JPH04156181A - Vector quantization method and vector quantizer - Google Patents

Abstract

PURPOSE:To retrieve an optimum code book vector with a few number of times of calculation, high efficiency and at a high speed by providing a distortion upper limit setting process, a lower limit setting process, a storage comparison process, a partial distortion calculation process and an upper limit revision process to the quantizer. CONSTITUTION:A code book generating section 1 generates a code book 100 based on a training series and a sort table 200. Then an optimum vector retrieval section 2 retrieves an optimum code book vector. A distortion calculation section 22 calculates distortion of a code book vector Y and an input vector X based on the code book 100 and sets the sum of distortion of all dimensions as an upper limit. Since the distortion is calculated in the order of larger distortion, the calculation is stopped with a few number of calculation. Thus, the optimum code book vector is obtained efficiently.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to digital information processing technology, and particularly to a vector quantization method and a vector quantization device.

(Prior Art) Generally, when images, sounds, etc. are expressed as a function, quantization is a discretization operation regarding a specific value range. That is, the continuous function on the sample points obtained by sampling is expressed by converting it into a finite number of values. A specific method is a vector quantization method.

In this vector quantization method, for example, an input image is divided into a plurality of blocks and quantization is performed on one vector whose components are the pixels of each block.For example, for a specific block, Several representative image patterns (representative vectors) that are considered to occur frequently are determined, assigned identification codes, and stored in the encoder. This is called a codebook, and when an input signal is returned, the representative vector most similar to the pattern is searched from the codebook for each block, and its identification code is sent out. It is also applied to audio etc. in the same way.

A vector quantization device that performs such vector quantization uses, for example, a tree search method (tree search method) that uses a structured codebook to reduce the number of vectors to be searched for the purpose of speeding up the processing. A.Buzo, A.H.
,Gray, Jr., ,R.M., Gray, and J.
D. Markel: “5peech codin
g based upon vector quantity
zation+, IEEE Trans, Acoust
,,5peech &Signal Process,
, ASSP-28, 10, pp. 5B2-574. Oc
t.

1980) has been proposed.

Among the methods for reducing the number of distortion calculations used for vector quantization, there is a method of changing the search order according to the probability of occurrence of the vector (Japanese Patent Laid-Open No. 1-218
279 Publication), and a method using the average value within the search truncated vector (Kano, Aizawa, Harashima: “Fast search method for orthogonal transform domain for vector quantization, IEICE, pp. 1102-1
110, Vol J69-B.

No. 10.198B), etc. have been proposed to speed up vector quantization.

(Problem to be Solved by the Invention) However, among the vector quantization methods mentioned above, the tree search method is faster than other methods in terms of distortion calculation, but when the input vector is at the boundary of the subspace. (``Fast search method for orthogonal transform domain for vector quantization'' P,
1103 (see upper right corner)), there is a drawback that the optimal vector cannot be obtained from the codebook. Therefore, the tree search method may lack accuracy in terms of minimizing distortion.

Furthermore, even in the method of reducing the number of distortion calculations, it is necessary to search all codebook vectors at least once, or it is necessary to perform some kind of processing.

In the conventional search methods described above, depending on the properties of the resulting input vector, it is optimal to search for a coatbook vector corresponding to that vector.

It was not possible to satisfy the requirements of high efficiency and high speed.

Therefore, the present invention provides, for any input vector,
It is an object of the present invention to provide a vector quantization method that can search for an optimal codebook vector efficiently and at high speed with a small number of calculations, and a vector quantization device that executes the method.

(Means for Solving the Problem) In order to achieve the above object, the present invention, as a first means,
A lower limit value setting step in which the sum of the minimum distortions given by unsearched codebook vectors is set as the lower limit value of distortion for each dimension of the codebook vector and the input vector, and a lower limit value setting step for each dimension of the codebook vector and human vector. an upper limit value setting step in which the minimum distortion given by the codebook vector at the end of the search is set as the upper limit value of distortion; a storage comparison step in which the lower limit value and the upper limit value are stored and compared; a search termination step in which the search is terminated when the upper limit values are compared and they match; and when the search is determined not to have been completed in the search termination step, the code blur vectors and input vectors are sequentially selected from the low-dimensional components up to the predetermined order components. a distortion partial sum calculation step of calculating a partial sum of distortion; and a distortion calculation process that compares the partial sum with the upper limit value and aborts further distortion calculations when the partial sum is larger than the upper limit value. A vector quantization method can be provided by an initial step and an upper limit updating step of updating the partial sum as a new upper limit when the partial sum is equal to or smaller than the upper limit.

As a second means, for a vector consisting of multiple dimensions forming a codebook, there is a classification means for classifying a unique component value for each dimension as a base value, and a classification means for classifying a vector consisting of multiple dimensions forming a codebook as a base value, and a classification means for classifying a unique component value for each dimension as a base value. lower limit value setting means for calculating the sum of all dimensions of distortion with base values belonging to the codebook vector; storage comparison means for storing and comparing the lower limit value and the upper limit value; search terminating means for terminating the search when the value and the upper limit value are compared and when they match; and when the search is determined to be unfinished by the search terminating step, codebook vectors and input vectors sequentially from low-dimensional components to predetermined dimensional components; distortion partial sum calculation means for calculating a partial sum of distortion; and a distortion that compares the partial sum with the upper limit value and aborts further distortion calculation when the partial sum is larger than the upper limit value. It is possible to provide a vector quantization device having calculation aborting means and upper limit updating means for updating the partial sum as a new upper limit when the partial sum is equal to or smaller than the upper limit.

(Operation) According to the vector quantization method and vector quantization device configured as described above, it is possible to efficiently and quickly search for an optimal codebook vector for all input vectors.

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

FIG. 1 is a block diagram showing the configuration of a high-speed vector quantization device according to a first embodiment of the present invention.

The configuration of this high-speed vector quantization device is roughly divided into a codebook generation section 1 and an optimal vector search section 2.

In the codebook generation section 1, a training section 10 first generates a codebook 100 based on a given training sequence. Based on the codebook 100, the sorting unit 11 classifies (sorts) each component into base values (unique vector component amounts) and creates a sorting table 200.
At the same time, the base value number is added to the corresponding vector component of the codebook 10.

When input vector number 1 A distortion calculation unit 22 includes a dimensional distortion calculation unit 22a that calculates distortion for each dimension of , and a partial distortion calculation unit 22b that calculates a partial sum of the distortion.
, a lower limit calculation unit 23 that calculates the lower limit value of distortion using the sorting table 200, a memory 24 that stores the partial distortion amount, the lower limit value, and the minimum distortion amount as the upper limit value, and a partial sum of the upper limit value and the distortion. The comparison unit 25 uses the upper limit value and the lower limit value to determine the end of the search.

Further, in the component calculation order determining unit 21, the sorting table 20
Using the component average value of 0, the component average value and the input vector
The component calculation order table 300 is created so that calculations are performed in descending order of the difference from the components.

) Further, FIGS. 2(a) and 2(b) show schematic diagrams as an example of the format of the codebook 100 and sorting table 200 described above.

I FIG. 2(a) is composed of component values and base value numbers of codebook vectors divided by the number of dimensions and the number of codebook vectors. FIG. 2(b) is composed of base values and vector numbers classified and arranged according to base value numbers and dimension numbers corresponding to the codebook vectors.

[In Figure 2 (C), when the th order component of the codebook vector
) Jo, and the pointer indicating the base value closest to the input vector X (nearest base value) is "p".

It is indicated by rqJ. Then, the difference between the input vector X and the base value of the codebook vector is calculated as Ds(k) and Db(
k), the difference D5(k) represents the difference on the smaller side than X.

Next, the operation of the vector quantization apparatus of the first embodiment configured as described above will be explained. Here, the number of codebook vectors is "N", and the number of dimensions of the vector is rMJ
and the codebook vector indicated by index l is expressed as r Y i J.

First, the codebook generation unit 1 as a preprocessing unit generates a codebook 100 and a sorting table 20 based on the training sequence.
Generates 0.

Next, the optimal vector searching unit 2, which is an encoding unit, searches for an optimal codebook vector according to the flowchart shown in FIG. This procedure will be explained one by one using the constituent members and their reference numbers shown in FIG.

First, when the input vector A component calculation order table 300 is created so that component calculations are performed in order of the largest difference (hereinafter,
To simplify the explanation, the component that gives the largest difference is 1
Next, it is assumed that the components are rearranged in numerical order so that the component that gives the smallest difference becomes the M order.)

Next, the distortion calculation unit 22 extracts the codebook vector Y1 corresponding to index 1 from the codebook 100, calculates the distortion with the input vector Calculate the corresponding index min (here, it is set to 1)
is stored in the memory 25 (step Sl). That is,
The sum U of distortions in all dimensions is set as the upper limit.

On the other hand, the lower limit calculation unit 23 uses the sorting table 200 to
The differences D5(k) and Db(k) between the input vector X and the base value of the codebook vector are shown below (2).

Calculate using equation (3).

Ds(k)=(X(k) −Vk(p))...
(2) pb(k)-(Vk(q)-X(k))
-(3) However, as shown in Figure 2 (C), k represents the dimension of the vector, p and q are pointers indicating the base value closest to X (nearest base value), and Ds is the pointer that indicates the base value closest to X. represents the difference on the larger side.

Then, the lower limit value is expressed by the following formula (4), is calculated and stored in the memory 24 (step S2).

Subsequently, the comparison unit 25 compares the upper limit value U and the lower limit value stored in the memory 24 (step S3
).

At this time, if the upper limit value U and the lower limit value match (Yes)
, the upper limit value U is the minimum distortion for all codebook vectors, so the index min is output and the process ends (step S4).

On the other hand, if the upper limit value U and the lower limit value do not match (No),
The index i is incremented for comparison with the next codebook vector (step S5). That is, "1" is added to index i, where the jth component is "1".

The distortion calculation unit 22 extracts the codebook vector Yi from the codebook U00 with index 1, and calculates the distortion with the input vector X (step S6). Here, the partial sum Sj of distortion from the low-dimensional component to the j-th component is determined from the following equation (%) and sent to the comparison unit 25. The comparison unit 25 compares the partial sum Sj with the upper limit value U stored in the memory (step S7), and determines that Sj≧U
...When (6) is satisfied (Yes)
, aborts further distortion calculations.

That is, when the distortion calculation is terminated, the index i+1 corresponding to the next codebook vector is sent to the distortion calculation unit 22, and the target is shifted to the next codebook vector. Also, since the vector indicated by the index l does not give the minimum distortion (upper limit U), the number of vectors belonging to the base value corresponding to this vector is decreased by one (step s8).
, the process moves to step 513, which will be described later.

On the other hand, if the calculated distortion value cannot be truncated (No), the order j of the vector and the number of dimensions M are compared (step S9).
.

Here, if it is smaller than the vector dimension M (No),
Add "1" to the order j of the vector (step 510)
, return to step s6. That is, it is sent to the distortion calculation section 22 (step S10).

Furthermore, even if the vector order j becomes equal to the dimension number M, if the distortion partial sum Sj is smaller than the previous minimum distortion, then (
Yes), it is found that the vector indicated by the index min does not give the minimum distortion, so the number of vectors belonging to the base value corresponding to this vector is decreased by one (step 511).

This minimum distortion is stored in the memory 24 as a new upper limit value U, and the index i at this time is set as a new index mi.
n in the memory 24 (step 512).

In this way, the number of vectors belonging to the base value corresponding to the codebook vector that was found not to give the minimum distortion was decreased by 1, and the nearest base value to which the number of vectors belonging became "0" It is determined whether or not there is (step 813).

Here, if there is a nearest base value of "0" (Yes
), the corresponding nearest base value pointer rpJ or “q
'' to the base value to which the vector belongs (step 514), and the process returns to step S2 to recalculate the lower limit value.

On the other hand, if the number of vectors belonging to the nearest base value does not become "0", the lower limit value remains unchanged and step S3
Processing continues.

The procedure described above is repeated until the upper limit value U and the lower limit value coincide.

From the above, it is possible to obtain the same optimal solution as the optimal solution obtained by performing a complete search of the codebook vectors, and also to terminate the search midway. In addition, since the distortion calculation is adapted to the input vector and calculated in order of components that are likely to have a large distortion, the partial distortion can be maximized with the minimum number of calculations, making it possible to abort the calculations with a small number of calculations. As a result, the optimal codebook vector for the input vector can be efficiently obtained with a small number of calculations. Also, the distortion between the nearest base value and the input vector component in the lower limit calculation unit 23 is as follows.

Since each component can be calculated completely independently, it is possible to further speed up the calculations by processing them in parallel.

Next, FIG. 4 shows the configuration of a vector quantization device as a second embodiment of the present invention. The configuration of this device is composed of a codebook vector generating section 2 and an optimum vector searching section 2, as in the first embodiment.

This codebook vector generation unit 1 includes a network 40 that adds the training unit 10 and the sorting unit 11 of the first embodiment, and outputs an index corresponding to the vector when a vector of the codebook 100 is manually generated.
It is constituted by a network generation unit 12 that generates 0. The optimal vector search unit 2 uses the network 400 when the input vector The configuration includes a search order determination section 20 that determines the search order for the codebook 100.

Next, the operation of the apparatus having such a configuration will be explained.

In the second embodiment described above, the operation of the network generation section 12 of the codebook generation section 1 as a preprocessing section will first be explained with reference to the flowchart shown in FIG. 5(a).

That is, in the network generation section 12,
According to the flowchart in a), the backpropagation method (D.E., Rumelhart, G.E., Hint.
On.

and R1, Williams: Learn
g representations by back-
propagating errors, Nature
, Vol, 323-9゜pp, 533-538.198
A network 400 compatible with cordons, such as the configuration shown in FIG. 5(b), learned by
is generated.

This network 400 has an input layer consisting of a plurality of input units for manually inputting an input vector, an intermediate layer consisting of a plurality of intermediate units each connected to all the input units, and an intermediate layer each connected to all the intermediate units. and a plurality of output units corresponding to the number of connected codebook vectors.

Such a combination is a simple hierarchical combination in which one cordon vector corresponds to one output unit. In addition to such hierarchical connections, it is also possible to use some networks typified by neural networks, such as hierarchical connections in which units are mutually connected within each layer.

Next, a method for forming a network will be explained with reference to the flowchart of FIG. 5(a). Here, it is assumed that the initial value of the network coupling coefficient is set using a random value or the like.

First, index i is initialized to "1" (step 521). Next, the codebook vector Yi is input to a network that has not yet been set (step 522). Then, 1 is set to correspond to the input index i.
The coupling coefficient of the network is changed so that one output unit outputs "1" and the other output units output rOJ (step 823).

Then, it is determined whether index i has been processed from rlJ to rNJ (step 524).

Here, if the index i is less than N (NO), "1" is added to the index i (step 525), and the process returns to step S22.

Also, when the index L becomes the same number as N (Yes
), for any input index i, one index i is output corresponding to one output crab unit, and if the same index i is, it is determined whether the same output crab unit is always output. (step 526)
.

Then, if the index i does not correspond to one index i or one a crab unit, such as when the index i is still output to a different output crab unit (No), the process returns to step S1. If it is determined that the index l has been correctly output (Yes), it is assumed that learning has been completed, and the process of forming this network 400 is ended.

Next, in the optimal vector search unit 2, which is an encoding unit, the input vector X is first input to the search order determination unit 20.

In this search order determining unit 20, using the network 400 whose learning has been completed using the codebook vector,
The indexes are rearranged so that the search is performed in descending order of the firing amount (output amount) of the output crab units. Then, in accordance with this order, the processing shown in the flowchart of FIG. 3 is performed similarly to the first embodiment.

According to the second embodiment as described above, no matter how the codebook vectors are distributed, the distortion calculation starts from the codebook vector closest to the input vector, so most other codebook vectors This means that there is no need to perform distortion calculations, or even if distortion calculations are necessary, they can be aborted with fewer calculations. This makes it possible to significantly reduce the number of calculations required to find the optimal codebook vector for the input vector.

In the above explanation, the distortion calculation used square distortion, but it is also possible to use other distortions.

Further, the present invention is not limited to the one embodiment described above, and it goes without saying that various modifications and applications can be made without departing from the gist of the invention.

[Effects of the Invention] As described in detail above, according to the present invention, it is possible to provide a vector quantization method and an apparatus therefor that can search with high efficiency and at high speed.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a high-speed vector quantization device according to a first embodiment of the present invention, FIG. 2(a) is a schematic diagram as an example of a codebook, and FIG. 2(b) is a sorting table. A schematic diagram as an example, Figure 2 (C) is a conceptual diagram showing the movement of a pointer indicating the base value (nearest base value) closest to the input vector A flowchart in which the search unit searches for an optimal codebook vector, FIG. 4 is a diagram showing the configuration of the vector quantization device according to the second embodiment of the present invention, and FIG.
FIG. 5(a) is a flowchart showing an example of a method for forming a network, and FIG. 5(b) is a diagram showing an example of the configuration of the network. DESCRIPTION OF SYMBOLS 1... Codebook generation part, 2... Optimal vector search part, 10... Training part, 11... Sorting part, 12... Network generation part, 20... Search order determination part, 21 ...component calculation order determining unit, 22...distortion calculation unit, 22a...dimensional distortion calculation unit, 22b...
Partial distortion calculation unit, 23... Lower limit value calculation unit, 24...
Memory, 25...Comparison unit, 100...Codebook, 200...Sort table, 300...Component calculation order table, 400...Network. Procedural amendment written on March 9, 1998, 4 days

Claims (4)

[Claims] (1) An upper limit value setting step in which the minimum distortion is set as the upper limit value of the distortion after the search by comparing the input vector and the codebook vector is completed; a lower limit value setting step in which the sum of minimum distortions calculated for each dimension of the codebook vector and the input vector is set as a lower limit value of distortion; a storage comparison step in which the lower limit value and the upper limit value are stored and compared; a search termination step in which the upper limit value and the lower limit value are compared in the storage comparison step and the search is terminated when they match; and when the search is determined to be unfinished in the search termination step, the search is sequentially performed for each dimension up to a predetermined order component; a distortion partial sum calculation step of calculating a partial sum of distortions between the codebook vector and the input vector; and comparing the partial sum with the upper limit value, and when the partial sum is equal to or larger than the upper limit value. , a distortion calculation abort step of aborting further distortion calculations, and an upper limit value updating step of updating the partial sum as a new upper limit value when the partial sum is smaller than the upper limit value. vector quantization method. (2) The absolute value of the difference between the average value of the codebook vector for each dimension and the input vector is calculated for each dimension, and in the distortion partial sum calculation step, the components with the largest absolute value of the difference are calculated in order. Claim (1) characterized by comprising the step of setting the selection order so that
Vector quantization method described. (3) After completing the search by comparing the input vector and the codebook vector, there is an upper limit setting step in which the minimum distortion is set as the upper limit value of the distortion, and a vector consisting of multiple dimensions forming the codebook is set. On the other hand, a classification means that classifies a unique component value for each dimension as a base value, and a distortion between a base value belonging to a codebook vector that most closely approximates the input vector and provides an unsearched codebook vector and a minimum vector. a lower limit value setting means for calculating a sum over all dimensions of; a storage comparison means for storing and comparing the lower limit value and the upper limit value; and a storage comparison means for comparing the lower limit value and the upper limit value and determining whether they match. a search terminating means for terminating the search at a point in time; and a distortion for calculating a partial sum of distortions between a codebook vector and an input vector sequentially from a low-dimensional component to a predetermined order component when the search is determined to be unfinished by the search terminating means. a partial sum calculation means; a distortion calculation aborting means for comparing the partial sum and the upper limit value and aborting further distortion calculations when the partial sum is equal to or larger than the upper limit value; and the partial sum a vector quantization device, comprising upper limit updating means for updating the partial sum as a new upper limit when the partial sum is smaller than the upper limit. (4) average value calculation means for calculating the average value of vectors forming a codebook for each dimension; and difference calculation means for calculating the absolute value of the difference for each dimension between the average value by the average value calculation means and the input vector. A vector quantization device, comprising: calculation order determining means for determining an order so that the distortion calculations are performed in descending order of the absolute value of the difference by the difference calculation means.