JPH01198185A - Picture data compressing system - Google Patents

Abstract

PURPOSE:To prevent the generation of a block boundary in a decoded picture, and besides, to enable the compression of a high compression rate capable of obtaining high picture quality by performing frame/frame conversion to orthogonal-transform a whole picture at a time without dividing the picture into blocks. CONSTITUTION:The whole picture is orthogonal-transformed at a time, and variable bit length is assigned to each transformed component in proportion to the logarithmic value of the absolute value of the transformed component. The respective transformed components are compressed and encoded according to a number of assigned bits. Here, a bit assignment table to be additional information is high-efficiency-encoded by, for instance, Quad-tree expression. For instance, picture data which has the luminance value of two-dimensional arrangement and is inputted from a picture input device 81 is orthogonal- transformed by a cosine transformer 82, and is stored once in a buffer 83. It is encoded by a bit assignment table encoder 84. The transformed component is obtained through a reverse quantizer 91, and is stored in the buffer 92, and the luminance value of the decoded picture is obtained by a reverse cosine transformer 93, and is outputted to a picture output device 94.

Description

Detailed Description of the Invention [Field of Industrial Application] The present invention relates to highly efficient encoding of image data, and in particular to an image data compression method suitable for achieving high image quality and high compression rate through transform encoding. .

[Prior art] ゛In the conventional transform coding, “Adaptive coding of monochromatic and color images J” by Jien et al.
adaptive coding of monoclll
roma and colorimages, IEE
E C0N-25, 11, pp1285-1292 (1
977)), one image can be divided into, for example, 16
It was divided into blocks of x16 pixels, and orthogonal transformation was performed on each block. This method has the disadvantage that when the compression rate is increased, block boundaries become noticeable in the decoded image. Furthermore, although the larger the block size, the higher the image quality, the additional information associated with it becomes enormous, and as a result, a large compression effect cannot be expected.

[Problem S that the invention attempts to solve]

The above conventional technology does not take into consideration the highly efficient retention of additional information as the block size increases, and has problems such as the block size cannot be expanded and block boundaries appear in the image. .

An object of the present invention is to enable efficient storage of additional information and compression to eliminate block boundaries and obtain high image quality.

[Means to solve the problem]

The purpose of eliminating block boundaries is achieved by means of performing frame-frame transformation, which performs orthogonal transformation on the entire image at once without dividing one image into blocks.

The purpose of enabling compression to obtain high image quality is to first obtain high image quality by using frame conversion.
Furthermore, to enable compression, there is provided a means for allocating a variable bit length to each transform component, and a means for compressing data by high-efficiency encoding using, for example, quad-tree representation of the bit allocation table. This is achieved by:

[Effect]

Frame frame transformation operates to orthogonally transform the entire image at once. The means for allocating the variable bit i to each transform component thus obtained operates, for example, by allocating the variable bit i in proportion to the logarithm of the absolute value of the transform component.

Each transform component is compressed and encoded according to the number of bits assigned to it. Here, the bit allocation table serving as additional information is subjected to high-efficiency encoding using, for example, Quad-tree expression, and is compressed and encoded.

By the above operation, block boundaries do not occur in the decoded image, and compression can be performed to obtain high image quality at a high compression rate.

(Example〕 An embodiment of the present invention will be described below with reference to FIGS. 1 to 8.

FIG. 1 is a flowchart showing the encoding processing procedure according to the present invention. First, in block 1, the image data is 2
Input as a luminance value string in a dimensional array. This is f (m
, n). However, the 0 image data f (m, n) where m, n=1, . . . , N and the image size is NXN is subjected to orthogonal transformation in block 2. The transformation is performed on the entire image at once. The converted value is F (u, v)
If you write

F (u, v) is given by the following equation.

...(1) u, v, = 1. ..., N22 and .psi. is an N-th orthonormal matrix that provides orthogonal transformation. Also, ψt is its transposed matrix. Specifically, <P
is a Hadamard transformation matrix or a cosine transformation matrix.For example, a cosine transformation matrix is given by the following equation.

1,j,=l,...,N The following description will be made assuming that cosine transformation has been performed.

In block 3, parameters of allowable distortion for the decoded image are input. Let this be D (>O).

In block 4, the number of bits to be allocated to each F (u, v) is determined according to this D. That is, D is also a parameter for controlling the compression rate of image data.

Before going into the explanation of block 4, I would like to give two points of caution. The first point is that the converted value F(1,1) of the (1,1) component is not subject to subsequent compression processing. The reason is that in cosine transformation, a DC component (a value twice the average brightness of the original image) appears in F(1,1), so if compression is applied to this depending on D, the average brightness of the decoded image becomes the same as the original image. This is because the brightness does not match the average brightness of the image. So F(1゜1
) is given by the following equation.

C(1,1)=NINT(F(1,1)/2)...
(3) However, NINT is converted into an integer by rounding off.

From equation (3), it can be seen that the number of bits required for C(1,1) is equal to the number of bits required to represent the gradation of the original image. The second point gives an idea of encoding the remaining F(11,V). This is because this is necessary for the explanation of block 4, which will be described later. However, this is just a concept, and the exact encoding method will be given later.

First, the number of bits allocated to encode each F (u, , v) is given by the following equation.

B(u,v)=INT(logz(IP(u,v)1/
D))...(4) Here, INT is converted into an integer by rounding down.

From equation (4), bits are assigned to 't F (
When u*v) is written as (u, v) in F, its rend (range) is as follows.

2LD< l Fx(u,v) l <2+ID
・C5) When encoding (u, v) into F, 1 bit is used to determine sign or negative, and K-1 bits represent lFg(u,v)l. This is C
Let (U, V) be, this can be determined as follows. Incidentally, an example is shown in FIG. 2 with 1=4.

First, it is assumed that each IFx(u,v)l is uniformly distributed within the range of equation (5), and this range is equally divided into 2IC-1 parts. Then, label each division point, 1 F
Find the nearest division point less than or equal to g(u,v)l and label it as CIC(u e v), that is, obtain a linear expression.

Cg(u,v)==INT((lFg(u,,v)I
2KD)/2D)...(6) 81Fx(u,v)I =Cx(u*v)X2D+2KD
=-(7) Furthermore, from both equations (6) and (7), it can be seen that the given D is the average quantization error in the decoded image.

In this way, for each transform component F (u, v), its encoded data C (u, v) and the number of allocated bits B (u,
v) is obtained, but B (u, v) for all components
The amount of information it holds is enormous. Therefore, in general, the bit allocation table (B (u, v),) is
) has a high autocorrelation, so this will be used to compress the bit allocation table.

Compression methods include reversible compression, which completely restores the original information, and
There is lossy compression that allows decoding errors. As a method for reversibly compressing the bit allocation table, the obtained bit allocation table 1-0 is divided into bit planes (for example, if the number of allocated bits is 0 to 8 bits, there will be three bit planes).

A possible method is to perform reversible compression such as run-length encoding on each bit plane. Also, by setting a single image model and assuming that the bit allocation table used when performing full frame conversion is known, the error between this and the bit allocation table of the image to be compressed is calculated, and this error table can be used in the above method or as follows. It is also possible to compress using the method shown below.

In this embodiment, an example will be described in which irreversible compression is used to achieve high compression of the bit allocation table. Furthermore, although there are various ways to represent the bit allocation table (one of which is two-dimensional array representation), we will use Quad-tree representation here.

By the way, allowing a decoding error in compressing the bit allocation table means allowing the value of B (u, v) to be different from the original value. Due to this.

F (
u, v) exist. As is clear from both equations (6) and (7), the quantization error caused by this F(u, v) is at most 2B(''?')D
Take. This becomes larger as B (u, v) becomes larger.

This results in a large distortion in the decoded image.

In order to suppress this maximum quantization error to the order of D and to obtain effective compression, a method is adopted in which the bit allocation table is created simultaneously with the quad-tree representation. The method is described below, but this is based on the F (
It is based on the encoding concept of u, v).

The processing procedure in block 4 of FIG. 1 is shown in FIG.
First, in block 31, parameters for decoding errors are input. Let this be R. Since R is the maximum quantization error parameter mentioned earlier, its value should be at most several times D.
Next, in block 32, node settings are performed. As an initial setting, the root node (root) of the Quad-tree is set to the area created by all the two-dimensionally arranged I F (u, V) l.

This is designated as S. In block 33, it is determined whether S is a branch node or a leaf node. This judgment is as follows.

, the following judgment formula is given.

Ml-2KLD<R...(8) and 2"D-M2<R+2D
(9) This determination formula determines whether or not the K2 bit (including the sign bit) is allocated to all I F (u, V) I within the area defined by S.

That is, I 1” (u, V) assigned K2 bits
The range of l is [2D to 2D, ID to 2), but
This range is expanded to [D-R-20 to 2, 20+R to 2], and it is determined whether all I F (u, v) l in S falls within this range.Here, (8 ) and (9
The reason why the values on the right side of the equation ) are different is to ensure that the maximum quantization errors at both ends of the range are equal, as can be seen from the example in FIG. 6(a). Moreover, the effect obtained by expanding the range in this way is, for example, for S in a wide area, I F (u
, V) It is sufficient to assume that I exists at each of the four corners of S.

S that satisfies both equations (8) and (9) is then set as a leaf node having a value of 2 in block 34. Otherwise, S is divided into four parts in block 35 and the process returns to block 32 for each region. This process ends when all nodes become leaf nodes, and Quad-tre
The e-expressed bit allocation table is completed.

In the block diagram of Fig. 1, Quad-tree encoding is performed, in which one bit is added to identify branch nodes and leaf nodes, and deep-first
It is sufficient to order the data according to its structure and write it as a one-dimensional data string.

FIG. 4 is given as an example of the above. Here, (a) in the same figure is a Pi1-allocation table expressed as a region image, (b) is a QuBd-t, rea expression, and (c) in the same figure.
is the encoded result. Each number is the number of allocated bits (K
2), Eo is a branch node and El is a 1-bit code representing a leaf node.

Since there is no branch node in the lowest layer node in the Quad-tree, this code is omitted.

Also, to represent 2 to each, L1 bits are uniformly given. Here, Ll is given by a linear equation due to the nature of orthogonal transformation.

L 1 = I NT (logL + 1)
-(10) Here, L is the number of bits required to represent the gradation of the original image.

In block 6, the bit table is decoded. The contents of the decoding are shown in FIG. If it is a single leaf node that returns to The same operation as above is repeated four times. After passing through blocks 53 and 54, it is determined whether all the data has been read, and if there is any remaining data, the process returns to block 51. If it has been completed, the decoding of the Quadtree is completed at this point. If this is converted into a region image, a two-dimensional array (B (us v)) will be obtained.As an example above,
Please refer to FIG. 4 in the order of (c)→(b)→(a).

In block 7, the cosine transform value F (u, v) is encoded, and the above-mentioned equation (6) can be used as is for this purpose. However, at this point, a decoding error has occurred in B (u, v), so a more accurate formula is as follows.

FIG. 6 shows a conceptual diagram of this encoding with B (u, v)=4. Figure (a) shows the range of IF(u,V)l to which 4 bits are allocated, and the range of IF(u,V)1 and C(
The relationship between u and v) is shown. FIG. 3(b) is an image of an encoded one-dimensional data string. Here So is positive, Sl
is a 1-bit sign (sign bit) representing a negative value.

As described above, the image data is cosine-transformed to become a coded data string of the transformed component and a coded data string of the bit allocation table of all components, and these data are output in block 8 of FIG. 1, and the compression is completed. .

The procedure for obtaining a decoded image from encoded data is shown in the flowchart of FIG. In block 71, encoded data is input, and in block 72, the bit table is decoded. Block 73 decodes the transformed components. Here, the sign bit is used to determine the sign bit.

The absolute value of the conversion component is obtained from the above-mentioned equation (7). Block 74 performs a full frame reverse direct exchange. The inverse cosine transform is given by the following equation.

ta@ n==1*-tN ”・(
12) Output at 75.

FIG. 8 is a block diagram of an encoding device based on the present invention. Image data having a two-dimensional array of brightness values inputted from the image input bag h 81 is subjected to orthogonal transformation in a cosine transformer 82 and temporarily stored in a buffer 83. The bit allocation table encoder 84 performs encoding according to the procedure shown in FIG.

There are two output lights, one is a communication path or storage means 88
The other is a bit allocation table decoder 85. The buffer 86 stores a decoded two-dimensional pitch allocation table. The quantizer 87 creates encoded data of the transform component in the manner shown in FIG. 6, and outputs it to one communication channel or storage means 88. For decoding, first the bit allocation table is decoded by the decoder 89, and then the bit allocation table is decoded by the decoder 89.
Store at 0. The transformed components are obtained through an inverse quantizer 91 and stored in a buffer 92. Inverse cosine transformer 93
The brightness value of the decoded image is obtained by the image output device 94.
is output to.

〔Effect of the invention〕

The compression effect according to the present invention is quantitatively expressed.

When the image size is N×N pixels and each pixel has L pit gradation, it is assumed that each pixel is expressed by C (<L) bits by full frame conversion and bit allocation to the conversion components. What is the compression ratio at this time?

C/L...(i) However, a bit allocation table is required as additional information here. As mentioned above, this is logzL+1 bit per pixel, and the overall compression ratio is (C+logzL+1)/L...(it
) is 0. According to the present invention, this additional information is compressed.

If the compression ratio is set to 1/, the overall compression ratio is (C+ (
l ogzL+1)/K)/L...'(in). Specifically, L=8. If C=0.5, then =20 is generally possible, and what was 45780 in (...) can be achieved to 7780 by the present invention (iii).

[Brief explanation of the drawing]

FIG. 1 is a flowchart of an embodiment of the present invention, FIG. 2 is a diagram showing an example of a uniform quantizer for cosine transform values, and FIG. 3 is a flowchart showing a procedure for creating a bit allocation table of quad-tree representation. FIG. 4 is a diagram showing an example of representation and encoding of a bit allocation table, FIG. 5 is a flowchart showing the decoding order of the bit allocation table, and FIG.
The figure shows an example of a quantizer for cosine transform values and an example of encoding, FIG. 7 is a flowchart showing the procedure for obtaining a decoded image, and FIG. 8 is a block diagram of the encoding device according to the present invention. 2 buds in the process 2 buds 3 prisoners 4~-33 works! Inner dAr4 Inner Port

Claims (1)

[Claims] 1. In an image data compression method using transform encoding,
An image data compression method characterized by performing orthogonal transformation on the entire image at once, assigning a variable number of bits to each transform component during quantization of the transform component, and compressing data using the assigned number of bits. 2. The above-mentioned data compression process is the first image data compression method consisting of a process of expressing the data to be compressed in a tree structure.