JPH01276981A - Orthogonal transform coding system for picture data - Google Patents

Abstract

PURPOSE:To compress an input picture at a fixed compressing ratio irrespective of the content of the input picture by comparing an input picture pattern with plural representative picture patterns stored in a memory beforehand and finding the closest representative pattern, and then, coding a high-frequency area by using the number indicating the representative pattern. CONSTITUTION:The pattern of orthogonally transformed picture data of several sheets of pictures selected as a training sequence in each block of a high-frequency area is stored in a code book as a model. A quantization circuit 14 for high-frequency component selects an output vector which gives the minimum distortion by comparing the pattern of an input vector with patterns of output vectors stored in the code book and outputs the index of the selected output vector. Then a coding circuit 15 for high frequency component outputs the index outputted from the circuit 14 after coding. Therefore, the picture quality and compressing ratio of vector quantization depend upon the number of the output vectors stored in the code book and the way of output vector selection.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to an image data encoding method for the purpose of data compression, and particularly to an image data orthogonal transform encoding method using orthogonal transform.

<Prior Art> A two-dimensional full-color still image has a huge amount of information, but this image data has a large amount of redundancy, and by suppressing this redundancy, it is possible to compress the image data.

Conventionally, one such image compression method uses Fourier transform to remove redundancy of image data.

There are methods of converting each pixel value into a spatial frequency component using orthogonal transformations such as Hadamard transformation, Barr transformation, cosine transformation, and Karhunen-Loeve transformation (KL transformation).

In this type of image compression method, energy is concentrated in specific spatial frequency components due to the above-mentioned transformation, so a large number of bits are allocated to spatial frequency components with high energy, and a small number of bits are allocated to spatial frequency components with low energy. By allocating , the amount of data can be compressed. FIG. 8 shows a block diagram of an image data orthogonal transform encoding device implementing the above-described method.

This image data orthogonal transform encoding device is as follows.

works. In other words, full-color still image data (
For example, red (R), green (G), blue (B)
3 primary colors 1 image size is NXN) is the preprocessing circuit!
The signal is converted into a luminance signal (Y) and color difference signals (I, Q). The reason for converting to the YIQ system in this way is as follows. In other words, the RGB system has a high degree of redundancy due to the high correlation between color components, but the YIQ system uses an orthogonal system with a low correlation between luminance and color difference, so redundancy is significantly reduced. In the system, Y due to human visual characteristics
This is because even if the accuracy of the I and Q components is considerably lower than that of the other components, it is not noticeable, so it is possible to roughly encode the I and Q components and reduce the total amount of code. RGB to Y
Conversion to the IQ system is performed using the following equation.

Next, the data converted into the YIQ system is converted into spatial frequency components having mutually independent conversion axes in an orthogonal conversion circuit 2. Then, the obtained orthogonal transformed image data is input to the encoding circuit 3, which allocates bits according to the energy of each frequency component, and performs quantization and encoding according to the number of allocated bits. and record/playback device 4.
to be recorded.

To reproduce this data and display it on the image reproduction device 7,
The data recorded in the recording/reproducing device 4 is read out and input to the decoding circuit 5, and the orthogonal transform image data decoded by the decoding circuit 5 is processed by the orthogonal transform circuit 2 by the inverse transform circuit 6. After performing the inverse transformation of the above conversion, the conversion from the YIQ system to the RGB system is performed to obtain an RGB signal. Then, a reproduced image is displayed on the image reproduction device 7 based on this RGB signal. Conversion from the YIQ system to the RGB system described above is performed using the following formula.However, in the above method, bit allocation is performed only based on the magnitude of the energy of the spatial frequency component, so even high frequency components have large energy. In some cases, a large number of bits can be allocated.

Therefore, human visual characteristics are sensitive to low frequency components (flat parts with little change in pixel values in the spatial coordinate system), but sensitive to high frequency components (edge parts with rapid changes in pixel values). The inventor first divided the spatial frequency region into a low frequency region and a high frequency region, noting that even if there is some error, it is not noticeable. We proposed an orthogonal transform coding method for image data that can increase the compression rate by allocating a small number of bits to high frequency components. In this image data orthogonal transform encoding method, when encoding information in the low frequency region, this low frequency region is divided into multiple blocks, the average value of orthogonally transformed image data for each block is calculated, and this The number of bits for each block is calculated and encoded based on the average value. On the other hand, when encoding information in a high frequency domain, the high frequency domain of the orthogonal transform data is
It is decomposed into bit planes of ~5 bits, and run length encoding is performed for each bit plate.

<Problems to be Solved by the Invention> However, the image data orthogonal transform encoding method that divides the spatial frequency region of the orthogonally transformed image data into a low frequency region and a high frequency region and encodes each region, When encoding information in the high frequency domain, the high frequency domain is decomposed into bit brains and run length encoded, so the amount of data when encoding the high frequency domain is reduced. Depends on the content of the image. That is, when an image with relatively dispersed energy is encoded, the amount of data increases and the compression rate is considerably reduced.

Therefore, an object of the present invention is to divide the spatial frequency domain into a low frequency domain and a high frequency domain, divide the plurality of high frequency domains into blocks, and perform vector quantization for each block to reduce the amount of data. It is an object of the present invention to provide an image data orthogonal transform coding system capable of compressing image data in a high frequency domain at a constant compression rate without depending on the image.

Means for Solving the Problems> In order to achieve the above object, the present invention provides an image data orthogonal transformation in which two-dimensional full-color still image data is orthogonally transformed, and spatial coordinates are converted into spatial frequencies and encoded. In the encoding method, a plurality of region patterns are provided that divide the spatial frequency region into a low frequency region and a high frequency region, each using a predetermined threshold value, and the absolute value of orthogonal transformation data corresponding to the low frequency region of each region pattern is determined. The sum is calculated, and the low frequency region of the region pattern in which the sum of absolute values is the maximum is identified as the low frequency region of the orthogonal transformation data, and
The high frequency region of the above area pattern is identified as the high frequency region of the above orthogonal transform data, the low frequency region of the above orthogonal transform data is divided into a plurality of blocks, and the average value for each of the above blocks of the above orthogonal transform data is calculated. , while allocating a large number of bits to blocks with a large average value, and allocating a small number of bits to blocks with a small average value, encoding the low frequency domain in the orthogonal transformed data,
The high frequency region of the orthogonal transformation data is divided into a plurality of blocks, and the pattern of the orthogonal transformation data in this block is compared with the representative pattern of the plurality of representative image data stored in memory in advance. The present invention is characterized in that a representative pattern closest to the pattern of orthogonally transformed image data is found, and a high frequency region in the orthogonally transformed data is encoded using a number representing this representative pattern.

<Example> Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments.

FIG. 1 is a block diagram showing an embodiment of an image data orthogonal transform encoding device according to the present invention. This image data orthogonal transform encoding device includes a preprocessing circuit 11, an orthogonal transform circuit 1
29 Frequency domain selection circuit 13° Quantization circuit for high frequency components 14. Encoding circuit for high frequency components 15. Bit allocation calculation circuit 16. Quantization circuit for low frequency components 17. Encoding circuit 18 for low frequency components. Recording and reproducing device 19. Decoding circuit 20, inverse conversion circuit 21, and image reproduction device 22
It is roughly structured by.

The image data orthogonal transform encoding device having the above configuration operates as follows.

First, the encoding process will be explained.

The preprocessing circuit 11 processes full-color still image data (each pixel is red (R), green (G), and blue (B).
) data of image size NXN expressed by three signals),
It is used to convert Y (luminance signal) to IQ (color difference signal). The image data subjected to YIQ transformation by the preprocessing circuit 11 is orthogonally transformed by the orthogonal transformation circuit 12. This orthogonal transformation includes many transformations such as Fourier transformation, Hadamard transformation, Barr transformation, and Karhunen-Loeve transformation (KL transformation). In this example, the discrete cosine transform (hereinafter referred to as
The following explanation will be given using DCT as an example. therefore,
In the orthogonal transformation circuit 12, the DOT is performed and the signals are transformed into spatial frequencies having mutually independent transformation axes.

FIG. 2 shows an outline of the conversion by OCT.

In FIG. 2, (x, y) is the position of a pixel on the screen, and WXy is the pixel value at the position (x, y) on the screen. Also, (u, v) is the frequency component after converting to spatial frequency, and Wuv is (u, v) in the spatial frequency domain.
It is the value of the DCT coefficient of the component. From Wxy above to WLIV
The conversion formula to is expressed by the following formula.

(U, V=O, l,..., t1) After converting the spatial coordinate system into a spatial frequency system by DCT according to the above equation, the frequency domain selection circuit 13 converts the low frequency It is separated into high frequency region and high frequency region.

That is, the energy distribution of spatial frequency components in the spatial frequency domain differs depending on the image data. Therefore, in order to perform adaptive processing even for image data with different energy distributions of spatial frequency components, the method shown in Fig. 3 (
Set four types of area patterns for determining the low frequency area of image data converted into a spatial frequency system (orthogonal transformed image data), as illustrated by two-dimensional arrays in a) to FIG. 3(d). . This area pattern is divided into a low frequency area (shaded area) and a high frequency area by a predetermined threshold value of the spatial frequency component. The area pattern in this embodiment is set so that the low frequency area is 174 of the total area.

Based on the energy distribution of the input orthogonal transformed image data, the optimal region pattern is selected from the above four types of region patterns as follows. First, the hatched portion (low frequency region) of each area pattern in FIGS. 3(a) to 3(d)
The sum of absolute values of DCT coefficients of orthogonally transformed image data corresponding to is calculated for each image data. Next, from among the above four types of area patterns, the area pattern having the maximum value of the above-mentioned sum is selected. Then, the low frequency region in this selected region pattern is determined as the low frequency region of the orthogonal transformed image data. As the low frequency region, regions as shown in FIG. 3(e) to FIG. 3(1) may be set.

Furthermore, the selection of the above-mentioned low frequency region will be explained in detail by giving an example. Figure 4 shows image data of size 16
FIG. 3 is a diagram expressing DCT coefficients in a two-dimensional array when orthogonally transformed to an X16 spatial frequency system. In this case, the sum of the absolute values of the DOT coefficients in FIG. 4 corresponding to the shaded areas (low frequency regions) in each area pattern in FIGS. 3(a) to 3(d) is as shown in FIG. 3(a). 1028 in the low frequency region of
21, 102892 in the low frequency region of Fig. 3(b),
In the low frequency region of FIG. 3(c), it is 103467, and in the low frequency region of FIG. 3(d), it is 103401. Therefore, the hatched portion of the area pattern in FIG. 3(c) with the largest sum value is selected as the low frequency area of the orthogonal image data in FIG. 4.

On the other hand, as the high frequency region, a portion (other than the shaded portion) that is not selected as the low frequency region in the region pattern of FIG. 3(c) is selected.

Next, a method of quantizing and encoding information in the low frequency region performed by the bit allocation calculation circuit 164, childization circuit 17 and encoding circuit 18 will be described.

As shown in FIG. 5, the low frequency region of the orthogonal transformed image data of FIG.
DCT coefficients are divided into blocks each having one unit. Here, it is known that data after image data is orthogonally transformed by DCT follows a Gaussian distribution or a Laplace distribution.

In this way, since the orthogonally transformed image data after DCT follows the Gaussian distribution or the Laplace distribution, it is appropriate to take the logarithmic value of the absolute value of each DCT coefficient of this orthogonally transformed image data and perform logarithmic compression. Then, this logarithmically compressed DC
Quantization is performed by calculating the average of the T coefficients for each block. The average value of the DCT coefficients in each block after logarithmic compression is calculated using the following equation.

Δ Here, σpq is the average value of the DCT coefficients of the block located at (p, q) in the spatial frequency domain, and Wij is the DCT coefficient value in the block. The above quantization is uniform quantization.

Based on the average value of the DCT coefficients obtained for each block in this manner, the bit allocation calculation circuit 16 calculates the number of bits to be allocated to each block as follows.

That is, assuming that the DCT coefficients follow a Gaussian distribution, the bit allocation that minimizes the mean square error is expressed by the following equation.

Here, Ftpq is the number of bits allocated to the block located at (p, q) in the spatial frequency domain, R is the average number of bits allocated to each block in the low frequency domain, and RI2 is lo
This is the average value of the entire low frequency region of gs I Wij I.

In this way, when the total number of bits to be allocated to each block is determined, the DCT coefficient Wij is converted by the quantization circuit 17 to the sign of Wij and the logs l W ij I− sum pq
Quantization is performed by uniform quantization of . Therefore, as information on low frequency components that are sensitive to human vision, Wij, △
The sign of △, logt l Wtj l-σpq and σp
Three types of information q are output.

In this way, in the case of information in the low frequency domain in the spatial frequency domain, the number of bits is distributed and encoded according to the average value of the DCT coefficients in block units. , it is possible to allocate and encode a large number of bits to blocks with large energy.

Therefore, it is possible to perform encoding at a high compression rate while suppressing image deterioration.

Next 1. A method for quantizing and encoding information in a high frequency domain performed by the quantization circuit 14 and encoding circuit 15 for high frequency components will be described.

First, the principle of vector quantization will be briefly explained. A two-dimensionally arranged signal space to be quantized is divided into a plurality of blocks. It is assumed that the block size of the above block is nXn. In that case, the n2-dimensional signal space Rn of the above block! Input vector x=(x,...
・n1-dimensional signal space Rn for X n t )! M of
Let the divisions be RI...RM, and the output vector yi = [yl] which is the representative point of the subspace Ry.

...yin! ] The set of ゛ is y = (i-'y,), and the index set of 4 is 1 = (1, 2,...M)
Then, vector quantization Q is expressed as a cascade of encoding C and decoding.

Q(")=4"if"i"eJ Q=D-C C:i'-*i "if"d(i": fi)≦d
(i': 7) "for" VjD:l→Yi Here, the distortion measure d(i: i) represents the distance between the input and output vectors, and when defined by the squared distortion measure norm.

Also, the set of the above output vectors yi = (Tr・
...Self, ) are determined in advance as follows. That is, using the pattern of each block in the high frequency domain of the orthogonally transformed image data of several images selected as a training sequence as a model, the n1-dimensional signal space R formed by the input vectors is created by clustering.
n! The optimal set of output vectors y Y = (i-i”y) is determined by dividing the cluster and calculating the center of gravity (output vector) of the cluster obtained by this division. The set of output vectors Y is stored in a codebook in a memory (not shown).

The quantization circuit 14 for high frequency components compares (pattern matching) the pattern of the input vector The output vector yi that gives the above-mentioned minimum distortion is selected, and the index i of the output vector Vi that gives the above-mentioned minimum distortion is output as a quantized value. Then, the high-frequency component encoding circuit 15 encodes the index i output from the quantization circuit 14 and outputs it.

When reproducing high frequency domain information encoded in this way, the decoding circuit 20
The receiver takes the quantized value i output from 4 and calculates the quantized value (the above index) i from the same codebook as the encoding side.
The output vector Vi corresponding to is extracted, and the above block is reconstructed from this output vector yi=(yH, . . . Win).

Therefore, the image quality and compression rate in vector quantization are
It depends on the number of output vectors stored in the codebook and how the output vectors are selected.

Furthermore, the output vectors are highly dependent on the images chosen as training sequences.

Therefore, when vector quantization is performed by selecting the entire spatial frequency domain as a training sequence, the original image is subjected to preprocessing such as normalization or orthogonal transformation in order to reduce dependence on the original image as much as possible. However, even then, it was not possible to adapt well to image quality in which the orthogonally transformed image data of the original image has various types of patterns.

However, in this embodiment, vector quantization is performed only on the high frequency region of orthogonally transformed image data obtained by orthogonally transforming the original image. As mentioned above, human visual characteristics are not very sensitive to high frequency regions. Furthermore, since energy is not concentrated in the high frequency region, the orthogonal transformed image data itself has many similar patterns. Therefore, the types of vectors are more limited than when the entire frequency range is selected as the training sequence. Therefore, 1. It is possible to encode the high frequency region of orthogonally transformed image data at a constant compression rate while suppressing image deterioration, and it is possible to obtain a restored image with high image quality.

Further, the quantization and encoding of the above-mentioned high frequency domain information will be explained in detail by giving an example. This quantization is done by quantizing the high-frequency components by 2'' level, assigning 1 bit to the sign of the DCT coefficient, and (a-1) bits to the DCT coefficient. A case will be explained in which OCT coefficients arranged two-dimensionally as shown in Fig. 4 are used as the converted image data. The quantization circuit 11 performs quantization as follows, for example.

That is, when 0≦IWij1×23, 0゛
When 23≦・IWijl×45 145≦1Wij
When l<91 2 When 91≦l Wij l As 3, 2
Quantize in bits. The results are shown in Figure 6(a) (
(The quantization value O is omitted.) Next, the quantization circuit 1
4, the high frequency region of FIG. 6(a) is divided into 12 blocks of block size 4×4 as shown in FIG. 6(c). Input vector Lt+X in each block above
+t... The elements of X13 and X44 are shown in FIG. 6(b). Here, it is assumed that the set Y of output vectors yi stored in the codebook is composed of four types of output vectors Yr*Y*+Ys and Y4, as illustrated in FIG. 6(d). Then, the quantization circuit 14 outputs one of the input vectors X (for example, Y) and the four output vectors Yr*Y**Ys.

By performing pattern matching with y4, the above distortion measure d(
Obtain the output vector (e.g., i) that minimizes x, i, ). Input vector X obtained in this way
A combination of the output vector yi and the output vector yi is shown in FIG. 6(e). Then, the quantization circuit 14 outputs a quantized value sequence as shown in FIG. 6(f) (that is, the index sequence i of the output vector Yi obtained as described above). Then, the encoding circuit 5 is replaced by the quantizing circuit 1.
The quantized value i output from 4 is encoded.

In this way, information in the high frequency domain in the spatial frequency domain is vector quantized in block units, so
In the high frequency domain, where human vision is insensitive and the types of vectors in orthogonally transformed image data are limited, the image quality and compression rate do not depend on the image content, and the compression rate remains constant without deteriorating the image quality. Encoding can be done with .

The information in the low frequency region and the high frequency region obtained as described above is stored in the recording/reproducing device 19 (magnetic tape, floppy disk, etc.).

Next, the decoding process will be explained.

The data read from the recording/reproducing device 19 is input to a decoding circuit 20, and the decoding circuit 20 performs decoding processing as described below.

In the case of information in the low frequency region, the number of bits allocated to each block is calculated using the same method as in the case of encoding described above from the average value σpq of the DCT coefficient coefficient rate for each block read from the recording/reproducing device 19. do. Then, determine the quantization width from the allocated number of bits and
01h l Wij l - Calculate the value of Opq and plan
By subtracting the value of q, logs l W ij I
seek. Furthermore, decoding is performed by calculating the power of 2 of logt I Wij I to obtain the value of 1WiN, that is, the DCT coefficient value.

In the case of information in a high frequency domain, the code book stored in the memory (not shown) of the decoding circuit 20, which is the same as the code book on the encoding side, is stored in accordance with the index i read from the recording/reproducing device 19, and the index i is read out. Extract the output vector Fi corresponding to . Then, the output vector yl extracted in this way is incorporated into the corresponding block in the high frequency domain in the spatial frequency domain in a two-dimensional array. The quantized values in the spatial frequency domain thus obtained are shown in FIG. 7(a). In addition, the result of decoding the quantized value in FIG. 7(a) is shown in FIG. 7(b).
Shown below. Here, the symbol l in FIG. 7(a).

2.3 are decoded as 32, 64, and 128, respectively.

As described above, after both low frequency domain information and high frequency domain information are decoded and DCT coefficients are calculated, these two types of DCT coefficients are combined to reconstruct a spatial frequency system, and this data is It is input to the inverse conversion circuit 21. The inverse transformation circuit 21 inversely transforms the spatial frequency system into a spatial coordinate system. Here, when the DCT coefficient is Wuv, the pixel value is Wxy, and the image size is NXN, the inverse discrete cosine transform is defined by the following equation.

(x, y = IO, 1, ..., tr) The pixel values calculated by the above formula are a luminance signal (Y) and a color difference signal (1, Q), so they are converted into red (R), Green
The data is converted into RGB signals of green (G) and blue (B) and input to the image reproducing device 22.
A restored image is obtained based on the RGB signals.

Note that parameters such as the quantization width, threshold value, and block size used in the image data orthogonal transform encoding method of the present invention can be determined experimentally or empirically.

<Effects of the Invention> As is clear from the above, the image data orthogonal transform encoding method of the present invention provides a plurality of area patterns that divide the spatial frequency area into a low frequency area and a high frequency area, and The sum of absolute values of the orthogonal transform data corresponding to the low frequency region of The high frequency region is specified as the high frequency region of the orthogonal transform data, and the low frequency region to which human vision is sensitive and the high frequency region to which human vision is insensitive are encoded using different encoding methods. It is possible to encode image data by reducing the number of bits for the high-frequency region, where the visual sense is insensitive, and increasing the number of bits for the sensitive low-frequency region, thereby suppressing deterioration of image quality and compressing image data at a higher compression rate than before. Can be compressed.

Further, the image data orthogonal transform encoding method of the present invention compares the pattern of the orthogonal transform data in the block of the high frequency domain of the orthogonal transform data with a representative pattern of a plurality of representative image data stored in a memory in advance. Therefore, the representative pattern closest to the pattern of orthogonally transformed image data in the above block is found, and the high frequency domain is encoded using the number representing this representative pattern, so that a constant compression is achieved regardless of the content of the input image. Orthogonal transform data in the high frequency domain can be compressed at a high rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of an image data orthogonal transform encoding device according to the present invention, and FIG. 2 is a DCT of image data.
Figure 3 is an explanatory diagram of each area pattern, Figure 4 is the DC after orthogonal transformation by DCT.
FIG. 5 is a diagram showing an example of a block in the low frequency domain. FIG. 6 is a diagram showing an example of vector quantization in the high frequency domain. FIG. 7 is a diagram showing an example of vector quantization in the high frequency domain. FIG. 8 is a block diagram of a conventional image data orthogonal transform encoding device. 11... Preprocessing circuit, 12... Orthogonal transformation circuit,
13... Frequency domain selection circuit, 14.17... Quantization circuit, 15.18... Encoding circuit, 16... Bit allocation calculation circuit, 19... Recording/reproducing device, 20... decoding circuit,
21... Inverse conversion circuit 22... Image reproducing device.

Claims (1)

[Claims] (1) In the image data orthogonal transformation encoding method, which performs orthogonal transformation on two-dimensional full-color still image data and converts the spatial coordinates into spatial frequencies and encodes them, the spatial frequency domain is converted to a low frequency by a predetermined threshold value. A plurality of area patterns are provided that are divided into areas and high frequency areas, and the absolute value sum of the orthogonal transformation data corresponding to the low frequency area of each of the area patterns is calculated, and the low frequency area of the area pattern where the above absolute value sum is the maximum is calculated. Identifying a frequency domain as a low frequency domain of the orthogonal transformation data, identifying a high frequency domain of the area pattern as a high frequency domain of the orthogonal transformation data, and dividing the low frequency domain of the orthogonal transformation data into a plurality of blocks. Then, calculate the average value for each block of the orthogonal transformed data, and allocate a large number of bits to blocks with a large average value, while allocating a small number of bits to blocks with a small average value, and then The low frequency region of the orthogonal transform data is encoded, the high frequency region of the orthogonal transform data is divided into multiple blocks, and the pattern of the orthogonal transform data in this block and multiple representative image data stored in memory in advance are The representative pattern closest to the pattern of the orthogonal transformed image data in the block is determined by comparing the representative pattern of A featured image data orthogonal transform encoding method.