JPH08212193A - Data conversion system and device and data transfer system and device using the conversion system and device - Google Patents

Abstract

PURPOSE: To provide a data conversion system/device which can extract the features with optional accuracy and can compress the data against an optional k-dimensional numerical vector and also to provide a data transfer system/device which uses the data conversion system/device. CONSTITUTION: An error evaluation arithmetic unit 13 reads an input vector, a candidate vector and a basis vector out of an input data register 11, a candidate vector memory 12 and a basis vector register 14 respectively. Then the unit 13 calculates the least squares error reduction value against the candidate vector in regard of the input and basis vectors and then adds and holds the candidate vector of the least squares error reduction value into the register 14 as a new basis vector to output it to the outside. A load coefficient calculation device 15 reads an input vector and a basis vector out of the registers 11 and 14 respectively and calculates the load coefficient of the basis vector to output it to the outside.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data conversion method and apparatus for converting a large number of numerical data into different numerical data that can be restored to the original numerical data, and a data transfer method and apparatus using the same. It is about.

[0002]

2. Description of the Related Art As a conventional data conversion method, data conversion using Fourier transform or Hadamard transform has been performed. Although in these conversion represents a k-dimensional numerical vector _{Z = [z 1 z 2 ...} z k] of k numerical data of k previously given k-dimensional basis vector B
_i = [b _{i, 1} b _{i, 2} ... b _{i, k} ] (i = 1, 2,
,, k) as a linear sum of Z = Σ _i a _i B _i (where
For each component, z _j = Σ _i a _i b _{j, i} = a ₁ b _{j, 1} +
a ₂ b _{j, 2} + ... + a _k b _{j, k} ) such that a ₁ , a
₂ ,…, _ak is calculated, and this numerical data set {a
₁ , a ₂ , ..., A _k } are used as the converted numerical data. In this case, the k-dimensional numerical vector Z = [z to be converted
Depending on the value of ₁ z ₂ ... z _k ], some values of a _i may be extremely small values. In such a case, m (m <k) a _i less than k The values alone can be used to represent the original k-dimensional numerical vector. Utilizing this property, the characteristics of the k-dimensional numerical vector can be extracted and data can be compressed. But,
In most cases, k values of a _i are required to represent an arbitrary k-dimensional numerical vector, so that there are cases where features cannot be extracted well or data cannot be compressed well.

In the vector quantization method, a candidate vector is determined in advance, a candidate vector closest to the input data is selected, and the number assigned to the candidate vector is used as the converted numerical data. In this case, since it is converted into one fixed vector, there is a drawback that the conversion error may become large depending on the input data.

[0004]

SUMMARY OF THE INVENTION It is an object of the present invention to provide a data conversion system and apparatus capable of performing feature extraction and data compression with arbitrary accuracy on an arbitrary k-dimensional numerical vector, and data using the same. It is to provide a transfer system and a device.

[0005]

In order to achieve the above object, the present invention is to convert a plurality of numerical data into another plurality of numerical data, and then convert the converted numerical data into the original numerical data. In order to be able to restore the data approximately, a plurality (k) of numerical data before conversion is represented by a k-dimensional numerical vector, and this is composed of m (m <k) k-dimensional numerical vectors. In the data conversion method in which the linear sum of the basis vectors is approximated and the m basis vectors and the corresponding m weighting coefficients are converted, the number of m basis vectors is more than 10 times as large as k. One from each of the n number of k-dimensional numerical vector candidates, a minimum of 2
The k-dimensional numerical vector with the largest amount of power error reduction is selected and selected, and the number of candidate k-dimensional numerical vectors is given a one-to-one corresponding number to select m bases. It is characterized in that each vector number of the vector and each weighting factor are converted into numerical data.

The present invention is also a data converter for realizing the above data conversion method, which comprises an input data register, a basis vector selection device and a weighting factor calculation device, and the basis vector selection device is a candidate vector memory. And an error evaluation arithmetic unit and a base vector register, the input data register holds a plurality of component values of the input vector, and the error evaluation arithmetic unit inputs the input vector from the input data register, the candidate vector from the candidate vector memory, and the base vector. If there are already selected basis vectors from the register, read the basis vectors, calculate the least squares error reduction amount corresponding to the candidate vector with respect to those input vectors and the basis vectors, and calculate the least squares error reduction amount. With the candidate vector that maximizes as the new basis vector The weight coefficient calculation device reads the input vector from the input data register and the basis vector from the basis vector register, calculates the weight coefficient of the basis vector, and outputs the value to the outside. It is characterized by that.

Further, the present invention is a method for generating values of components of n candidate k-dimensional numerical vectors by calculation from each vector number in the above data conversion method. First, the numerical value of the vector number is initialized. The initial value is calculated by the generating function, the initial value is converted by the vector component generating function to calculate the value of the first component of the vector, and the first component value is further converted by the vector component generating function to generate the second component value. Is calculated, and the same calculation is sequentially repeated to sequentially calculate and generate k numerical data up to the k-th component value.

The present invention is also a data conversion device for realizing the above-mentioned data conversion system, which has a vector component generation device, calculates and outputs an initial value by inputting a number given to a vector, With this initial value as input, calculate and output the numerical value of the first component of the vector, and then feed back the first component value to the input and repeat the same operation to output the second component value, which is fed back. By sequentially outputting each component of the vector, n candidate k-dimensional numerical vectors are generated.

Further, in the data transfer system of the present invention, each vector number and each weighting factor of m basis vectors, which are the data converted by the above data conversion system, are input to a decoding device via a communication path and decoded. In the digitizing device, vector component data corresponding to the vector number is read from the candidate vector memory as a base vector, and the sum of products of all the base vectors of the base vector and the corresponding weighting factor is calculated, and the calculated value is the original value. It is characterized by outputting as an approximate value of numerical data.

The present invention is also a data transfer device for realizing the above data transfer method, which comprises the above data conversion device, a communication path and a decoding device, and the decoding device is a candidate vector memory and a base vector register. And a weighting coefficient register and a product-sum calculation device, and the data conversion device converts the input data into each vector number of m basis vectors and each weighting factor and sends them out to the communication channel, and the decoding device to the communication channel. The vector number and the weighting coefficient are received from, the vector number is input to the basis vector register, and the weighting coefficient is input to the weighting coefficient register respectively. In the basis vector register, the vector component data corresponding to the vector number is read from the candidate vector memory and used as the basis vector. The sum-of-products calculation unit temporarily holds the basis vector and weighting factor read from the basis vector register. It calculates the sum of products of the weighting factors read from the register, and is characterized in that outputs the calculated value as an approximation of the input data.

Further, in the data transfer system of the present invention, each vector number and each weighting factor of m basis vectors, which are the data converted by the above-mentioned data conversion system, are inputted to the decoding device through the communication path and decoded. The initial value is calculated by the initial value generation function from the numerical value of the vector number in
The initial value is converted by the vector component generation function to calculate the value of the first component of the vector, the first component value is further converted by the vector component generation function to calculate the second component value, and similar calculations are sequentially performed. Iteratively calculates k numerical data up to the k-th component value, sets the calculated value as a base vector, calculates the sum of products of all the base vectors of the base vector and the corresponding weighting coefficient, and calculates the calculation. The value is output as an approximate value of the original numerical data.

The present invention is also a data transfer device for realizing the above data transfer method, which comprises the above data conversion device, a communication path and a decoding device, and the decoding device is a vector component generation and weighting coefficient register. The data converter converts the input data into each vector number of m basis vectors and each weighting factor and sends out to the communication channel, and the decoding device outputs the vector number from the communication channel. Receives the weighting factor and assigns the vector number to the vector component generator,
The weighting factor is input to each weighting factor register, the vector component generating device calculates the initial value from the vector number by the initial value generating function, and the initial value is converted by the vector component generating function to calculate the first component value of the vector. Then, the first component value is converted by the vector component generation function to calculate the second component value, and the same calculation is repeated successively to sequentially calculate k numerical data up to the kth component value, and the calculated value Is used as the basis vector, and the product-sum calculation unit calculates the product sum of the basis vector read from the vector component generation unit and the weight coefficient read from the weight coefficient register, and outputs the calculated value as an approximate value of the input data. It is a feature.

[0013]

According to the first and second aspects of the invention, the k-dimensional numerical vector Z is Y = a ₁ B ₁ + a ₂ B ₂ + ... + a _j B _j
J basis vectors {B ₁ , B ₂ ,
, B _j } candidates of n (n >> k) k-dimensional numerical vectors {C ₁ , C ₂ , ...
C _n } is prepared, and the least square error reduction amount is calculated one by one from among them, and the one having the largest amount, that is, the approximation error ‖Z−Y‖ can be minimized. A combination of the m number of k-dimensional basis vectors required to finally obtain a desired precision by selecting a dimensional numerical vector {B ₁ , B ₂ , ..., B _m } = {C _s1 , C _s2 , …, C
select _sm }. The number of candidate vectors corresponding to each of the selected m basis vectors {s1,
, sm} and the weight vector W = [a _s1 a _s2 ... _Asm ] having each weighting factor as a component are converted into numerical data, which is a data conversion method. Since a large number of basis vectors are prepared in this conversion method, for any k-dimensional numerical vector Z, only m basis vectors less than k are used with sufficient accuracy.
A combination of basis vectors that can be approximated to Z−Y∥˜0 can be selected from {C ₁ , C ₂ , ..., C _n }.

According to the third and fourth aspects of the invention, the n component candidate vectors (k-dimensional numerical vectors {C ₁ , C ₂ , ...
C _n } is generated by calculation from the numerical value of the number given to each vector of the value of each vector component. That is, C _i = [c _{i, 1} c _{i, 2} ... C _{i, k}
], The initial value generation function x = g (i) and the recursive function y = f (x) are expressed, and x = g (i) is calculated by inputting the vector number i in the vector component generation device and the initial value is calculated. Output x. Then, with x as input, c _{i, 1} =
f (x) is calculated and the first component c _{i, 1} is output.
c _{i, 1} is input and c _{i, 2} f (c _{i, 1} ) is used to output the second component c _{i, 1} and such iterative calculation is performed for the k th component c _{i, k.}
The values of the k components are determined by performing successive calculations up to = f (c _{i, k-1} ). In this way, the candidate n
It is not necessary to store all the numerical data of n × k components as k-dimensional numerical vectors {C ₁ , C ₂ , ..., C _n }, and if the numbers si are specified, the functions f and g are Just by storing it, each basis vector {B ₁ , B ₂ , ..., B
The value of the component of _m } = {C _s1 , C _s2 , ..., C _sm } can be set. Various functions can be applied as the function f. One example is a function that gives a pseudo-random number, which may be particularly effective.

According to the fifth and sixth aspects of the present invention, the data converted by the data conversion method according to the first aspect is input to the decoding device via the communication path, and the approximate value of the original input data is output. System and equipment. That is, the vector number of the base vector and the weighting factor are transferred, the vector data corresponding to the vector number is read from the candidate vector memory in the decoding device, the product sum with the weighting factor is calculated, and the original before conversion is performed. Output the approximate value of numerical data. Conventionally, only the vector number is transferred, or the basis vector is fixed and only the weighting factor is transferred. In this method, since the approximately optimal combination of basis vectors and the weighting factor for them are transferred, it is possible to efficiently transfer compressed data to arbitrary data.

According to the seventh and eighth aspects of the present invention, the data converted by the data conversion method according to the third aspect is input to the decoding device via the communication path, and the approximate value of the original input data is output. System and equipment. Although the data transfer method and the apparatus according to the fifth and sixth aspects are similar, the vector data corresponding to the vector number is not read from the memory but is generated by calculation as a function of the number. That is, the apparatus has a basis vector component generation device, and the corresponding vector data is output by performing a function calculation using the vector number as an initial input value. Then, the product-sum calculation device calculates the product sum of the vector component corresponding to the vector number and the corresponding weighting factor, and outputs the approximate value of the original numerical data before conversion. Therefore,
Compared with the data transfer system and device according to claims 5 and 6,
The device can be made smaller because it does not require a large memory.

[0017]

Embodiments of the present invention will now be described in detail with reference to the drawings. (Embodiment 1) The operation of the apparatus according to the present invention will be described in detail with reference to the embodiment of FIG.

The configuration of the data conversion device 100 of this embodiment comprises an input data register 11, a candidate vector memory 12, an error evaluation calculation device 13, a basis vector register 14, and a weighting factor calculation device 15. Basis vector selection device 1
01 is the candidate vector memory 12 and the error evaluation calculation device 1
3 and the base vector register 14. The input data register 11 stores the input vector Z, and the candidate vector memory 12 stores n (n >> k) vectors {C ₁ ,
Store C ₂ , ..., C _n }. The error evaluation calculation unit 13 evaluates the error ∥Z−Y∥ when one of the candidate vectors is added as a base vector, and selects the candidate vector predicted to have the smallest error as one of the base vectors. The base vector register 14 is the error evaluation computing device 1
Basis vector selected in 3 {B ₁ , B ₂ , ..., B _m }
, {C _s1 , C _s2 , ..., C _sm } are stored and a combination of vector numbers {s1, s2, ..., sm} is output. Load coefficient calculation unit 15 minimizes the square error ‖Z-Y‖ of Z and _{Y Y = a 1 B 1 +} a 2 B 2 + ... + a m B m of the load vector _{W = [a s1 a s2 ...} a _sm ] and output. However, ‖Z−Y‖ = Σ _{j = 1} ^k (z _j −y _j ) ² .

The detailed operation of the error evaluation calculation device 13 is shown in FIG. 2 as a flow chart. (Step 1) First, the number j of the basis vector is set to 0. (Step 2) The value of j is increased by 1, and the number of trials Nt of the candidate vector selection is increased.
And the comparison value Δ (Cg) of the least square error reduction amount is set to zero. (Step 3) Increase the value of Nt by 1.
(Step 4) Select one Ci from the candidate vectors.
(Step 5) The minimum squared error reduction amount Δ (C
Calculate i). (Step 6) Δ (Ci) is Δ (C
If it is larger than g), go to step 7, otherwise jump to step 8. (Step 7) Let the candidate vector Ci be the most influential candidate Cg of the base vector. (Step 8) If the number Nt of trials is less than the limit Nc of number of trials, the process returns to step 3, otherwise the process proceeds to step 9. (Step 9) C in which the least squared error reduction amount is maximum
Register g as the basis vector Bj. (Step 1
0) Y = a ₁ B ₁ + a ₂ B ₂ + ... + a _j B _j error ‖
The minimum value of Z−Y∥ is evaluated, and if the value is not smaller than the evaluation reference Ec, the process returns to step 2, and if it is smaller, the selection of the base vector is ended. The required basis vectors are selected one by one based on this flowchart.

FIG. 3 shows a flowchart for evaluating the least square error reduction amount Δ (Ci) with respect to Ci in (Step 5). First, in (step 5-1), if j = 0 regarding the base vector number j (step 5-
4) and if j ≠ 0, proceed to (step 5-2). In (step 5-2), the j basis vectors {B ₁ , B ₂ , ..., B _j } already selected are column vectors (k × j) matrix X _j = [B ₁ B ₂ ... B _j ] is set. Then, in step 5-3, the projection matrix Pc _j
= I−X _j (X _j ^T X _j ) ⁻¹ X _j ^T is calculated. However,
X _j ^T is a transposed row example in which the rows and columns of X _j are exchanged, (X _j ^T
X _j ) ⁻¹ is the inverse matrix of the matrix X _j ^T X _j . However,
In (step 5-4), the projection matrix Pc _j = I is set.
After that, in (step 5-5), the least square error reduction amount is Δ (C _i ) = (Z ^T Pc _j C _i ) ² / (C _i ^T Pc _j
C _i ) is calculated according to the vector and matrix operation rules. The least squared error reduction amount is an error ‖Z− when a candidate vector C _i to be selected next is added as a base vector.
It is an amount representing the amount of decrease in the minimum value of Y /. In addition,
The minimum value of the error ‖Z−Y‖ is calculated as Pc _j Z.

The basis vectors {B ₁ , B ₂ , ..., B _m } = {C _s1 , C _s2 , ...
C _sm } is registered in the base vector memory. In the load coefficient calculation device 15, the load vector W is obtained by calculation with W = (X _j ^T X _j ) ⁻¹ X _j ^T Z, and W = [a _s1 a _s2 ...
a _sm ] is output. (Example 2) In Example 2, a plurality of input vectors Z (Nz
, The same basis vector {B ₁ , for any of the converted vectors {Y ₁ , Y ₂ , ..., Y _Nz } of each vector {Z ₁ , Z ₂ , ..., Z _Nz }. This is an embodiment for conversion as a linear sum of B ₂ , ..., B _m }. in this case,
The configuration and operation are the same as in the first embodiment except for (step 5-5), so only (step 5) is shown in FIG. In (Step 5-5 ′) in the second embodiment, the candidate vector C
_The least squared error reduction amount for _i is Δ (C _i ) = Σ _{q = 1} ^Nz
It is determined as (Z _q ^T Pc _j C _i ) ² / (C _i ^T Pc _j C _i ). This Δ (C _i ) is the square error of each vector ‖Z _q
The sum of −Y _q ‖ Σ _{q = 1} ^Nz ‖ Z _q −Y _q ‖ is an amount that corresponds to the minimum amount of decrease, and by selecting the vector C _i that maximizes this amount of decrease, each vector is selected. It is a criterion for selecting the most suitable basis vector in the transformation of. In this way, a common basis vector suitable for Nz input vectors can be selected. (Embodiment 3) The operation of the apparatus according to the present invention will be described in detail with reference to the embodiment of FIG. The data conversion device 500 of the present embodiment is different from the configuration of the data conversion device 100 of FIG. 1 in that the candidate vector memory 12 in the basis vector selection device 101 is replaced with the candidate vector register 51 and the vector component generation in the basis vector selection device 501. Only the device 52 is provided. That is, the value of the component of the candidate vector is not stored in the memory as it is, but the value of the component of the candidate vector is output from the number of the candidate vector by the vector component generation device 52 and is temporarily stored in the candidate vector register 51. It is configured to be retained in. Therefore, only the operations of the candidate vector register 51 and the vector component generating device 52 will be described below.

First, a vector number is generated from the candidate vector register 51 and sent to the vector component generator 52.
As shown in FIG. 6, the configuration of the vector component generation device 52 includes an initial value generation unit 61 and a vector component calculation unit 62.
The initial value generation unit 61 uses the candidate vector C _i = [c _{i, 1} c
_{For i, 2} ... C _{i, k} ], an initial value c _{i, 0} = g (i) is calculated from the number i using a function g. The vector component calculation unit 62 uses the recursive function f to set the value of the j-th component of C _i as c _{i, j} = (c _{i, j−1} ) and j = 1 to j = k.
Generate up to calculation. A flowchart of this operation is shown in FIG. In (step 1), the initial value c _{i, 0} = g (i) is calculated with j = 0. Next, in (step 2), j is incremented by 1 and the value of the j-th component of C _i is generated by calculation with c _{i, j} = (c _{i, j-1} ). In (step 3), if j is smaller than k, the process returns to step 2, and if j is not smaller, the process ends. The value of the component thus calculated is the candidate vector register 5
It is temporarily held at 1 and is output to the error evaluation calculation device 13.

As the functions g (x) and f (x), functions having a finite range of functions and strong non-linearity are suitable. For example,
Function px mod that is often used for generating pseudo random numbers
It is preferable to use q (remainder when px is divided by q), a sine function sin (px), a cosine function cos (px), or the like among trigonometric functions. However, p and q are constants, and p is preferably a number sufficiently larger than 1. When the device is configured with a digital computer and the remainder calculation of the function px mod q is performed, if the divisor q is set to 2 ^s , the value of the product px is extracted as the function value output by extracting only the value of the lower s bits. Therefore, complicated division becomes unnecessary and high-speed arithmetic processing is possible. (Embodiment 4) FIG. 8 shows an embodiment of a candidate vector component generation device 52 utilizing the remainder calculation of the function px mod q. However, q = 2 ^s . The configuration is composed of a multiplication device 81 and an input switching device 82 that perform px calculation. Multiplier 8
The coefficient is input to the input IA of 1 and the output of the input switching device 82 is connected to IB. The vector number is input to one input of the input switching device 82, and the output of the multiplication device 81 is connected to the other input. The values of the components of the candidate vector are obtained from the output of the multiplier 81.

To operate this, first, the coefficient Pi is input to IA, the vector number Si is selected by the input switching device 82, and this value is input to IB. After this, the multiplication device 8
1 calculates the product of Pi and Si, and outputs the result of the lower S bits ignoring the upper bits. This value is C _{i, 0} . Next, the coefficient P is input to IA, the input switching device 82 selects the output C _{i, 0} of the multiplication device 81, and this value is input to IB. Then, the multiplication device 81 calculates the product of P and C _{i, 0} , and the result of the lower S bits ignoring the upper bits is C
Output as _{i, 1} . After this, similarly, C _{i, 2} , C
_{i, 3} , ..., C _{i, k} are calculated and output one after another. (Embodiment 5) The operation of the apparatus according to the present invention will be described in detail with reference to the embodiment of FIG. The data transfer device of the present embodiment is a device that inputs the data converted by the data conversion method according to claim 1 to a decoding device via a communication path and outputs an approximate value to the original input data. That is, the data conversion device 100, the communication path 91, and the decoding device 92 of FIG. 1 are included. The data conversion device 100 uses the vector Z
Is the vector number of the basis vector {s1, s2, ..., S
m} and the weighting factor {a _s1 a _s2 ... _Asm } are combined and output to the communication path 91. Decoding device 9
2 receives the vector number and the weighting factor from the communication path 91, and inputs the vector number to the base vector register 93 and the weighting factor to the weighting factor register 94, respectively. In the base vector register 93, vector numbers {s1, s2, ...,
The vector data corresponding to sm} is read from the candidate vector memory and the basis vector {C _s1 , C _s2 , ..., C _sm }.
Is held temporarily. However, the candidate vector memory 95
1 is the same as that of the candidate vector memory 12 of FIG. The product-sum calculation unit 96 calculates the product sum of the vector corresponding to the vector number and the corresponding weighting factor as the vector Y = a _s1 C _s1 + a _s2 C _s2 + ... + a _sm C _sm and outputs it. This vector Y is an approximate value of the original numerical vector Z before conversion. (Embodiment 6) The operation of the apparatus according to the present invention will be described in detail with reference to the embodiment of FIG. The data transfer device 500 of the present embodiment is a device that inputs the data converted by the data conversion method of claim 3 to the decoding device via the communication path and outputs the approximate value of the original input data.
That is, the data conversion device 500 and the communication path 11 of FIG.
1. Decoding device 112. Data converter 500
Is the vector number of the base vector {s1, s
, ..., Sm} and the weighting factor {a _s1 a _s2 ... A _sm } are converted into a combination and output to the communication path 111.
The decoding device 112 receives the vector number and the weighting factor from the communication path 111, and inputs the vector number to the vector component generating device 113 and the weighting factor to the weighting factor register 114, respectively. The vector component generation device 113 is the same as the vector component generation device 52 of FIG. 5, and generates the component data of the basis vector corresponding to the vector number {s1, s2, ... The basis vectors {C _s1 , C _s2 , ..., C _sm } are temporarily held. The product-sum calculation unit 115 calculates the product sum of the vector corresponding to the vector number and the corresponding weighting factor as vector Y = a _s1 C _s1
+ A _s2 C _s2 + ... + a _sm Calculate as C _sm and output. This vector Y is an approximate value of the original numerical vector Z before conversion.

As described above, according to the _first and _second aspects of the present invention, an arbitrary k-dimensional numerical vector Z is a linear sum of basis vectors {B ₁ , B ₂ , ..., B _m } and Z to Y. =
_{a 1 B 1 + a 2 B} 2 + ... + a m When converting to B _m, the appropriate basis vectors {B ₁ from a large number of candidate vectors,
Since B ₂ , ..., B _m } are selected as necessary, it is possible to perform conversion with a small error ‖Z−Y‖ with sufficient accuracy by using m basis vectors that are less than k. In addition, as shown in the second embodiment, a plurality of vectors {Z ₁ , Z ₂ ,
Since it is possible to select an appropriate common basis vector for ..., Z _q }, in particular, each vector of {Z ₁ , Z ₂ , ..., Z _q } is a very similar vector and the distribution is biased. In some cases, the number of base vectors required is smaller, the storage amount of base vectors and the amount of calculation for conversion can be reduced, which is effective for data compression.

According to the inventions of claims 3 and 4,
A data table in which all the components of _m basis vectors {B ₁ , B ₂ , ..., B _m } are recorded is not required, and the memory amount can be significantly reduced, which is useful for downsizing the device.

According to the fifth and sixth aspects of the invention, the data transfer amount can be reduced by performing the data transfer by utilizing the invention of the first aspect. According to the inventions of claims 7 and 8, the amount of memory can be significantly reduced as compared with the case of the invention of claim 5, by performing data transfer using the invention of claim 3. Useful for downsizing the device.

[0028]

As described above, according to the present invention, a data conversion system and apparatus capable of performing feature extraction and data compression with respect to an arbitrary k-dimensional numerical vector with arbitrary accuracy, and the device are used. A data transfer system and device can be provided.

[Brief description of drawings]

FIG. 1 is a configuration explanatory view showing an embodiment of claims 1 and 2 of the present invention.

FIG. 2 is a flowchart showing an example of basis vector selection of the present invention.

FIG. 3 is a flowchart for evaluating a least square error reduction amount Δ (Ci) according to the first embodiment of the present invention.

FIG. 4 is a flowchart for evaluating a least square error reduction amount Δ (Ci) according to the second embodiment of the present invention.

FIG. 5 is a structural explanatory view showing an embodiment of claims 3 and 4 of the present invention.

FIG. 6 is a configuration explanatory diagram showing an example of a candidate vector component generation device of FIG.

FIG. 7 is a flowchart showing an example of the operation of FIG.

8 is a configuration explanatory diagram showing an example of a candidate vector component generation device in FIG.

FIG. 9 is a structural explanatory view showing an embodiment of claims 5 and 6 of the present invention.

FIG. 10 is a structural explanatory view showing an embodiment of claims 7 and 8 of the present invention.

[Explanation of symbols]

11 ... Input data register, 12 ... Candidate vector memory, 13 ... Error evaluation calculation device, 14 ... Basis vector register, 15 ... Weighting factor calculation device, 51 ... Candidate vector register, 52 ... Vector component generation device, 61 ... Initial value generation Part, 62 ... Vector component calculation part, 81 ... Multiplier, 8
2 ... Input switching device, 91 ... Communication path, 92 ... Decoding device,
93 ... Basis vector register, 94 ... Weighting coefficient register, 95 ... Candidate vector memory, 96 ... Sum of products arithmetic unit,
100 ... Data conversion device, 101 ... Basis vector selection device, 111 ... Communication channel, 112 ... Decoding device, 113 ... Vector component generation device, 114 ... Weighting coefficient register, 11
5 ... Sum of products operation device, 500 ... Data conversion device, 501 ...
Basis vector selection device.

Claims (8)

[Claims] 1. Before converting a plurality of numerical data so that the original numerical data can be approximately restored from the converted numerical data after the data is once converted into another plurality of numerical data. A plurality of (k) numerical data of is represented by a k-dimensional numerical vector, and this is approximated by a linear sum of basis vectors consisting of m (m <k) k-dimensional numerical vectors. In the data conversion method for converting m basis vectors and m corresponding weighting coefficients, 1 is selected from among n number of k-dimensional numerical vector candidates in which m basis vectors are more than 10 times as many as k. One by one, minimum 2
The k-dimensional numerical vector with the largest amount of power error reduction is selected and selected, and the number of candidate k-dimensional numerical vectors is given a one-to-one corresponding number to select m bases. A data conversion method characterized in that each vector number of a vector and each weighting factor are converted into numerical data. 2. A data conversion device for realizing the data conversion method according to claim 1, comprising an input data register, a basis vector selection device, and a weighting factor calculation device, and the basis vector selection device is a candidate vector. An input data register holds a plurality of component values of an input vector, and an error evaluation arithmetic unit receives an input vector from an input data register, a candidate vector from a candidate vector memory, and a base vector register. If there is already a selected basis vector from the vector register, read that basis vector, calculate the least squared error reduction amount corresponding to the candidate vector with respect to those input vectors and basis vectors, and reduce the least squared error reduction. The candidate vector with the maximum amount is set as a new basis vector The weighting factor calculator reads the input vector from the input data register and the basis vector from the basis vector register, calculates the weighting factor of the basis vector, and outputs the value to the outside. A data conversion device characterized by the above. 3. The data conversion method according to claim 1, wherein a value of a component of n candidate k-dimensional numerical vectors is generated from each vector number by calculation, and the numerical value of the vector number is first initialized. The initial value is calculated by the value generation function, the initial value is converted by the vector component generation function to calculate the value of the first component of the vector, and the first component value is converted by the vector component generation function and the second component is calculated. A data conversion method characterized in that a value is calculated, and the same calculation is sequentially repeated to sequentially calculate and generate k numerical data up to the k-th component value. 4. A data conversion device for realizing the data conversion method according to claim 3, further comprising a vector component generation device, calculating an initial value by inputting a number assigned to the vector, and outputting the initial value. , Input the initial value to calculate and output the value of the first component of the vector, and then feed back the first component value to the input and repeat the same operation to output the second component value. A data conversion device characterized by generating n candidate k-dimensional numerical vectors by successively outputting each component of the vector. 5. The vector number and the weighting factor of m basis vectors, which are data converted by the data conversion method according to claim 1, are input to a decoding device via a communication path, and the decoding device The vector component data corresponding to the vector number is read from the candidate vector memory as a basis vector, the sum of products for all basis vectors of the basis vector and the corresponding weighting factor is calculated, and the calculated value is the original numerical data. A data transfer method characterized by outputting as an approximate value. 6. A data transfer device for realizing the data transfer method according to claim 5, comprising the data conversion device according to claim 2, a communication path and a decoding device, wherein the decoding device is a candidate vector. It has a memory, a base vector register, a weighting coefficient register, and a product-sum calculation device, and the data conversion device converts the input data into each vector number of m base vectors and each weighting factor, and sends it out to the communication path for decoding. The digitizer receives the vector number and the weighting factor from the communication channel, inputs the vector number to the basis vector register, and inputs the weighting factor to the weighting factor register, respectively. In the basis vector register, the vector component data corresponding to the vector number is input from the candidate vector memory. It is read and temporarily stored as a basis vector, and the multiply-accumulate operation unit is loaded with the basis vector read from the basis vector register. A data transfer device, which calculates a product sum with a weighting coefficient read from a weighting coefficient register and outputs the calculated value as an approximate value of input data. 7. The vector number and each weighting factor of m basis vectors, which are data converted by the data conversion method according to claim 3, are input to a decoding device via a communication path, and the decoding device The initial value is calculated from the numerical value of the vector number by the initial value generation function, the initial value is converted by the vector component generation function to calculate the value of the first component of the vector, and the first component value is further calculated by the vector component generation function. Convert and calculate the second component value, and repeat the same calculation one by one to sequentially calculate the k numerical data up to the kth component value, and use the calculated value as the basis vector, and the basis vector and the corresponding load A data transfer method characterized by calculating the sum of products for all basis vectors with coefficients and outputting the calculated value as an approximate value of the original numerical data. 8. A data transfer device for realizing the data transfer method according to claim 7, comprising the data conversion device according to claim 4, a communication path and a decoding device, wherein the decoding device is a vector component. It has a generation / weighting coefficient register and a product-sum calculation device, and the data conversion device converts the input data into each vector number of m basis vectors and each weighting factor and sends out to the communication path, and the decoding device communicates. The vector number and the weighting factor are received from the road, the vector number is input to the vector component generating device, and the weighting factor is input to the weighting factor register, and the vector component generating device calculates the initial value from the vector number by the initial value generating function. Calculate the first component value of the vector by converting the initial value with the vector component generation function,
Further, the first component value is converted by the vector component generation function to calculate the second component value, and the same calculation is repeated successively to sequentially calculate the k numerical data up to the kth component value, and the calculated value is the basis. A vector, the product-sum calculation device calculates a product sum of the basis vector read from the vector component generation device and the weight coefficient read from the weight coefficient register, and outputs the calculated value as an approximate value of the input data. Data transfer device.