JPH04222122A - Data compressor - Google Patents

Abstract

PURPOSE:To improve arithmetic accuracy for compressing data. CONSTITUTION:A data conversion part 31 is provided to execute data conversion or inverse data conversion arithmetic based on a prescribed coefficient to a numerical example, and a quantizing part 32 is provided to execute quantizing arithmetic. The coefficient of the data converting arithmetic part 31 is changed to an integral ratio and a gain change caused by the change into the said integral ratio is adjusted by the quantizing part 32.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data compression device used for compression processing of image data, and more particularly to a data compression device with improved calculation accuracy during data compression.

0002

2. Description of the Related Art DCT (discrete cosine transform) is becoming the mainstream of high-efficiency encoding technology for images based on ISDN and CD-ROM. Not limited to this DCT, if the average number of bits per pixel is reduced by high-efficiency encoding, the quality of the image deteriorates, and if the compression rate is increased, the image quality deteriorates. The problem when compressing the current standard television signal to 1.5 Mbit/s is:
These are deterioration of the contour portion and block distortion occurring in block units (for example, 8×8 pixels) processed by DCT. When reproducing pixels by inverse transformation, all DCT outputs in a block are linearly summed, but if there is information loss in even one of the 64 DCT outputs of a block consisting of 8 x 8 pixels, Deterioration occurs in the reproduced pixels of the entire block.

[0003] Therefore, in order to reduce such block distortion as much as possible, reference document IEEE TRANSAC
TIONS ON ACUSTICS,SPEECH
, AND SIGNAL PROCESSING. V.O.
L. 37. No. 4. APRIL1989(The L
OT Transform Coding Without
ut Blocking Effects, HENRIQ
UE S. MALVAR, DAVID H. STAEL
The LOT operation disclosed in IN) has been proposed. Figure 5
1 shows a LOT arithmetic device that performs this LOT arithmetic processing, and is a block diagram of a one-dimensional LOT. In FIG. 5, 1 is a LOT calculation device, 2 and 3 are DCT devices, and various calculation units shown in FIGS. 6 to 9 are connected to the DCT devices 2 and 3. Here, FIG. 6 shows subtraction c=a+
(-b), FIG. 7 shows the calculation to add c=a+b, FIG. 8 shows the calculation to adjust a predetermined gain (for example, 1/2), and FIG. 9 shows the calculation to perform vector rotation. It shows. The outputs of the DCT devices 2 and 3 are even (ev
En: even number) outputs 0, 2, 4, 6 and odd (odd number) outputs 1, 3, 5, 7 are added and subtracted, and finally only the odd components are vector-rotated by the butterfly calculator shown in Figure 9. and becomes LOT data. One-dimensional LO shown in Figure 5
In the T configuration, if 16 pixels (X0 to X7, X0' to X7') are input to the DCTs 2 and 3 constituting the LOT calculation device 1, 8 data (Y0 to Y7) can be output by LOT calculation. That is, at the first stage of input, a one-dimensional DCT operation is performed to obtain 16 data, and after various operations are performed on these 16 data, vector rotation is performed to finally obtain 8 data. Since this LOT operation is one-dimensional, the output is 8×16 for a 16×16 input pixel, and the same LOT operation is performed by switching the vertical and horizontal directions again to obtain 8×8 data.

FIG. 10 is a block diagram of an image data compression device 11 using the LOT calculation device 1 shown in FIG. In FIG. 10, the image data stored in the image data memory 12 is subjected to two-dimensional DCT processing by the DCT calculation device 13 and output to the LOT calculation device 1. The LOT calculation device 1 is a DC
A one-dimensional (horizontal) LOT operation is performed on the data input from the T operation device 13, and data for one block line is stored in the block line buffer A14 for the LOT operation. After the operation for one block line is completed, the same operation is performed to store the block line buffer B15.
Store data in. Here, the DCT calculation device 13 temporarily stops operation, and the LOT calculation device 1 performs a vertical one-dimensional LOT calculation on the data in the block line buffer A 14 and the block line buffer B 15, and the quantization device 16 The quantizer 16 quantizes the data and outputs the quantized data to the compressed data memory 17. Again, DCT calculation device 1
3 and LOT calculation device 1, DCT,L is calculated from the original data by
An OT operation is performed and data for one block line is written to the block line buffer A14. Then, block line buffer B15 and block line buffer A
14, a vertical LOT operation is performed and quantization is performed. Thereafter, the block line buffers are repeatedly switched to process one screen. In addition, in the reverse direction, image data memory 12 ⇔ compressed data memory 17, DC
The basic operation is the same, just by changing T⇔inverse quantization, LOT⇔ILOT, and quantization⇔IDCT.

By the way, not only the image data compression device 11 having the above-mentioned LOT operation device 1 but also ordinary data compression devices perform data conversion (FET, DCT, etc.) on a data vector (numerical string). After that, quantization is performed. FIG. 11 is a diagram showing an example of orthogonal data conversion and quantization in such a data compression device. In FIG. 11, 21 is a data conversion unit of a data compression device;
22 is a quantization unit, and the data conversion unit 21 receives input x0,
As shown in FIG. 12, x1, x2, and x3 are data-converted by performing multiplication using a trigonometric function as a multiplier, and output to the quantization section 22. The quantization unit 22 quantizes the converted data by the calculation shown in FIG. 13 and outputs quantized data y0, y1, y2, and y3. In this case, in the conventional data compression device, during data conversion operation in the data conversion section 21, quantization in the quantization section 22, etc.
For numerical values that cannot be expressed in binary (for example, 1/3), the lower digits are rounded to approximate the numerical value and perform calculations.

[0006]

[Problems to be Solved by the Invention] However, such conventional data compression devices are configured to perform rounding approximations many times during data conversion operations and quantization. As the number of times increases, the calculation accuracy deteriorates significantly, and the calculation bit width must be increased, resulting in an increase in circuit scale. That is, in the conventional data compression apparatus as shown in FIG. 11, input is multiplied using a trigonometric function as a multiplier. However, when configuring hardware, trigonometric function coefficients must naturally be expressed in binary decimal numbers, but in general, accurate values cannot be expressed in binary decimal numbers; Requires bit rounding. In this way, since the calculation uses coefficients that contain errors from the beginning, in order to maintain accuracy, it was necessary to preserve the output value as much as possible without rounding it and pass it on to the next stage of calculation. . This has the disadvantage that the data bus width and the circuit scale of the arithmetic unit become extremely large. For example, in FIG. 11 above, when the x input and cos coefficient are all 8 bits, the data bit width is shown in FIG. 14, and the data bit width t is:
t1: 17bit, t2: 26bit, t3: 35bit
It becomes t.

[0007]

[Means for Solving the Problems] In order to achieve the above object, the data compression device according to the present invention includes data conversion calculation means for performing data conversion or inverse data conversion calculation on a numerical string based on a predetermined coefficient. , and quantization means for performing a quantization operation that replaces a continuously changing code with a finite number of levels, the data compression device replacing the coefficients of the data conversion operation means with a ratio of integers. At the same time, the change in gain caused by replacing the integer with the ratio is adjusted by the quantization means.

[0008]

[Operation] The operation of the present invention is as follows. The coefficients of the data conversion calculation means are replaced with integer ratios, and the change in the gain is adjusted by the quantization means. Therefore, most of the calculations are performed using ratios of integers that do not include errors, which greatly improves the calculation accuracy and significantly reduces the circuit scale.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be explained below based on the drawings. 1 and 4 are block diagrams showing a data conversion section and a quantization section of a data compression apparatus according to the present invention, and are opposite views to FIG. 9. In FIG. 1, 31 is a data compression device 30
, and its quantization section 32, and the calculation coefficient of the data conversion section 31 (the number circled in the figure) is cos 0.
．． 13π and sin0.13π, cos0.16π and si
n0.16π is approximated by a ratio of integers as shown in FIG. Note that the ratio of integers does not necessarily have to be such a ratio, and a ratio with a larger number of digits may be used to replace the ratio with a more accurate ratio. Furthermore, since the operation based on the integer ratio has a gain different from that of the operation that should be performed, the difference in gain is absorbed by the quantization operation. In the case of this practical example, since the output of such a calculation is the input to the next stage calculation, the gain adjustment calculation is performed once at z1, z2, and z3 in FIG. Note that gain matching in this case means that the gains of the inputs match, but does not mean that the gains of the output match. The gain ratio between these inputs is shown by Equations 1 and 2. z12:z22≒1:72+32
Formula 1z22:z32≒1:(92+5
2) x z2 Equation 2 As an example that satisfies Equations 1 and 2 above, z1:z2:z3=5:38:3 in Figure 1
A coefficient of 92 is given. Note that this is just one example, and it goes without saying that such a numerical value does not necessarily have to be used. Further, the change in each output gain caused by expressing it as a ratio of integers is absorbed in the quantization section 32. That is, the arithmetic coefficients of the data converter 31 are replaced with integer ratios, and the resulting changed gain is corrected by the quantizer 32. FIG. 3 is a diagram showing an inverse data conversion section and a quantization section in inverse transformation of the data compression apparatus, and shows an example of performing the inverse transformation of FIG. 1. In FIG. 3, 41 is the data compression device 3
0 is an inverse quantization unit, and 42 is its inverse data conversion unit. The numerical examples are upside down or reversed from those in FIG. In the case of inverse conversion, as in the case of FIG. 1, the calculation coefficients of the inverse data conversion unit 42 are replaced with integer ratios as shown by the numbers circled in FIG.
The inverse quantization unit 41 adjusts to absorb this change in gain.

Next, the operation of this practical example will be explained. During data conversion calculation (see Figure 1) In the conventional example, the coefficient co required to calculate y3
When s0.13π and sin0.13π, cos0.16π and sin0.16π, etc. are expressed as binary decimal numbers, they are infinite decimals as shown in Figure 2, and the accuracy deteriorates due to rounding errors. Furthermore, the accuracy of the output data was considerably degraded because the coefficients with degraded accuracy were used four times in calculations, but in this practical example, by replacing the coefficients of the data converter 31 with ratios of integers, All calculations by the data converter 41 are made possible by integer calculations. Specifically, the calculation coefficients cos0.13π and sin0.13π,co of the data conversion part shown in FIGS. 11 and 12 are
s0.16π and sin0.16π are expressed as infinite decimals as shown in FIG. Now, this cos0.13π and si
If we take the ratio of both n0.13π, cos0.16π and sin0.16π, the ratios are 7:3 and 9:15, respectively. In this way, calculations are performed using values obtained by replacing calculation coefficients with ratios of integers. The change in gain caused by replacing the calculation coefficients with integer ratios is determined by the quantization units z1, z2, z3 and θ0, shown in FIG.
Absorption is performed by quantization at θ1, θ2, and θ3. In Figure 1, z1 is z before performing the calculation in Figure 1 (A).
The size is adjusted by multiplying by 5 by 1. Similarly, z2 is used to adjust the gain in FIG. 1(A), and z3 is used to adjust the gain in FIG. 1(C). By doing this, calculations are performed using integers until they are output to the quantization unit (32), but depending on the coefficients, these z1, z2, z3 can be changed to θ0, θ1, θ2, θ
3 may be calculated all at once. Here, the quantization value is, for example, in the case of θ0, the original quantization number 1
The value θ0 = /2 multiplied by the value 7/cos0.13π
0.13π/(7×2), and in the case of θ1, the gain z2,9/co is changed to 1/3 of the original quantization number.
Numerical value θ1 multiplied by the reciprocal of s0.16π = cos0.16
π/(38×9×3). In this case, the quantization value will be very small, but it can be reduced to a small number of bits by performing floating point arithmetic.

During the inverse data conversion operation (see FIG. 3) During the inverse data conversion operation, the calculation coefficients of the inverse data conversion section 42 are replaced with integer ratios using the same concept as in the case of FIG. All calculations are performed using integer ratios, and the gain adjustment caused by changing to integer ratios is absorbed in the inverse quantization unit 41. There is a fractional coefficient 315/2 immediately before x3 in FIG. 3, but since the denominator in this case is a power of 2, the actual operation only requires bit shifting. As explained above, in this embodiment, the calculation coefficients of the data conversion section 31, the data conversion section 31, and the inverse data change section 42 are changed to the ratio of integers, and the change in gain is applied to the quantization section and the inverse quantization section. Since the coefficients are absorbed by the quantization section, calculations using coefficients containing errors are performed only once in the quantization section, and other calculations can be performed using ratios of integers that do not include rounding errors. As a result, the data bit width in calculations before quantization can be significantly reduced as shown in Figure 4 compared to the conventional method shown in Figure 14, and high calculation precision can be achieved with a small bus width. Obtainable. Since it is possible to realize a data compression device with high calculation accuracy and a small circuit scale, it is suitable for application to a data compression device that performs image compression or audio compression using DCT.

[0012] In this embodiment, the coefficient is, for example, 7:3.
Although we have shown an example where it is a ratio of integers, it is not limited to this,
It goes without saying that any integer ratio may be used as long as it is expressed as a ratio of integers. Furthermore, any type of data conversion may be used as long as it performs data conversion on a data vector (numerical example), such as FFT, DCT, L
It is applicable to orthogonal data conversion such as OT. Furthermore, although the gain of the coefficients of the data conversion section is absorbed by the quantization section, it goes without saying that any quantization section may be used as long as the gain can be adjusted.

[0013]

[Effects of the Invention] According to the present invention, the arithmetic coefficients of the data conversion arithmetic means are replaced with integer ratios, and the resulting change in gain is adjusted by the quantization means, so that the arithmetic accuracy can be improved. In addition, by reducing the operation bit width, the circuit size can be significantly reduced.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram showing a data conversion section and a quantization section of a data compression device according to the present invention.

FIG. 2 is a diagram for explaining the operation of replacing coefficients with integer ratios in the data compression device according to the present invention.

FIG. 3 is a configuration diagram showing an inverse data conversion section and an inverse quantization section of the data compression device according to the present invention.

FIG. 4 is a configuration diagram showing the bit width of the data converting section of the data compression device according to the present invention.

FIG. 5 is a diagram showing a computing unit for butterfly computation of a conventional data compression device.

FIG. 6 is a diagram showing a computing unit for butterfly computation of a conventional data compression device.

FIG. 7 is a diagram showing a computing unit for butterfly computation in a conventional data compression device.

FIG. 8 is a diagram showing a computing unit for butterfly computation of a conventional data compression device.

FIG. 9 is a diagram showing a computing unit for butterfly computation of a conventional data compression device.

FIG. 10 is a block diagram of a conventional image data processing device.

FIG. 11 is a configuration diagram showing a data conversion section and a quantization section of a conventional data compression device.

FIG. 12 is a diagram showing an arithmetic unit for arithmetic operations in a data conversion section of a conventional data compression device.

FIG. 13 is a diagram showing an arithmetic unit for arithmetic operations in a quantization section of a conventional data compression device.

FIG. 14 is a diagram showing the bit width of a data conversion section of a conventional data compression device.

[Explanation of symbols]

30 Data compression device 31 Data conversion section 32 Quantization section 41 Inverse quantization section 42 Inverse data conversion section

Claims (1)

[Claims] 1. A data conversion operation means for performing data conversion or inverse data conversion operation on a numerical string based on predetermined coefficients, and a quantization operation for replacing continuously changing codes with a finite number of levels. quantization means for performing the data compression operation, wherein the coefficients of the data conversion calculation means are replaced with integer ratios, and the change in gain caused by the replacement with the integer ratio is A data compression device characterized in that the data compression device is adapted to be adjusted by.