JPH0472960A - Picture coding system - Google Patents

Abstract

PURPOSE:To obtain a reproduced picture with high picture quality by discriminating totally a code quantity and picture quality so as to select optimum conversion thereby reducing the code quantity more than that of the adaptive KL conversion coding system employing a variable length code in a conventional system. CONSTITUTION:An input picture is divided into NXN rectangles by a block division circuit 101, a conversion matrix T1 is called by a conversion matrix memory 106, an orthogonal transformation circuit 102 applies orthogonal transformation to the block and a quantization circuit 103 applies quantization to the result. A transformation coefficient quantized is coded by using a coding table 108 at a coding circuit 104, subject to inverse quantity by an inverse quantization circuit 107 and an error calculation circuit 109 calculates a quantization error. The quantization error and the code quantity are fed to an evaluation calculation circuit 110, in which an evaluation value R1 is calculated, Then a matrix T2 is called from the conversion matrix memory 106 to apply similar processing and to calculate the evaluation value R2. In the case of R1>R2, the code and the index stored in the memory 105 and the index memory 112 are replaced into those by the T2.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to an image encoding system, and particularly to improvements in encoding technology when performing adaptive orthogonal transformation and variable length encoding.

(Prior art) Generally, in a method for compressing and encoding an image and transmitting it,
The method of dividing an input signal into several small units (called blocks), performing orthogonal transformation, and encoding the transform coefficients is called orthogonal transform encoding.

In particular, when the orthogonal transformation is one adapted to the input source, K
Karhunen-Loevetransform
rm). The principle of KL conversion will be explained below.

When the autocorrelation matrix generated from the source of a certain input block is written as RXX, this is a symmetric matrix, so it can be expressed as an orthonormal matrix and a diagonal matrix as follows.

Here, T is an orthonormal matrix consisting of eigenvectors corresponding to eigenvalues of RXX, and e□, . . . e8 are eigenvalues of RXX, the average of which is equal to the variance of the input block source. Here, e1≧e2≧...≧e5 (
S is the number of elements in the block). Define T as the KL transformation corresponding to this input source.

KL transformation is said to be the most efficient transformation method in principle, but calculating it for each input source requires a huge amount of calculations and is difficult to implement, so it is necessary to prepare multiple transformations in advance and apply them according to the input block. A common method is to switch. This is called adaptive KL transformation. FIG. 7 shows an example of the adaptive KL transformation. This was organized by the Picture Coating Symposium Japan 88 (PC8J88); co-sponsored by the Image Engineering Research Committee of the Institute of Electronics, Information and Communication Engineers;
As described in ``Adaptive KL Transform Adaptive Coding Method J (7-11) (held from September 30, 1988 to October 2, 1988), the input image is divided into horizontal 8 vertical 8 two-dimensional blocks by the block division circuit 74. is divided into 75 for each transformation T
1. . . , calculate the amount of code when performing quantization and variable length coding using T, and select the index i of the transformation that minimizes the amount of code.

According to this index i, the KL transform T process is called from 76, and orthogonal transform circuit 77 performs orthogonal transform using Ti. All the transformed blocks are quantized (78) and Huffman encoded (79) using the same method and then transmitted.

In this way, methods for selecting the optimal transformation for each input block include, in addition to the method that minimizes the amount of code as described above, the Euclidean method of selecting the autocorrelation matrix of the block and the autocorrelation matrix that is the source of the transformation. Those that minimize the distance (for example, Japanese Patent Application No. 63-223015) and those that minimize the quantization error during decoding (for example, Japanese Patent Application No. 63-22301)
Publication No. 5) etc.

However, these methods cannot take into account both the code amount and quantization error, which are important requirements during encoding, and therefore select a KL transform that is not optimal for either of the two, resulting in poor decoding. There was a problem in that the quality of the image deteriorated.

(Problem to be Solved by the Invention) As described above, in the adaptive KL transform coding method using variable length coding, the conventional selection method did not consider both the amount of code and the image quality, so a suboptimal transform was selected. There is a problem that this may result in an increase in the amount of code or deterioration of the decoded image.

The present invention was made to improve the above problem, and examines both the code amount and quantization error for each input block, compares the performance for both with a certain weight, and if the code amount is the same, the quantization error This is an encoding method that optimizes both the amount of code and image quality by making it possible to select a conversion with a small amount of code, and a conversion with a small amount of code if there is no difference in quantization error.

[Structure of the invention]

(Means for Solving the Problem) When selecting a plurality of orthogonal determinants to be used in orthogonal transformation applied to each block during encoding, the present invention evaluates the image quality of the block when each orthogonal determinant is used. A circuit is added to check the code amount of the block when it is encoded using the determined quantization error and its transformation. Through this, the effect on the amount of code and the effect on image quality when using the orthogonal determinant is investigated, and the orthogonal determinant is selected by comprehensively evaluating both.

(Operation) In the image encoding method according to the present invention, first, the screen is
It is divided into blocks of X 8 , 16X16, etc., and then each block is transformed using each of orthogonal determinants prepared in advance. Quantize the transformed blocks and calculate the quantization error for each transform. It also encodes the quantized block and calculates the amount of code. The quantization error and code amount are added together with a certain weight, the transform with the smallest value is selected, and the block encoded with this is transmitted. The conversion number (index) is also separately encoded and transmitted.

By selecting an orthogonal determinant based on both the code amount and quantization error, it is possible to obtain a reproduced image with better overall image quality and code amount than the encoding method using the conventional selection method. .

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

FIG. 1 is a block diagram showing a first embodiment to which the encoding method of the present invention is applied.

In the transformation matrix Tkk=□l-1, an encoding table and an encoding table for input files are created in advance and stored in the transformation matrix memory 10.
6. Encoding table memory 108. It is prepared in the encoding table memory 111. Here, K is the number of conversions, for example, K = 256. Figure 2 shows the transformation matrix.

3 is a flowchart outlining a method for creating an encoding table and an index encoding table. Alternatively, existing conversions or tables may be used. Note that the index encoding table can also be created based on the frequency of use of each conversion of the input image. In this case, the created table is transmitted to the receiving side.

The encoding procedure will be explained below. The input image is divided into NXN rectangles by the block division circuit 101. For example, N is 8. Next, the transformation matrix T1 is called from the transformation matrix memory 106, orthogonal transformation is performed on the block in the orthogonal transformation circuit 102, and quantization is performed in the quantization circuit 103. The quantization is linear quantization with a step size of 16, for example. The quantized transform coefficients are sent to the encoding circuit 104.
, the data is encoded using the encoding table 108 , inverse quantization is performed in the inverse quantization circuit 107 , and a quantization error is calculated in the error calculation circuit 109 . The quantization error and code amount are sent to the evaluation value calculation circuit 110, where a value R1 is calculated using the following equation.

Here, λ is a weighting coefficient. For example, let λ=0.7.

Here, quantization error≠ is just an example, and any measure that can evaluate the image quality during playback may be used. For example, it may be an error calculated in consideration of visual characteristics. The "code amount" may also be, for example, entropy, as long as it indicates efficiency during transmission.

The encoded block is stored in the memory 105 and the index 1 of the transformation is stored in the index memory 112. Further, T2 is called from the transformation matrix memory 106 and similar processing is performed to calculate R2. If R□≦R2, move on to the next conversion T3, and if R1) R2, rewrite the code and index stored in the memory 105 and index memory 112 to those according to T2, change the value of R1 to the value of R2, and perform the next conversion. Moving on to conversion T3. After repeating this process up to T, the code stored in the memory 105 is transmitted, the index stored in the index memory 112 is encoded and transmitted using the encoding table 111, and the next block is processed. Move to. Alternatively, if the index encoding table is designed according to the input image, the index is encoded and transmitted after the input image has been encoded.

The decoding procedure will be explained below. The received code is decoded in decoding circuit 114 using encoding table 118 . The encoding table 118 is the encoding table 10
Same as 8. Furthermore, the inverse quantization circuit 115 dequantizes the code, and the inverse transformation circuit 116 converts the separately received index code ■ into the transformation matrix memory 120 using the value i decoded in the decoding circuit 121 using the encoding table 119. The transformation T1 is called to perform inverse transformation, and the block composition circuit 117 restores the image to an output. Here, the encoding table 119 is encoding table 1.
It is the same as 11. Transformation matrix memory 120 is identical to transformation matrix memory 106.

The theory of creating KL transformations and the theory of creating Huffman codes are well known, for example, as described in "Digital Signal Processing of Images" by Takahiko Fukinuki, published by Nikkan Kogyo Shinbunsha.

The step size of the quantum machine can be changed depending on the amount of information by examining the properties of the input image.

In this case, it is necessary to transmit the step size to the receiving side as additional information. The quantizer does not have to be a linear quantizer, nor does it need to use the same step size for all transform coefficients; for example, the step size for low-frequency coefficients can be set smaller than 16, and the step size for high-frequency coefficients can be set larger than 16. You can also.

FIG. 3 is a block diagram showing a second embodiment to which the encoding method of the present invention is applied.

In Figure 3, 301 to 321 are 101 to 12 in Figure 1.
It has the same function as 1. Quantizer Qi (j=1...,
L) is set in plurality, for example, four, and is stored in the quantizer memory 322 in advance.

The encoding procedure will be explained below. The input image is divided into NXN rectangles by a block division circuit 301. For example, N is 8. Next, in the quantizer selection circuit 323, a quantity representing the amount of information of the block, for example, an intra-block variance value, is calculated, and a quantizer for quantizing the block is selected based on this value. For example, quantizer Q□,...
l Q 4, quantization step size 8, 16 .
32.64 linear quantizer, if the variance value of the block is O~1000, then Q,,1.001~2
If it is 000 then Q2. If it is 2001~3000 then Q3.3
If it is 001 or more, select Q4. The index J of the selected quantizer is transmitted in 2 bits. Further, a quantizer is called from the quantizer memory 322 according to the index,
This is used during quantization in the quantizer 303. After selecting a quantizer, the transforms are sequentially retrieved from the transform matrix memory 306, and the optimal transform is selected and encoded and transmitted. This procedure is the same as the first embodiment, so the explanation will be omitted.

The decoding procedure will be explained below. Quantizer memory 324
is the same as 322. The received quantizer index value J is sent to the quantizer memory 324, and the quantizer Q, y
is called, and QJ is used during inverse quantization in the inverse quantization circuit 315. The rest of the decoding procedure is the same as in the first embodiment, so a description thereof will be omitted.

By adding a quantizer selection circuit in this manner, it is possible to perform quantization according to the information amount of the block.

Note that the quantizer does not need to use the same step size for all transform coefficients, and may vary, such as using a small step size for low-frequency coefficients and a large step size for high-frequency coefficients. It is also possible to use a nonlinear quantizer.

FIG. 4 is a block diagram showing a third embodiment to which the encoding method of the present invention is applied.

In Figure 4, 401 to 421 are 101 to 12 in Figure 1.
It has the same function as 1. For example, three quantizers Qj je□+-+L (L=3) are set and stored in the quantizer memory 422 in advance.

The encoding procedure will be explained below. The input image is divided into N×N rectangles by the block dividing circuit 401.N is, for example, 8. Next, similarly to the first embodiment, the transformations are sequentially recalled from the transformation matrix memory 406, and the orthogonal transformation circuit 402
Perform orthogonal transformation at .

Here, the index i of the current transformation T is sent to the quantizer selection circuit 423. The quantized coefficients of the previous block are also stored here, and the quantizer for the current transformation of the block is selected based on the index of the current transformation and the coefficients of the previous block. For example, quantizers Q, , Q2 . If Q3 is equipped with a linear quantizer with a quantization step size of X8.16.32, ■ If the index of the current transformation is 1, Q
1 ■ If the index of the current transformation is not 1 and the number of non-zero coefficients in the previous block is 2 or less Q2 ■ If the index of the current transformation is not 1 and the number of non-zero coefficients in the previous block is 3 or more Q3 Choose a quantizer. The selected quantizer is called from the quantizer memory 422 and used during quantization in the quantizer 463. The rest of the encoding procedure is the same as in the first embodiment, so a description thereof will be omitted.

The decoding procedure will be explained below. Quantizer selection circuit 425
is the same as 423 and stores the quantized coefficient value of the previous block. The index of the decoded transformation is added to this, and a quantizer is selected in the same way as during encoding. The quantizer memory 424 is identical to 422, from which the selected quantizer is called and used during dequantization in the dequantization circuit 415. The rest of the decoding procedure is the same as in the first embodiment, so a description thereof will be omitted.

By selecting a quantizer in this manner, quantization can be changed depending on the transform used or with reference to the pre-encoding result. The information required when selecting a quantizer may be only the index of the current transform, or may use the encoding results of not only the previous block but also past blocks. By using this information, the quantization method can be changed without increasing additional information.

FIG. 5 is a block diagram showing a fourth embodiment to which the encoding method of the present invention is applied.

In Figure 5, 501 to 521 are 101 to 12 in Figure 1.
It has the same function as 1. Quantizer Q J+ j =□1−
For example, three ILs (L=3) are set and the quantizer memory 52
Prepare in advance for step 2.

The encoding procedure will be explained below. The input image is divided into blocks as in the first embodiment, and orthogonal transformation is performed by sequentially recalling transformations from the transformation matrix memory 506. Here, the index i of the current transformation T is sent to the quantizer selection circuit 523.

523 also stores the index of the transformation selected in the previous block, that is, the block to the left or immediately above that has already been encoded, and based on the index of the current transformation and the index of the previous block, the index of the transformation of the block is stored. The quantizer for the current transform is selected. For example, quantizer Q1. Q2. Q3
If a linear quantizer with a quantization step size of 8 and 16.32 is provided, ■If the index of the current transform is 1, then Q1. ■If the index of the current transform is not 1 and any of the indices of the previous block is 1. In the case of Q2 ■ If the index of the current transformation is not 1 and none of the indices of the previous block is 1, select a quantizer as in Q3. The rest of the encoding procedure is the same as in the third embodiment, so a description thereof will be omitted.

The decoding procedure will be explained below. Quantizer selection circuit 525
is the same as 523 and stores the conversion index of the previous block. The index of the decoded block is added to this, and a quantizer is selected in the same way as during encoding.

Other decoding procedures are the same as in the third embodiment.

FIG. 6 is a block diagram showing a fifth embodiment to which the encoding method of the present invention is applied.

For the transformation matrix T k k=1+-+, an encoding table and an index encoding table are created in advance and stored in the transformation matrix memory 6.
10, encoding table memory 604. The encoding table memory 615 is prepared. Here, K is the number of conversions, for example, =256. The transformation matrix, encoding table, and index encoding table can be prepared in the same manner as in the first embodiment.

The encoding procedure will be explained below. The input image is divided into NXN rectangles by a block division circuit 601. N=8
shall be. The blocks are sequentially stored in the frame memory 605, and the information amount calculation circuit 602 calculates a quantity representing the amount of information of the block, for example, an intra-block variance value. After the amount of information has been calculated for all blocks in this way, the planned bit allocation amount for each block is determined. This means, for example, that the total number of blocks is M
.

The information amount is σJ k=1+-++4+planned total code amount is X,
When the planned bit allocation amount is XJJ=□r−+M,
Define as follows. Here, X'/M is an amount allocated in advance, and a fixed amount is given for the code for the (1, 1) component and the E OB (end of block) code that are absolutely necessary for every block. For example, X'/M is 10
Bit.

The planned bit allocation amount for each block determined in this way is stored in the pill allocation amount memory 603.

After the planned bit allocation is determined, the frame memory blocks are sequentially called and the transformation T is first read from the transformation matrix memory 610.
1 and performs orthogonal transformation in orthogonal transformation circuit 606. Next, quantization is performed in a quantization circuit 607. The quantizer is, for example, linear quantizer with a quantization step size of 16. Note that this quantizer can also be switched from among a plurality of quantizers as in the second to fourth embodiments. The quantized transform coefficients are encoded in the encoding circuit 608 with reference to the encoded daple 604 . At this time, the code amount of the block is determined by the bit allocation amount memory 603.
If it exceeds the planned bit allocation amount of the block called out by adding the surplus bit amount at the time of encoding, the quantized transform coefficient is sent to the significant coefficient reduction circuit 611, where the significant coefficient That is, one non-zero coefficient is changed to zero. For example, the non-zero coefficient with the highest frequency or the non-zero coefficient with the smallest absolute value may be selected as the non-zero coefficient.

After reducing the significant coefficients in this way, the encoding circuit 60
8, the reduction is repeated until the code amount becomes equal to or less than the sum of the planned bit allocation amount and the surplus bit amount at the time of encoding. When the code amount is less than the planned bit allocation amount, the dequantization circuit 612 performs dequantization of the transform coefficient, and the error calculation circuit 614 calculates the quantization error of the block.

The quantization error is sent to the evaluation value calculation circuit 613, and R1 is calculated using equation (1) together with the code amount.

Note that the difference between the code amount and the amount obtained by adding the surplus bit amount at the time of previous encoding to the planned bit allocation amount is stored in the encoding circuit 608 as the surplus bit amount.

The encoded block is stored in the memory 609 and the index 1 of the transformation is stored in the index memory 616.

Next, T2 is called from the transformation matrix memory 610 and similar processing is performed to obtain R2. The following processing is the same as that in the first embodiment, so the explanation will be omitted.

The decoding procedure is the same as in the first embodiment, so its explanation will be omitted.

By determining the bit allocation amount to be allocated to each block in advance in this way, fixed rate encoding can be performed without exceeding the planned total code amount.

In the above embodiments, the transform coefficient encoding circuit is written in a manner that does not depend on the transform index, but it can also be switched depending on the transform index. It is also possible to switch based on information other than the conversion index. In this case, additional information about the encoding method is required.

〔Effect of the invention〕

As explained above, with the image encoding method according to the present invention, it is possible to comprehensively judge the code amount and image quality and select the optimal transformation. It is possible to reduce the amount of code compared to the encoding method, and to obtain high-quality reproduced images.

FIG. 2 is a block diagram showing embodiments 1.2, 3, and 4.5 related to the encoding method, and FIG. 2 shows an example of creating a transformation matrix, encoding table, and index encoding table in the first embodiment. FIG. 7 is a block diagram for explaining the conventional image encoding method.

Agent: Patent Attorney Noriyuki Chika

Claims (1)

[Claims] (1) In an image encoding method in which an image is divided into blocks and subjected to orthogonal transformation and then variable encoded, a plurality of orthogonal determinants used in the orthogonal transformation are prepared in advance, and each orthogonal determinant is used to perform orthogonal transformation and An image encoding system characterized in that it has means for checking the code amount and image quality during encoding, and selects the orthogonal determinant based on the code amount and image quality determined by the means.