic encoding on the symbols output from the one-dimensionalization circuit 13 using a configuration using a Qcog, which will be described later. Further, the arithmetic encoding circuit 15 outputs codes subjected to adaptive arithmetic encoding.

A MUX (multiplexer) 17 multiplexes the code in the index and the code output from the arithmetic coding circuit 15, and outputs the multiplexed code to a decoding means (not shown).

FIG. 2 is a block diagram showing details of one embodiment of the arithmetic coding circuit 15 using Qcog.

The standard deviation memory 101 stores the standard deviation corresponding to the variance value of the symbol based on the index I output from the conversion selection circuit 3 (and the position (ij) in the block of the symbol output from the one-dimensionalization circuit 13). σh (i, j)
is determined and output.

The appearance probability memory 103 stores the symbol output from the one-dimensional circuit 13 as the last symbol (a
The probability p that poB/End Of Block) is
k (i, j) is determined and output.

The binarization circuit 105 converts the symbols output from the one-dimensionalization circuit 13 into one or more binary symbols,
This is where it will be output. The binarization circuit 105 also transmits d, which will be described later, to the Q index selection circuit 107.

[Q index selection circuit] 07, the binarized symbol b is a p, and the symbol a is a p. If it is informed that there is no symbol indicating whether the value is 3 or not, an index S classified into classes is determined and output based on the standard deviation σh (i+J) output from the standard deviation memory 101. Further, when the Q index selection circuit 107 is informed that the binarized symbol b is a symbol indicating whether or not a is a E (18), the Q index selection circuit 107 selects the probability p * ( t + j) is determined and output.

The Q index memory 109 stores M types of Q indexes, and stores the Q index q corresponding to the index S output from the Q index selection circuit 07,
This outputs the following.

The MPS estimated value memory 1]1 stores M types (7) MPS estimated values, and the Q index selection circuit]0
The MPS estimated value m corresponding to the index S output from 7 is output.

[Q coder] 13 uses the Q index q6 outputted from the Q index fumori 1.09 and the MPS estimated value m6 outputted from the MPS estimated value memory 111, and uses the 2 This is where the value symbols are encoded. In addition, the Q coder]]3 is a symbol MPS with a high probability of appearance and a symbol LPS with a low probability of appearance.
The Q value determined by the probability of appearance of a symbol is used to perform encoding adapted to the probability of appearance of a symbol.

One embodiment of the present invention is thus constructed, and will now be described in detail.

The input image is divided by the block division circuit 1 into blocks of N laterals×N pixels. For the divided blocks, a transformation selection circuit 3 selects a transformation suitable for this block. Furthermore, the conversion selection circuit 3 outputs an index 1 (indicating the selected conversion).

Here, the conversion selection method in the conversion selection circuit 3 is described in Akamine et al.: "Adaptive KL Transform Coding Method with Adaptive Bit Allocation" Proceedings of the 1989 Institute of Electronics, Information and Communication Engineers Autumn Conference D-42. . According to this, a representative autocorrelation matrix may be created in advance, and the determination may be performed based on the distance between this representative autocorrelation matrix and the manually created autocorrelation matrix of the block. This representative autocorrelation matrix is created by creating an autocorrelation matrix for a set of blocks from the vectors included in each block, grouping them using the LBG algorithm, and then A representative autocorrelation matrix may be obtained and created for each case.

Note that as a set of blocks, a set of blocks obtained by dividing several test images by the same means as the block dividing circuit 1 may be used.

The index l( outputted from the conversion selection circuit 3) is encoded by the encoding circuit 5. The encoding circuit 5 performs fixed-length encoding with the same code length for all transformation indices. Alternatively, variable-length encoding may be performed using a variable-length code table created in advance.Alternatively, for all blocks of the image to be encoded, the selected transformation may be checked, and the transformation A Huffman code may be created according to the appearance frequency of , and before the index k is encoded and transmitted to the decoding means, the correspondence between the conversion index and the created Huffman code may be transmitted.

In addition, the index l( outputted from the conversion selection circuit 3 is also input to the conversion matrix memory 7, and this index 1(
A transformation matrix T corresponding to is output from the transformation matrix memory 7.

On the other hand, the block output from the block division circuit 1 is
Orthogonal transformation circuit 9 performs orthogonal transformation based on transformation matrix Tk. Coefficients of the transformed block C(i,j)
is quantized by the quantization circuit 11, and a symbol A(i, j) corresponding to this quantized value is output. The quantization in the quantization circuit 11 may be linear quantization or some kind of nonlinear quantization. Further, this quantization may be performed differently depending on the index and (i, j), or the same quantization may be performed. For example, only C(1,1,) may be quantized with a smaller quantization width than other coefficients.

Symbol A(i.

j) are the symbols a(1,), a(2)...
・, is output as a (L). Figure 3 shows (i,
This is an example of a table that associates j) with the order of output from the one-dimensionalization circuit 13. The same order may be used for all blocks. Also, a different order may be used depending on the index 1 selected by the conversion selection circuit 3. In this case, the order may be determined by outputting the symbol corresponding to the coefficient with a large variance value first. .

FIGS. 4(A) and 4(B) show an example of the order in which the output is performed in descending order of the variance value. Fourth
Figure (A) shows the variance value of each coefficient, and Figure 4 (B) shows (i, j
) and the output order] 3.

Note that the symbols a (1,), a (2), ・
..., a(L) may include all of the symbols A(i, j) input to the one-dimensional circuit 13, or the number may be determined in advance and only that number may be included. You can also do this. That is, if the number is L in advance, from the one-dimensional circuit]3, the first to Lth symbols a (
1,) , a (2) , . . . , a(L) may be output. Here, the number may be the same for all blocks, or may be different depending on the selected transformation.

Also, L8゜8th and subsequent symbols all have a quantization value of 0.
In the case of symbols corresponding to
HO2 symbols a (1) , a (2) ,
-a (LEOll 1) and only aEoB may be output.

The symbols output from the one-dimensional conversion circuit 13 are input to the binarization circuit 105 and the Q index selection circuit 107 in the arithmetic coding circuit 15. Further, (i, j) indicating the position of the symbol within the block is stored in the standard deviation memory 101,
Appearance probability memory 103 and Q index selection circuit 1
07 is input.

In addition, the index l outputted from the conversion selection circuit 3 is manually stored in the standard deviation memory 10 and the appearance probability memory 103.

In the standard deviation memory 101, the index 1 (and the position (l) of the symbol to be encoded within the block) are stored.

j), the standard deviation σh (i, j) corresponding to the variance value of that symbol is determined. The determined standard deviation σ* (t, j) is then output to the Q index selection circuit 107.

In the appearance probability memory 103, when the symbol a to be encoded is the first symbol a(1) output from the one-dimensionalization circuit 13 in the block, the probability that it is a EOB is p(a (1,) =aEOB) is output to the Q index selection circuit 1.07. In addition, if the symbol a to be encoded is the second or subsequent symbol, the r previous symbol a
is not a symbol corresponding to a quantization value of 0, the conditional probability that the symbol a is a EOB is output to the Q index selection circuit 107.

Symbol a input to the binarization circuit 105 is binarized and output to the Q kneader 113 as a binary symbol b.

Here, the symbol a may be binarized by expressing the quantized value represented by the symbol a as a fixed-point number and arranging the quantized values from the most significant bits. Alternatively, binarization may be performed using the flowcharts shown in FIGS. 5(A) to 5(F). However, the fifth
In the figure, r3 log, dbjt, and D are constants.

Note that in the binarization circuit 105, the multilevel symbol a is converted into one or a plurality of binary symbols. At that time, when a binary symbol indicating whether or not a is EOB is output, it is set as d-D, and otherwise, the output binary symbol corresponds to one multilevel symbol and is output as d-D. Let d be the position indicating which symbol among the value symbols. This d is then transmitted to the Q index selection circuit 1.07.

In the Q index selection circuit 107, if the transmitted d is not D, an index S classified into classes is determined based on the standard deviation σk (i, j) outputted from the standard deviation memory 1.01. Furthermore, when d is D, the probability pi (1, J) output from the appearance probability memory 103
The classified index S is determined by.

Here, an example of the method for determining the index S in the Q index selection circuit]07 is shown in FIG. In the same figure,
This is the case where b is a symbol corresponding to BOB judgment. Note that σ and 8 are σ.

The maximum value of (i, j), i.e. (Immx = m a X ” U* (1, j)
1. In the example shown in FIG. 6, when d is not D, classes are classified based on the coefficient standard deviation σ* (i, j), that is, the variance value, and when d is D, EOB
Classification is performed according to the conditional probability p(i, j).

Further, another example of the method for determining the index S is shown in FIG. 7.8.9. However, in FIG. 8.9, σ□ is an arbitrarily determined positive number, which may be determined as the standard deviation of the average for all transformations of each sequence.

The index S classified in this way is output from the Q index selection circuit]07 and input to the Q index memory]09 and the MPS estimated value memory]1.

In the Q index memory 109 to which the index S is input, the Q index q corresponding to this index S is
, is selected from M types and output to the Q coder]]3. In the Q index memory 109, each time the encoding of one binary symbol is completed, if the Q index is updated, the updated Q index is stored as q.

Also, the MPS estimated value memory 1 to which the index S is input
11, the MPS estimated value m corresponding to this index S is selected from among M types and is output to the Q coder]]3. Even in this MPS estimated value memory 11, if the MPS estimated value is updated each time encoding of one binary symbol is completed, the updated MPS estimated value is m6
It is stored as .

The binary symbol b output to the Q kneader 113 is encoded by the Q kneader 113. At this time, Q kneader 11
3, the Q value determined by the Q index q8 and the MPS estimated value m are used.

Here, the Q value is selected from a Q table as shown in FIG. 10, for example. Current Q value (Qe Val
ue) indicates which value in the Q table it is and is determined by the Q index (Qe 1ndeX).

The Q index and the MPS estimated value are adaptively updated in accordance with normalization processing during encoding.

FIGS. 11(A) to 11(D) show an example of an initialization method and an updating method for the Q index, Q value, and MPS estimated value.

In addition, FIG. 12 shows the input symbols, output symbols, and
and an example of the Q value.

The code encoded by the Q kneader 113 is MUXl
, 7, and is multiplexed with the code output from the encoding circuit 5.

When the encoding of the image is completed, the signal 20 outputted from the block dividing circuit is input to the Q kneader 113, and thereby the encoding completion process is performed.

In this way, since the probability of appearance of a symbol differs depending on the statistical properties of the coefficients caused by the local properties of the image, encoding efficiency can be improved by using different Q indexes and MPS estimates accordingly. In this example, the symbols to be encoded are divided into several classes according to the variance value or appearance probability, and the encoding efficiency is improved by using different Q indexes and MPS estimation values for each class. .

Note that in this embodiment, the adaptation method is changed depending on the variance value or the appearance probability, but the invention is not limited to this. For example, by deleting the appearance probability memory 103 and leaving only the standard deviation memory 101, it is possible to change the adaptation method depending only on the variance value. The method for determining the index S in the current Q index selection circuit 107 is shown in FIG. 13.1.4.

In addition, standard deviation memory] 01 and appearance probability memory 10
3 and directly inputting the index k to the Q index selection circuit 1-07, the adaptation method can be changed depending on the selected orthogonal transform and the order in which the symbols are encoded.

The method for determining the index S at this time is as shown in FIG.

Furthermore, standard deviation memory 101 and appearance probability memory 1
By deleting 03 and replacing them with a probability distribution memory that stores multiple occurrence probability distributions p-(i, j, d), the adaptation method can be changed depending on the symbol occurrence probability distribution. . FIG. 16 shows the method for determining the index S in this case.

Furthermore, by adding a probability distribution memory to the standard deviation memory 101 and the occurrence probability memory 103 and storing a plurality of occurrence probability distributions 2 p (i, j, d), the adaptive method can be adapted according to the variance value. It is possible to determine whether to change or change depending on the appearance probability based on the positions i and j of the coefficient to be encoded within the block and the quantization value. The method for determining the index S at this time is as shown in FIG.

Finally, standard deviation memory 10] and appearance probability memory 1
03 are deleted together, and multiple bit allocation amounts gb (i,
By arranging a bit allocation amount memory that stores j, d), it becomes possible to change the adaptation method depending on the bit allocation amount of the symbol. The method for determining the index S at this time is shown in flowchart 1 in FIG. In addition, in FIG. 18, g III N
t is gb(i+j.

d) is the maximum value.

[Effects of the Invention] As described above, according to the present invention, by using adaptive arithmetic coding, it is possible to perform coding suitable for the local characteristics of an image. Therefore, images can be encoded efficiently, and the amount of code can be reduced without deteriorating the image quality. Furthermore, it is possible to reduce the encoding processing time and reduce the burden on hardware.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the image encoding system of the present invention, FIG. 2 is a block diagram showing details of the arithmetic encoding circuit shown in FIG. 1, and FIGS. Tables, flowcharts, or Q tables for explaining one embodiment of the invention, FIGS. 13 to 18 are flowcharts 1 for explaining the invention in detail. 15... Arithmetic initialization circuit 10]... Standard deviation memory 103... Appearance probability memory] 05... Binarization circuit] 07... Q index selection circuit 109... Q index memory 111...・MPS estimated value memory 113...Qcog

Claims (7)

[Claims] (1) Divide the image into multiple blocks, convert the divided blocks by selecting an appropriate transformation for each block,
An image coding method that quantizes coefficients corresponding to pixel values in a transformed block and then adaptively arithmetic encodes symbols corresponding to the quantized values, the method comprising: An image coding method characterized in that a corresponding adaptation method is varied depending on the local properties of the block. (2) The image encoding method according to claim 1, wherein the local property is the selected orthogonal transform and the order in which the symbols are encoded. (3) The image encoding method according to claim 1, wherein the local property is a variance value of the coefficient. (4) The local property is an appearance probability distribution of the symbol determined by a probability density distribution of the coefficient and a quantization method used to quantize the coefficient. 1) Image encoding method described. (5) Claim (5) characterized in that the local property is a variance value of the coefficient or a probability of appearance of the symbol.
1) Image encoding method described. (6) The image encoding method according to claim (5), wherein whether the local property is a variance value of the coefficient or an appearance probability of the symbol is determined by the quantization value. (7) The image encoding method according to claim (1), wherein the local property is a bit allocation amount for the coefficient.