JP3015446B2 - Vector quantization method and vector quantization apparatus - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to digital information processing technology, and more particularly, to a vector quantization method and a vector quantization device.

(Prior Art) Generally, when an image, a sound, or the like is expressed as a function,
The operation of discretization with respect to a specific range is quantization. That is, a continuous function on sampling points obtained by sampling is represented by being converted 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 having pixels of each block as components. For example, for a specific block, Several representative image patterns (representative vectors) that are considered to occur at a high frequency are determined, assigned identification codes, and stored in the encoder. This is called a codebook, and when an input signal results, a representative vector most similar to the pattern is searched from the codebook for each block, and the identification code is transmitted. The same applies to voice and the like.

For the purpose of speeding up the processing, for example, a vector quantization apparatus that performs such vector quantization includes, for example, a tree search method using a structural codebook to reduce the number of searched vectors ( A.Buzo, AHGray.Jr., R.
M. Gray, and JDMarkel: “Speech coding based upon v
ector quantization ", IEEE Trans.Acoust., Speech & S
ignal Process., ASSP-28, 10, pp. 562-574, Oct. 1980)
Has been proposed. Also, among the methods for reducing the number of times of calculating the distortion used for vector quantization,
A method of changing the search order according to the probability of occurrence of a vector (Japanese Patent Laid-Open No. 1-218279) and a method of using an average value in a vector for search termination (Chuide, Aizawa, Harashima: Search method, IEICE, pp.110
2-1110, Vol J69-B, No. 10, 1986), etc.
The speed of vector quantization is increased. (Problems to be Solved by the Invention) However, among the above-described vector quantization methods, the present search method is faster than the other methods in the speed of distortion calculation, but the input vector is at the boundary of the subspace. In the vicinity ("High-speed search method for orthogonal transform domain of vector quantization" P.11
In some cases, for example, there is a drawback that the optimal vector cannot be obtained from the codebook. Therefore,
Tree search methods may lack accuracy in the sense of minimizing distortion.

Even in the method of reducing the number of distortion calculations, it is necessary to perform at least one comparison or some processing on all the codebook vectors.

In the conventional search method as described above, the search for the coat book vector corresponding to the resulting input vector cannot satisfy the optimum, high efficiency, and high speed depending on the properties of the input vector.

Therefore, the present invention provides a vector quantization method capable of searching for an optimum codebook vector efficiently and quickly with a small number of calculations for any input vector, and a vector quantization method for executing the method. It is intended to provide a device.

(Means for Solving the Problems) In order to achieve the above object, the present invention provides, as a first means, a method in which the minimum distortion in the search by comparing the input vector with the codebook vector is completed. An upper limit value setting step to set an upper limit value, and a lower limit to set a sum of a minimum distortion calculated for each dimension of the codebook vector giving the minimum distortion and a codebook vector not yet searched and the input vector as a lower limit value of the distortion. A value setting step, a storage comparing step of storing and comparing the upper limit value and the lower limit value,
A search termination step of terminating the search when the upper limit value and the lower limit value are compared and matched by the storage comparison step, and when the search termination step determines that the search has not been completed, a code for each component is sequentially transmitted to a predetermined next component. A partial distortion calculation step of calculating a partial sum of distortion between the book vector and the input vector,
Comparing the partial sum with the upper limit value, when the partial sum is equal to or greater than the upper limit value, a distortion calculation discontinuing step of discontinuing further distortion calculation, and the partial sum is equivalent to the upper limit value. Alternatively, when it is smaller, an upper limit value updating step of updating the partial sum as a new upper limit value can provide a vector quantization method.

As a second means, for a vector composed of a plurality of dimensions forming a codebook of an upper limit value setting means, a classification means for classifying a unique component value as a base value for each dimension, And a lower limit value setting means for calculating a sum of distortions in all dimensions of an unsearched codebook vector and a base value belonging to a codebook vector providing a minimum distortion, storing and comparing the lower limit value and the upper limit value Storage comparing means, and a search terminating means for terminating the search when the lower limit value and the upper limit value are compared and matched by the storage comparing step, and when the search terminating step has not been completed, sequentially for each component, Partial distortion calculating means for calculating a partial sum of distortions of the codebook vector and the input vector up to a predetermined next component, and comparing the partial sum with the upper limit value, When the partial sum is equal to or larger than the upper limit value, a distortion calculation terminating means for stopping further distortion calculation when the partial sum is equal to or smaller than the upper limit value. It is possible to provide a vector quantization device having upper limit value updating means for updating as an upper limit value.

(Operation) According to the vector quantization method and the vector quantization apparatus configured as described above, it is possible to efficiently and quickly search for the 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 a 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 unit 1 and an optimal vector search unit 2.

In the codebook generation unit 1, first, the training unit 10 generates a codebook 100 based on a given training sequence. Then, based on the codebook 100, the sorting unit 11 sorts (sorts) each component into a base value (a unique vector component amount), creates a sorted table 200, and assigns the number of the base value to the corresponding value of the codebook 100. Is added to the vector component.

Then, when the input vector X is input, the optimal vector search unit 2 determines a component calculation order using the average value of the sort table 200, and a component calculation order determination unit 21. A distortion calculation unit 22 including a dimensional distortion calculation unit 22a that calculates distortion for each dimension and a partial distortion calculation unit 22b that calculates the distortion partial sum; and a lower limit calculation unit that obtains a lower limit value of the distortion using the sort table 200. twenty three,
A memory 24 that stores the partial distortion amount, the lower limit value, and the minimum distortion amount as the upper limit value, determines the termination of the distortion calculation using the upper limit value and the partial sum of the distortions, and searches for the search using the upper limit value and the lower limit value. And a comparison unit 25 for determining the end.

Further, the component calculation order determining section 21 sorts the
The component calculation order table 300 is created so that the component calculation is performed in descending order of the difference between the component average value and the component of the input vector X using the component average values of the component.

FIGS. 2A and 2B are schematic diagrams showing an example of the format of the code book 100 and the sort table 200 described above.

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

In FIG. 2 (c), the k-th codebook vector
When the next component is classified as a base value, the t-th base value counted from the smallest one is represented by “Vk (t)”, and a pointer indicating the base value (closest base value) closest to the input vector X is These are indicated by “p” and “q”. Then, the difference between the input vector X and the base value of the codebook vector is
As Ds (k) and Db (k), the difference Ds (k) represents the difference on the side smaller than X.

Next, the operation of the vector quantization device of the first embodiment configured as described above will be described. Here, the number of codebook vectors is “N”, the number of dimensions of the vector is “M”, and the codebook vector indicated by index i is represented by “Yi”. First, a codebook generation unit 1 as a preprocessing unit generates a codebook 100 and a sort table 200 based on a training sequence.

Next, according to the flowchart shown in FIG. 3, the optimal vector search unit 2 as an encoding unit searches for an optimal codebook vector. This procedure will be described sequentially using the components shown in FIG. 1 and their reference numbers.

First, when the input vector X is input to the component calculation order determination unit 21, as an initial setting, the component calculation order determination unit 21 uses the component average value of the sort table 200 to calculate the component average value and the components of the input vector X. The component calculation order table 300 is created so that the component calculation is performed in the descending order of the difference (hereinafter, in order to simplify the description, the component giving the largest difference is the first order, and the component giving the smallest difference is the Mth order Are arranged in numerical order so that

Next, the distortion calculation unit 22 extracts the codebook vector Y1 corresponding to the index 1 from the codebook 100, calculates the distortion with respect to the input vector X, and calculates the sum U of the all-dimensional distortion by the equation (1). And the corresponding index min
(Here, 1) is stored in the memory 24 (step
S1). That is, the sum U of the distortions in all dimensions is set as the upper limit.

On the other hand, the lower limit calculation unit 23 uses the sort table 200 to calculate square distortions Ds (k) and Db (k) between the input vector X and the base value of the codebook vector as follows (2) and (3). Calculate by formula.

Ds (k) = (X (k) −Vk (p)) ² (2) Db (k) = (Vk (q) −X (k)) ² (3) However, FIG. 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 represents the difference on the side smaller than X.

Then, the lower limit L is calculated by the following equation (4): And stores it in the memory 24 (step S2).

Then, the comparing unit 25 compares the upper limit U and the lower limit L stored in the memory 24 (step S3).

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

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

The distortion calculator 22 extracts the codebook vector Yi from the codebook 100 at the index i and calculates the distortion with respect to the input vector X (step S6). here,
Partial sum Sj of distortion up to the j-th component in order from the low-dimensional component
Is given by the following equation (5). And sends it to the comparison unit 25. In this comparison unit 25,
The partial sum Sj is compared with the upper limit value U stored in the memory (step S7), and when Sj ≧ U (6) is satisfied (Yes), further distortion calculation is terminated.

That is, when the distortion calculation is discontinued, the index i + 1 corresponding to the next codebook vector is sent to the distortion calculator 22, and the target is shifted to the next codebook vector. Further, since the vector indicated by the index i does not give the minimum distortion (upper limit U), the number of vectors belonging to the base value corresponding to this vector is reduced by 1 (step S
8) The process proceeds to step S13 described later.

On the other hand, if the calculated distortion value is not ceased (No), the order j of the vector is compared with the number of dimensions M (step S9).

Here, if the number of dimensions of the vector is smaller than M (No),
"1" is added to the degree j of the vector (step S10),
Return to step S6. That is, it is sent to the distortion calculator 22 (step S10).

Further, even if the degree j of the vector is equal to the dimension number M, if the distortion partial sum Sj is smaller than the minimum distortion up to that point (Yes), it can be seen that the vector indicated by the index min does not give the minimum distortion. Decrease the number of vectors belonging to the base value corresponding to the vector by 1 (step S1
1).

This minimum distortion is stored in the memory 24 as a new upper limit value U, and the index i at this time is stored in a new index mi.
It is stored in the memory 24 as n (step S12).

In this manner, the number of vectors belonging to the base value corresponding to the codebook vector found not to give the minimum distortion is reduced by one, and the number of vectors belonging to the nearest base value becomes “0”. It is determined whether or not there is (step S13). Here, if there is a closest base value of “0” (Yes), it is shifted to a base value where a vector to which the corresponding closest base value pointer “p” or “q” belongs exists (step S14), and the lower limit is set. The process returns to step S2 to recalculate the value L.

On the other hand, if the number of vectors belonging to the nearest base value has not become “0”, the lower limit value L is left as it is and step S3
Continue processing.

The above procedure is performed until the upper limit U and the lower limit L match.

From the above, it is possible to obtain the same solution as the optimal solution obtained by performing a full search of the codebook vector,
In addition, the search can be terminated on the way. Also in the distortion calculation, the components are calculated in order from the component having a high possibility of increasing the distortion by adapting to the input vector. Therefore, the partial distortion is maximized with the minimum calculation, and can be terminated 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. Further, since the distortion between the nearest base value and the input vector component in the lower limit value calculation unit 23 can be calculated completely independently for each component, it is possible to further increase the speed by processing them in parallel.

Next, FIG. 4 shows the configuration of a vector quantization apparatus as a second embodiment of the present invention. The configuration of this device is composed of a codebook generation unit 1 and an optimum vector search unit 2 as in the first embodiment.

The codebook vector generation unit 1 generates a network 400 that, when a vector of the codebook 100 is input, outputs an index corresponding to the vector, in addition to the training unit 10 and the sorting unit 11 of the first embodiment. It is configured by the network generation unit 12. The optimum vector searching unit 21 is a component calculation order determining unit according to the first embodiment.
21, distortion calculator 22, lower limit calculator 23, memory 24, comparator 25
In addition, a search order determining unit 20 that determines the search order for the codebook 100 using the network 400 when the input vector X is input is added to the configuration.

 Next, the operation of the device having such a configuration will be described.

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

That is, in the network generation unit 12, the back propagation method (DERumelhart, GEHinton, and RJWilliams:
Learning representations by back-propagating erro
rs, Nature, Vol. 323-9, pp. 533-536, 1986) and the like, and a network 400 corresponding to a codebook such as the configuration shown in FIG. 5B is generated.

The network 400 includes an input layer including a plurality of input units for inputting an input vector, an intermediate layer including a plurality of intermediate units each connected to all the input units, and further, each including all the intermediate units. And a plurality of output units corresponding to the number of codebook vectors to be connected.

Such a combination is a simple hierarchical combination in which one codebook vector corresponds to one output unit. In addition to such a hierarchical connection, it is also possible to use some networks represented by a neutral network such as a hierarchical connection in which units are interconnected in each layer.

Next, a method of forming a network will be described with reference to the flowchart of FIG. Here, it is assumed that the initial value of the coupling coefficient of the network is set by a random number value or the like.

First, the index i is initialized to "1" (step S21). Next, the codebook vector Yi is input to the network (step S22). Then, the coupling coefficient of the network is changed so that one output unit outputs "1" and the other output unit outputs "0" so as to correspond to the input index i (step S23).

Then, it is determined whether or not the index i has been processed from "1" to "N" (step S24). Here, when the index i is less than N (No), "1" is added to the index i (step S25), and the process returns to step S22.

Also, when the index i becomes equal to N (Ye
In s), for any input index i, one index i is output corresponding to one output unit, and if the index i is the same, it is determined whether the same output unit is always output. (Step S2
6).

If one index i does not correspond to one output unit (for example, the index i is still output to a different output unit) (NO), the process returns to step S21. If it is determined that the index i is correctly output (Yes), it is considered that the learning has been completed, and the network 400 forming process ends.

Next, in the optimal vector search unit 2 which is an encoding unit, first, the input vector X is input to the search order determination unit 20.
The search order determination unit 20 rearranges the indexes using the network 400 for which learning has been completed by the codebook vector, so that the search is performed in the descending order of the firing amount (output amount) of the output unit. Then, according to this order, the processing shown in the flowchart of FIG. 3 is performed as in the first embodiment.

According to the second embodiment as described above, no matter how the codebook vectors are distributed, the distortion calculation is started from the codebook vector closest to the input vector. Is
It is not necessary to perform the distortion calculation, or even if the distortion calculation is required, it is necessary to terminate the calculation with a small number of calculations. As a result, the number of calculations for finding the optimal codebook vector for the input vector can be significantly reduced.

In the above description, the distortion calculation uses the squared distortion, but other distortions can be used.

Further, the present invention is not limited to the above-described embodiment, and it is needless to say 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 a device capable of performing high-speed search at high efficiency.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a high-speed vector quantization apparatus according to a first embodiment of the present invention, FIG. 2 (a) is a schematic diagram showing an example of a codebook, and FIG. FIG. 2C is a schematic diagram showing an example of a table. FIG. 2C is a conceptual diagram showing the movement of a pointer indicating a base value (nearest base value) closest to the input vector X in a certain dimension of the vector, and FIG. FIG. 4 is a flowchart showing how a vector search unit searches for an optimal codebook vector, FIG. 4 is a diagram showing a configuration of a vector quantization apparatus according to a second embodiment of the present invention, and FIG. FIG. 5B is a flowchart showing an example of the configuration, and FIG. 1 codebook generation unit 2 optimal vector search unit 10 training unit 11 sort unit 12 network generation unit 20 search order determination unit 21 component calculation order determination , A distortion calculator, 22a a dimensional distortion calculator, 22b a partial distortion calculator, 23 a lower limit calculator, 24 a memory, 25 a comparator, 100 a codebook, 200: Sort table, 300: Component calculation order table, 400:
…network.

Claims (4)

(57) [Claims] 1. After the search by comparison between an input vector and a codebook vector is completed, an upper limit setting step of setting the minimum distortion in the search as an upper limit of distortion, a codebook vector giving the minimum distortion, An end codebook vector, a lower limit value setting step of setting a sum of minimum distortions calculated for each dimension of the input vector as a lower limit value of the distortion, and a storage comparing step of storing and comparing the lower limit value and the upper limit value. A search termination step of terminating the search when the upper limit value and the lower limit value are compared and matched by the storage comparison step, and when the search termination step determines that the search has not been completed, a predetermined order is sequentially determined for each dimension. A distortion partial sum calculation step of calculating a partial sum of distortion between the codebook vector up to the component and the input vector; and comparing the partial sum with the upper limit value, When the value is equal to or greater than the upper limit value, a distortion calculation discontinuation step of discontinuing further distortion calculation, and when the partial sum is smaller than the upper limit value, an upper limit update for updating the partial sum as a new upper limit value. A vector quantization method comprising the steps of: 2. An absolute value of a difference between an average value of a codebook vector for each dimension and an input vector is calculated for each dimension,
4. The vector according to claim 1, wherein the distortion partial sum calculation step includes a step of setting an order of selection such that components having a larger absolute value of the difference are to be calculated in order. Quantization method. 3. An upper limit value setting step in which the search by comparison between the input vector and the codebook vector is completed, wherein a minimum distortion in the search is set as an upper limit value of the distortion, and a plurality of dimensions forming the codebook. Classifying means for classifying a unique component value as a base value for each dimension with respect to a vector, a base value belonging to a codebook vector which is closest to the input vector and which provides an unsearched codebook vector and a minimum vector. A lower limit value setting unit that calculates a sum over all dimensions of the distortion, a storage comparison unit that stores and compares the upper limit value and the lower limit value, and the upper limit value and the lower limit value are compared by the storage comparison unit. A search terminating means for terminating the search at the time of matching; A distortion partial sum calculating means for calculating a partial sum of distortions between the input vector and the input vector, 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 vector calculation means for terminating the distortion calculation, and an upper limit updating means for updating the partial sum as a new upper limit when the partial sum is smaller than the upper limit. Device. 4. An average value calculating means for dividing a vector forming a codebook into each dimension and calculating an average value for each dimension, and an average value calculated by the average value calculating means and a component of an input vector. A vector quantization apparatus, comprising: a calculation order determining unit that determines the order in which the distortions are calculated in descending order of the distortion indicating the difference.