JP3078917B2 - Quantization method and quantization device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an apparatus for quantizing an input group of data samples, and to a processing method performed by such an apparatus.

[0002]

2. Description of the Related Art As an apparatus for quantizing an input group of data samples, there has been proposed an apparatus for reducing an amount of memory necessary for holding data accompanying a compression of input data or a transmission rate required for data transmission. Various devices have been developed. In general, these devices often quantize 2 to 32 data samples together as an input group. At this time, each data sample is represented by a predetermined number of bits such as 8 bits. Therefore, as the number of data samples in a group and the number of bits of each data sample increase, the total number of bit strings included in one group also increases.

In the quantization process, a process of replacing a plurality of different bit strings existing at positions “close” to each other with one bit string is performed. That is, this one bit string forms one quantization group of the data sample. By limiting the total number of quantization groups, desired data compression can be performed. For example, it is possible to quantize an input group of data samples of millions of bit strings into one of several thousand quantization groups.

[0004] In order to quantize an input group of data samples in this way, a quantizer must perform a number of processes. Here, these processes depend on the complexity of the process performed by the quantization device. In addition, in order to execute the above processing, the quantization device must read all the quantization groups from the storage module. Here, the storage capacity of the storage module depends on the processing to be executed. Therefore, in order to perform the quantization processing at high speed, the degree of complexity of the processing to be performed must be reduced. In addition, in order to obtain an economical quantization device, the size of the storage module holding all the quantization groups must be reduced.

[0005]

SUMMARY OF THE INVENTION From such a viewpoint,
SUMMARY OF THE INVENTION It is a main object of the present invention to provide a new quantization method and a new quantization apparatus capable of quantizing an input group of data samples with a very small storage capacity and a small number of processes as compared with the conventional apparatus. .

[0006]

In order to solve this problem, a quantizing apparatus according to the present invention provides a data processing apparatus having substantially more than N possible values when N is a predetermined positive integer. An apparatus for quantizing sampled data for converting an input group of samples to a centroid of one of N regions divided from said values, an input circuit receiving said input group of data samples, and a product μ · k Are selected to arrive at a boundary orthogonal to the boundary between at least two of the N regions,
Storage means for storing a reference group μ of a plurality of data samples each of which does not exceed N / 2 and N-1 and a plurality of constants k; and one of the reference groups of the data samples connected to the storage means. and a control means for selecting one k _y of one mu _x and the constant, coupled to said storage means and said input circuit, a calculating means for calculating the inner product of the reference group mu _x and input groups of the data samples , connected to said computing means, comparison means for comparing said constant k _y and inner product result by the calculating means, connected to said comparing means, for defining the center of gravity by the respective all comparison result by the comparison means and a logic means, said control means is further connected to said comparing means, each node, as well as identify and k _y for the comparison means and mu _x for the computing means, the comparison means Comparison Indicates take either branches based on, and the mu _x and k _y are selected to correspond to the selection tree having a plurality of said nodes are interconnected by branches, the comparison and the calculating means Means and different μ to enter
characterized in that it repeatedly select the _x and k _y.

When N is a predetermined positive integer,
An apparatus for quantizing sampled data for transforming an input group of data samples having substantially more than N possible values into a centroid of one of N regions divided from said values,
An input circuit for receiving an input group of the data samples;
Storage means for storing N-1 constants k; register means for holding a reference group μ of N / 2 or less data samples; wherein the product μ · k is at least 2 of the N areas.
Orthogonal to one of the boundaries is selected to reach the boundary, and is connected to the input circuit and the storage means, and the respective inner products of the input group of the data samples and the reference group μ of all the data samples a calculating means for calculating in parallel, connected to said calculation means and said storage means, and selecting means for selecting one k _y of one R _x and the constant of the dot product result, are connected to the selection means, comparing means for comparing the constant k _y that the selected inner product results and the selected R _x, coupled to said selecting means, with each node, specifying the of R _x and k _y of said selection means, said It indicates take either branch based on the comparison result of the comparison means, each of R _x and k
and the _y, is selected to correspond to the selection tree having a plurality of said nodes are interconnected by branches, in order to select a different of R _x and k _y, and control means for activating the repeat said selection means And a logic unit connected to the comparison unit and defining the center of gravity from each of all the comparison results by the comparison unit.

[0008] The quantization method of the present invention, when N is a predetermined positive integer, divides an input group of data samples having substantially more than N possible values into N groups divided from said values. Receiving the input group of data samples from an electronic input circuit, wherein the product μ · k is orthogonal to at least two boundaries of the N regions. Storing a reference group μ and a plurality of constants k of a plurality of data samples selected to reach the boundary and not exceeding N / 2 and N−1 in the storage means, wherein one of the reference group of data samples and mu _x and step for selecting the one of said constant k _y, and step of obtaining the inner product of the reference group mu _x and input groups of the data samples from, obtained by the inner product Inner product result and _ky A step of comparing the bets, whether each node, as well as identify and k _y for the mu _x and the comparing step for said calculating step, taking either branch based on the comparison result of the comparing means shown, each mu _x and k _y are selected to correspond to the selection tree having a plurality of said nodes are interconnected by branches, each iteration mu _x and k _y
And a step of repeating the selection step, the inner product step, and the comparison step several times, and a step of combining all of the comparison results to define the center of gravity.

When N is a predetermined positive integer,
A quantization method for transforming an input group of data samples having substantially more than N possible values into a centroid of one of N regions divided from said values, comprising: Receiving from an electronic input circuit;
A step of storing one constant k in the storage means, and a step of holding the reference group μ of N / 2 or less data samples in the register means, wherein the product μ · k of the N areas Determining a respective inner product of an input group of the data samples and a reference group μ of all of the data samples, wherein the inner product is selected so as to reach the at least two boundaries orthogonally; one and of R _x and stroke for selecting one k _y of the constant, the selected inner product result R _x
Or a step of comparing the selected constant k _y, each node, as well as identify and of R _x and k _y for the selection step takes either branch based on the comparison result of the comparing step and are shown, and each of R _x and k _y is in is selected corresponding to the selection tree having a plurality of said nodes interconnected branches, while selecting and of R _x and k _y for each iteration, the The method includes a step of repeating the selection step and the comparison step a plurality of times, and a step of defining the center of gravity by combining all of the comparison results.

[0010]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first preferred embodiment of the present invention is an apparatus for sequentially converting an input group of data samples into one of N quantization groups, comprising a storage module, a control module, an arithmetic module, This is a device including a comparison module and a flip-flop. N-1 constants and a reference group of very few data samples are stored in the storage module. Process, the control module transmits an address signal to the memory module is started by reading the reference group mu _x and constants k _y of data samples. Then, we obtain an inner product of the reference group mu _x of the input groups and data sample of the data samples in the computation module, and compares the inner product result and the constant k _y in the comparison module. Here, if the inner product result is equal to or _larger than the constant ky, one of the flip-flops is set, and if it is equal to or _smaller than the constant, reset. Subsequently, the control module selects a different mu _x and k _y and before entering into the comparison module and the operation module. Selection process of the mu _x and k _y is performed based on the comparison result from this until the comparison module.
The result of the comparison process is held in each flip-flop, and the state of the flip-flop becomes a quantization group into which the input group of the data sample is transformed.

According to the above apparatus, the storage capacity of the storage module can be reduced by several thousand percent compared to the conventional quantizer. This is possible by limiting the total number of reference groups of data samples in the storage module to 2 log ₂ N or less. If r is the number of samples in the input group of data samples, it is possible to limit the total number of reference groups of data samples in the storage module to 2r or less.

A second preferred embodiment of the present invention is an apparatus for converting an input group of data samples into one of N quantization groups in parallel, comprising a storage module, a register module, a control module, and an arithmetic module. , A multiplexer module, a comparison module, and a flip-flop. N-1 constants are stored in the storage module and reference groups of very few data samples are stored in the register module. The processing is started by the processing module executing in parallel the processing for obtaining the inner product of the input group of the data sample and the reference group of the data sample. Then, at the same time the control module selects the dot product result R _x to send an address signal to multiplexer module, the control module selects the constant k _y by transmitting address signals to the memory module. Next, a comparison between the inner product result and the constant k _y in the comparison module. Here, if the inner product result is equal to or _larger than the constant ky, one of the flip-flops is set, and if it is equal to or _smaller than the constant, reset.
Subsequently, the control module selects a different inner product result of R _x and k _y and before entering into the comparison module. This R
Selection process of _x and k _y is performed based on the comparison result from this until the comparison module. The result of the comparison process is held in each flip-flop, and the state of the flip-flop becomes a quantization group into which the input group of the data sample is transformed.

According to the above apparatus, the storage capacity of the storage module can be made very small. This is made possible by limiting the total number of reference groups of data samples in the register module to 2 log ₂ N or less. If r is the number of samples in the input group of data samples, the total number of reference groups of data samples in the register module is 2
It is also possible to limit it to r or less.

Hereinafter, a preferred embodiment of the present invention, that is, an example of processing for quantizing an input group of data samples and an example of an apparatus for executing quantization processing will be described with reference to the accompanying drawings. First, as a simple example, a case in which an input group of data samples includes two samples S1 and S2 will be described. Each dot 10 in FIG. 1 indicates a specific input group consisting of one data sample S1 and one data sample S2. Here, the value of the data sample S1 is represented by the S1 axis, and the value of the data sample S2 is represented by the S2 axis. In general, the input group of each data sample can have any number of samples. For example, each group may be a group of 4 × 4 16 samples extracted from the image array.

Each data sample in FIG. 1 is obtained by sampling the output electric signal from the sensor. For example, this sensor senses a reflected signal from a radar, an optical image, an infrared image, a multi-frequency spectrum, or an acoustic signal. Then, all samples output from the sensor are converted to digital form,
It is represented by a predetermined number of bits such as 8 bits.

These digital samples are then grouped in a predetermined manner into pairs S1 and S2. For example,
Samples S1 and S2 of one group are the first and second samples from the sensor, respectively, and sample S1 of another group is
And S2 can be grouped to be the third and fourth samples from the sensor, respectively. In addition, each sample S
1 may be grouped as samples from one type of sensor, such as an optical image sensor, and each sample S2 may be grouped as a sample from another type of sensor. The data sample acquisition and grouping process described above is continued until the overall distribution of the data samples becomes characteristic of a particular output signal from the sampled sensor.

Subsequently, as shown in FIG. 2, the data sample group of FIG. 1 is divided into N multiple regions 20. Then, all the data sample groups included in a certain area are quantized to the center of gravity (average value) of the area. That is, each centroid is a quantization group of N data samples, and the input group of all data samples is converted to one of these quantization groups.

As shown, the region 20 has various sizes and shapes. According to the present embodiment, the region 20 is formed by two types of boundary lines 21 and 22. Boundary line 21 is first
Are parallel to each other in the direction of
Are border lines parallel to each other in the second direction. FIG.
In order to divide the data sample group 10 shown in FIG. 3 into a plurality of regions 20, a repetitive process as shown in FIGS. 3A to 3D is performed.

First, a first boundary line 21-1 for dividing the entire data sample group 10 of FIG. 1 into two regions 20-1 and 20-2 is determined. The center of gravity of the data sample group located in the region 20-1 is 20-1C, and the region 20-C
The center of gravity of the data sample group located within 2 is 20-2C
It is. Quantizing all data sample groups in region 20-1 to centroid 20-1C results in distortion D1. The distortion D1 is proportional to the sum of the squares of the distances between all the data sample groups in the area 20-1 and the center of gravity 20-1C. Similarly, when all the data sample groups in the region 20-2 are quantized to the centroid 20-2C, the sum of the squares of the distances between all the data sample groups in the region 20-2 and the centroid 20-2C is obtained. Distortion D2 proportional to
Occurs. Here, these distortions D1 and D2 change depending on the position and the inclination of the boundary line 21-1. Therefore, in order to reduce the distortion, the position and the inclination of the boundary line 21-1 are adjusted until the sum of the distortions D1 and D2 reaches the allowable value or less.

When the final position of the boundary 21-1 is determined, the direction of the boundary is changed to the first reference group μ ₁ of the data sample.
Expressed by Here, the first reference group mu ₁ street, data samples shown in FIG. 3A, the length perpendicular to the boundary line 21-1 is defined as a vector from the origin of "1". On the other hand, the position of the boundary line 21-1, k ₁ mu ₁ border 21
Expressed in constant k ₁ when crossing the 1.

Next, as shown in FIG.
1, the region 20-1 is divided into two regions 20-3 and 20-
Divide into four. The center of gravity of the data sample group located in the region 20-3 is 20-3C, and the center of gravity of the data sample group located in the region 20-4 is 20-4C. Quantizing all data sample groups in region 20-3 to centroid 20-3C is proportional to the sum of the squares of the respective distances between all data sample groups in region 20-3 and centroid 20-3C. The distortion D3 occurs. Similarly, area 20
Quantizing all the data sample groups in the region 20-4 to the center of gravity 20-4C, the distortion D4 proportional to the sum of the squares of the distances between all the data sample groups in the region 20-4 and the center of gravity 20-4C. Occurs. Therefore, in order to reduce the distortion, the boundary line 22 is set until the sum of the distortions D3 and D4 reaches the allowable value or less.
Adjust the position and tilt of -1.

Once the final position of the boundary line 22-1 has been determined, the direction of the boundary line is determined by the second reference group μ ₂ of the data sample.
Expressed by Here, as shown in FIG. 3B, the second reference group μ ₂ of the data sample is defined as a vector from the origin having a length of “1” and being perpendicular to the boundary line 22-1. on the other hand,
Position of the boundary line 22-1, k ₂ mu ₂ is expressed by the constant k ₂ when crossing the boundary line 22-1.

Subsequently, regions 20-2 and 20- in FIG.
A process of dividing each of 3, 20-4 into two regions is performed, and the divided regions are further subjected to a dividing process. Such division processing is repeatedly performed until the area division as shown in FIG. 2 is performed. At this time, the boundaries 21 and 22 that divide the region are selected so as to be parallel to the boundaries 21-1 and 22-1. That is, all the boundaries 21 and 22 are selected to be orthogonal to μ ₁ and μ ₂ , respectively.

FIGS. 3C and 3D show an example in which the data sample group is divided into several areas by a boundary line having a fixed direction. In Figure 3C, the region 20-3 by a boundary line 22-2 which is orthogonal to the mu ₂ are regions 2
It is divided into 0-5 and 20-6. Here, the position and the inclination of the boundary line 22-2 can be represented by k ₃ μ ₂ . Further, in FIG. 3D, region 20-5 by a boundary line 21-2 which is orthogonal to the mu ₁ is a region 20-7 20
−8. Here, the boundary line 21-2 is
It can be represented by k ₄ μ ₁ .

In order to generate N plural regions 20 as shown in FIG. 2, a total of N-1 boundary lines 21 are formed.
And 22 are required. At this time, the position and tilt of the N-1 boundaries, can be expressed by the product of a constant k _i and mu ₁ or mu _2. Therefore, state of the area division illustrated in FIG. 2 is represented by holding the the reference group mu ₁ of the N-1 constant k _i and two data samples mu ₂ and the digital memory.

FIG. FIGS. 4A-4C, following the above process, the input group V _i of any data samples in Figure 2 of N
FIG. 10 is a diagram illustrating a state of a process of quantizing to a barycenter of one of a plurality of regions. Here, V _i is a vector. First, the input groups V _i and μ ₁ of the data samples
Find the inner product with Here, the inner product, the product of the S ₁ value of S ₁ value and mu ₁ of V _i, is that the product of the value obtained by adding the S ₂ value of S ₂ value and mu ₁ of V _i. As R ₁ in FIG. 4A, V
shows the inner product result of _i and μ _1. Subsequently, the result the comparison process is performed between the R ₁ and a constant k _i. Here, the result R ₁ is a constant k _i
The case is at least, input group V _i of the data sample is determined to belong to the region 20-1. On the other hand, if the result R ₁ is equal to or less than the constant k _i is input group V _i of the data sample is determined to belong to the region 20-2.

FIG. 4A shows an input group V _{i of} data samples.
Belongs to the area 20-2, but as shown in FIG.
0-2 is divided into two regions by a boundary line 22-X. Therefore, subsequently, the processing of the input group V _i of data samples to determine belongs to either side of the boundary line 22-X performed. Here the boundary line 22-X is represented by k _x μ _2, to determine the areas belonging by obtaining an inner product between the input group V _i and mu ₂ data samples similar to the above processing. R ₂ in FIG. 4B is a dot product result between V _i and mu _2. Therefore, the comparison process is performed with the result R ₂ and a constant k _x. Here, if the result R ₂ is equal to or greater than the constant k _x ,
Input Group V _i of the data sample is determined to belong to the region 20-M. On the other hand, results when R ₂ is less than the constant k _x is input group V _i of data samples area 20-N
Is determined to belong to

According to FIG. 4B, input group V _i of data samples belong to the region 20-M. Incidentally, as shown in FIG. 4C, the region 20-M by the the border 22-Y expressed in k _y mu ₁ is divided into two areas 20-P and 20-Q. Therefore, the input group V _i of the data sample
There to decide belongs to either the region 20-P and the region 20-Q, obtaining the inner product between V _i and mu _1. Here, when the inner product result R _y is less than the constant k _y is input group V _i of the data sample is determined to belong to the region 20-P. On the other hand, if the result R _y is less than the constant k _y is input group V _i of the data sample is determined to belong to the region 20-Q.

The above inner product processing and comparison processing are repeated.
And the input vector V_i Is determined,
Data sample input group V_iIs the quantum center of gravity
Be transformed into FIG. 5 is a flowchart of the above-described quantization processing.
It is shown in the format. This flowchart shows that 30
-1, 30-2, 30-3, etc.
With two branches, “YES” and “NO”
It has a tree structure. For each of these nodes,
Input group V_i And μ₁ Or μ_Two Calculate inner product with
And the inner product result and the constant k_i Is performed. here
And the inner product result is a constant k _i If this is the case, the “YES” branch
Otherwise, the "NO" branch is selected
You.

Note that, in the flowchart of FIG.
3A to 3D show the values of μ and k in FIG.
This shows the processing result. That is, the highest node 30-
3A as k and μ of 1₁ And μ₁ To the node 30-
K of FIG. 3B as k and μ of 2 _Two And μ_Two To the node 30-
K in FIG. 3C as k and μ in FIG._Three And μ_Two Is given. Ma
In addition, as k and μ of the node 30-3, k of FIG._x And μ_Two 
As k and μ of the node 30-5, and k of FIG._y And μ_Two
Is given.

By descending the node of the tree shown in FIG. 5 from the top node 30-1 to the end node, the area to which the input group of the data sample belongs is determined. Here, if the tree is symmetric, the tree will have N = 2 ⁿ endpoints. In this case, the number of times of descending the tree node is n, and the highest node 30-1
To descend the tree from to the end node, n inner product processes and n comparison processes are performed.

As the quantization process, a flowchart shown in FIG. 6 can be used in addition to the process shown in FIG. The flowchart of FIG.
The tree structure has a plurality of nodes such as 2, 40-3, and two branches “YES” and “NO” from each node. However, that tree of Figure 6 differs from the tree of FIG. 5, reference group mu ₁ and input group V _i of advance data samples before prior to the processing of the highest node 40-1, advance to calculate the inner product of the mu ₂ Is that the inner product result is held in registers REG ₁ and REG ₂ . That is, for each node of the tree, the value of the register REG ₁ or REG ₂ is compared with the constant k _i, and a branch is selected based on the result of the comparison.

Both the processing shown in FIG. 5 and the processing shown in FIG. 6 include a case where two or more data samples are included as a group and two or more are used as a reference group that defines a boundary between a plurality of regions 20. Applicable. Here, “r” is the number of data samples for each group, “s”
Is the number of reference groups that define the boundary between the regions 20. Then, in FIG. 5, the reference group in each _{node, μ 1, μ 2, ...} , μ selected from among mu _s
x , and the inner product of V _i and μ _x is the first sample of V _i and μ
the product of the first sample of _x, the sum of the product of the product of the second sample of the second sample and mu _x of V _i, ... the final sample and mu _x final sample of V _i.

Similarly, when the processing shown in FIG. 6 is generalized,
Dot product processing for the input group V _i of the data samples is performed for all groups from the reference group μ ₁ to μ _s ,
Holding these inner product results from the register REG ₁ to REG _s. These inner product processes are performed at the highest node 40-1.
Conducted prior to treatment and subsequently while comparison between the value and the constant k _i of the REG for each node of the tree, processing continues until the end point node of the tree.

The main feature of the generalized process of FIG. 6 is that the number of inner product processes is minimized when the number of nodes from the highest node to the end node is larger than the number of reference groups. That is, in the case of a tree structure having symmetry, the processing of FIG. 6 minimizes the number of inner product processing when n is s or more. On the other hand, in the process of FIG.
When the number of nodes from the highest node to the end node is smaller than the number of reference groups, the number of inner product processes is minimized.

[0036] As another feature of the process of FIG. 5 and FIG. 6, reference group mu ₁ with constant k _i, ..., include that small amount of memory required to hold the mu _s. In order to hold the constant k _i , it is sufficient if there is enough memory to hold N−1 scalar amounts, and to hold reference groups, it is sufficient to hold s × r scalar amounts. If there is.

One method of reducing the amount of memory by limiting "s" (total number of reference groups) is to keep "s" to 2 log ₂ N or less. As another method, “s” can be suppressed to 2r or less. By providing a constraint on "s" in this manner, it is possible to reduce the memory amount by several thousand percent compared to the case where "s" is unlimited and (N-1). At the same time, according to the computer simulation, the increase in distortion caused by constraints on "s" is not more than 5 percent, rarely a problem when N is 2 ¹⁰ or more. Furthermore, it is also possible to eliminate the increase in distortion here simply by increasing N.

As described above, if the direction of the boundary line of the quantization area 20 is not constant, the reference group μ _i (i = 1,..., N−
1) must be given. In this case, in order to hold all the reference groups, a memory amount for holding (N−1) × r scalar amounts is required. Since it is desired to reduce the number of nodes (N-1) in the tree to several thousands in order to reduce the distortion, the increase in the amount of memory here is extremely large.

As an actual numerical example of the reduction of the memory amount, as shown in FIG. 5 or FIG.
Is 2 ^16, that is, 65,536. Consider a 4 × 4 16-pixel group from the image array as an input group of data samples, and assume that there are 16 reference groups that define N regions in which the input group of data samples is quantized. That is, N = 65,536 and r = s = 16. At this time, in order to hold the reference group in the processing shown in FIG. 5 or FIG. 6, it is sufficient if there is enough memory to hold 16 × 16 scalar amounts. On the other hand, when the direction of the boundary of the quantization area is not constant, a memory amount for holding 65,535 × 16 scalar amounts is required.

FIG. 7 shows the overall configuration of an electronic circuit for executing the quantization processing of FIG. This circuit includes a plurality of modules numbered 50 to 59. The module 50 is an electronic sensor that measures a physical phenomenon and outputs an electric output signal corresponding to the measured phenomenon to the lead 50a. For example, the sensor 50 measures a reflected pulse from a radar, an optical image, an infrared image, an acoustic signal, a signal obtained by mixing these, and the like.

The module 51 is a sampler that periodically samples an output signal from a sensor on the lead 50a. The module 52 is a hold circuit, which is a circuit for receiving a sample signal from the sampler 51 and temporarily holding the sample signal. Specimens in module 52, selectively accessed by the input group V _i by a module 53 is produced. The input group V _i of each data sample is accessed in response to a control signal START from a module 54 which is a sequence controller such as a microprocessor. The input group of each data sample is sent from module 53 to module 55 via bus 53a.
The module 55 is an arithmetic unit, and is a part that executes inner product processing. A second input bus 56a from the memory module 56 is connected to the module 55.
The memory 56 holds all the reference groups μ ₁ ,..., Μ _s and receives the address signal A from the controller 54.
1 is selectively accessed.

The result of the inner product is sent from the module 55 to the module 57 as a comparison circuit via the bus 55a.
At the same time, the comparison circuit 57 is connected to the output bus 58 of the module 58.
Input a constant k _i from a. Module 58 Hamou is one memory, the constant k _i in the memory 58 is selectively accessed by the address signal A2 from the controller 54.

The module 57 has an output line 57a,
A digital comparison signal “C” is generated. If the inner product result on the bus 55a is equal to or greater than the constant k _i , the comparison signal becomes “1”, and otherwise becomes “0”. The comparison signal C is supplied from the module 57 to the n flip-flops 59-1,.
It is sent to the module 59 having 59-N. Clocks are individually input to these flip-flops by the controller 54.

The processing is started when the controller 54 sends a START signal to the module 53. In response, module 53 sends the dot product module 55 outputs an input group V _i of data samples to an output bus 53a. Module 54 then sends the A1 address signal to memory 56 to access one of the reference groups μ ₁ ,..., Μ _s . Further, the module 54 sends the A2 address signals to the memory 58, accesses one of the constants k _i.

Subsequently, the modules 55 and 57 execute the above-described processing to generate the first comparison signal C1. The first comparison signal C1 is clocked by the controller 54 and input to the flip-flop 59-1. Based on the state of the comparison signal C1 in the flip-flop 59-1, the controller 54 sends the A1 signal to the memory 56 to select a new reference group and, at the same time, sends the A2 signal to the memory 58 and sends a new constant k _i . select. As a result, a second comparison signal C2 is generated, and the comparison signal C2
Is clocked by module 54 and input to flip-flop 59-2.

Next, flip-flops 59-1 and 59-
2, based on the previous comparison signals C1 and C2, module 54 generates new A1 and A2 signals. Then
The third comparison signal C3 is clocked and input to the flip-flop 59-3. Such processing is repeated until all the flip-flops 59 are loaded with the comparison signal C. When all of the flip-flop 59 to the comparison signal C is loaded, the index I ₁ input group V _i of data samples that state is input of the flip flop 59 indicates the center of gravity of the region 20 to be quantized, ..., I _n Becomes

FIG. 8 shows the overall configuration of an electronic circuit for executing the quantization processing of FIG. Some of the modules in FIG. 8 are the same as those in the circuit of FIG. 7, and have the same reference numerals. These common modules are 50, 51, 52, 53, 55, 57, 5
8 and 59. However, the circuit of FIG. 8 includes "s" arithmetic modules 55. Further, a plurality of registers 61-1 in FIG. 8, ..., includes a 61-s, each reference group mu _1, ..., holds the mu _s. Further, FIG.
Includes a multiplexer 62.

The arithmetic modules 55-1,..., 55-s proceed in parallel. Each operation module has an input group V _i of data samples and a reference group μ ₁ . …, Execute inner product processing with one of _μs . The inner product result calculated by the operation module is sent to the multiplexer 62, and one of the
One is selected by multiplexer 62 and sent to comparison module 57 via output 62a. Here, the selection of the inner product result by the multiplexer 62 is performed by the controller 5.
4 based on the address signal A1. Also,
Selection of Constant k _i to be sent to the comparison module 57 is performed based on the address signal A2 from the controller 54.

The two preferred processing examples for quantizing an input group of data samples have been described in detail above. Also, two suitable examples of electronic circuits for performing such a quantization process have been described in detail. However, various changes and modifications can be made to these preferred processing examples and circuit examples without departing from the spirit of the present invention.

For example, in the quantization processing of FIGS. 5 and 6, any number of nodes or any number of end-point nodes may be used. In the processing of FIGS. 5 and 6, the tree structure may be symmetric or asymmetric. Further, in the processing of FIGS. 5 and 6, it is possible to simplify the inner product processing by adding a constraint that only one of the reference groups μ has a non-zero element. The constraints on such a reference group, mu ₁ is on the S1 axis, mu
₂ on the S2 axis, ..., mu _r will be located on the Sr-axis, the inner product processing can be performed in a single multiplication.

Further, in the quantization circuits of FIGS. 7 and 8, it is possible to increase the total number "s" of the reference groups to N / 2. Even in this case, when the direction of the boundary line of the area 20 is not fixed. Less than half the amount of memory. The present invention is not limited to the preferred processing examples and circuit examples described above, but is indicated by the gist set forth in the claims.

[0052]

According to the present invention, it is possible to provide a novel quantization method and a new quantization apparatus capable of quantizing an input group of data samples with a very small storage capacity and a small number of processes as compared with the conventional apparatus. That is, the storage capacity of the storage module can be reduced by several thousand percent as compared with the conventional quantization method and apparatus. This is possible by limiting the total number of reference groups in the storage module to 2 log ₂ N or less. If r is the number of samples in the input group of data samples, it is possible to limit the total number of reference groups in the storage module to 2r or less.

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of an input group of data samples to be quantized, each group being composed of two samples.

FIG. 2 is a diagram showing how an input group of data samples shown in FIG. 1 is divided into a plurality of regions using the quantization method of the present invention.

FIG. 3A is a diagram showing a processing result until an area shown in FIG. 2 is obtained;

FIG. 3B is a diagram showing a processing result until an area shown in FIG. 2 is obtained.

FIG. 3C is a diagram showing a processing result until the area shown in FIG. 2 is obtained.

FIG. 3D is a diagram showing a processing result until the area shown in FIG. 2 is obtained.

FIG. 4A shows an input group V _{i of} an arbitrary data sample shown in FIG. 2;
FIG. 8 is a diagram for describing a preferred example of processing for quantizing into one of the regions.

FIG. 4B shows an input group V _{i of} an arbitrary data sample shown in FIG.
FIG. 8 is a diagram for describing a preferred example of processing for quantizing into one of the regions.

FIG. 4C shows an input group V _{i of} an arbitrary data sample shown in FIG. 2;
FIG. 8 is a diagram for describing a preferred example of processing for quantizing into one of the regions.

FIG. 5 is a flowchart illustrating a quantization processing procedure in FIGS. 4A to 4C.

FIG. 6 is a flowchart illustrating another preferred quantization processing procedure as a configuration method different from that of FIG. 5;

FIG. 7 is a diagram illustrating a quantizer that executes the quantization processing of FIG. 5;

FIG. 8 is a diagram illustrating a quantizer that executes the quantization processing of FIG. 6;

Continuation of the front page (56) References JP-A-59-77730 (JP, A) JP-A-61-43875 (JP, A) JP-A-62-76992 (JP, A) JP-A-64-44500 (JP, A) , A) JP-A-2-69073 (JP, A) JP-A-2-109468 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) H03M 7/30

Claims (22)

(57) [Claims] 1. An input group of data samples having substantially more than N possible values, where N is a predetermined positive integer, the centroid of one of N regions divided from said values. An input circuit for receiving an input group of said data samples, said product μ · k being orthogonal to and at a boundary between at least two of said N regions. Storage means for storing a reference group μ and a plurality of constants k of a plurality of data samples which are selected to arrive, each not exceeding N / 2 and N-1; and Control means for selecting one of the reference groups μ _x of the data samples and one of the constants _ky , which is connected to the input circuit and the storage means, wherein the input group of the data samples and the reference group μ _x Calculating means for calculating the inner product of Serial connected to the calculating means, and comparing means for comparing the constant k _y and inner product result by the calculating means, connected to said comparing means, for defining the center of gravity by the respective all comparison result by the comparison means logic and means, the control means is further connected to said comparing means, each node, the comparison of k _y with identifying and said comparing means for mu _x and said comparing means for said computing means Indicate which branch to take based on the result, each μ _x and k _y is selected corresponding to a selection tree having a plurality of the nodes interconnected by branches, and quantization apparatus characterized by repeatedly selecting a different mu _x and k _y are input to the comparing means. 2. The quantization apparatus according to claim 1, wherein the total number of reference groups of data samples in said storage means is ₂ log ₂ N or less. 3. The quantization apparatus according to claim 1, wherein, if r is the number of samples in an input group of data samples, the total number of reference groups of data samples in said storage means is 2r or less. 4. An input group of data samples having substantially more than N possible values, where N is a predetermined positive integer, the centroid of one of N regions divided from said values. Receiving the input group of data samples from an electronic input circuit, such that the product μ · k reaches orthogonal to at least two boundaries of the N regions. And each is N
/ And two and reference groups of data samples (N-1) does not exceed the mu and step for storing a plurality of the constant k in the storage means, and one mu _x reference group of data samples from said memory means a step of selecting one of said constant k _y, the reference group and the input group of data samples mu _x
A step of obtaining the inner product of the step of comparing the inner product result and k _y obtained by the inner product, each node, and k _y for the mu _x and the comparing step for said calculating step identified as while, indicates take branches either on the basis of the comparison result of the comparing means, each mu _x and k
and the _y, is selected to correspond to the selection tree having a plurality of said nodes are interconnected by branches, while selecting the mu _x and k _y for each iteration, the selection step and the inner product step and comparison step Is repeated several times, and a step of defining all the comparison results to define the center of gravity is provided. 5. Only one of each reference group of the data sample
5. The method according to claim 4, wherein only one sample has a non-zero value. 6. The method according to claim 4, wherein each input group of data samples is a group of pixels. 7. The method of claim 4, wherein each input group of data samples is a sample group from a signal reflected from a radar. 8. The method of claim 4, wherein each input group of data samples is a sample group from an infrared signal.
The quantization method as described. 9. The method of claim 4, wherein each input group of data samples is a sample group from an audio signal. 10. The method of claim 4, wherein each input group of data samples comprises a mixture of samples from a multi-frequency spectrum. 11. The quantization method according to claim 4, wherein the total number of reference groups of data samples in said storage means is ₂ log ₂ N or less. 12. When N is a predetermined positive integer, an input group of data samples having substantially more than N possible values is divided into a centroid of one of N regions divided from said values. An input circuit for receiving an input group of said data samples; a storage means for storing N-1 constants k; and a reference to N / 2 or less data samples. Register means for holding a group μ; wherein the product μ · k is selected so as to reach a boundary orthogonal to at least two boundaries of the N regions, and the input circuit and the storage means Computing means for calculating in parallel the respective inner products of the input group of the data samples and the reference group μ of all the data samples; and one of the inner product results connected to the computing means and the storage means R _x and before Selection means for selecting one k _y constants, connected to said selecting means, the selected inner product result R _x
Comparing means for comparing the selected constant k _y, connected to said selecting means, each node, as well as identify and of R _x and k _y of the selection means, based on the comparison result of the comparing means and indicates take either branch, selecting each of R _x and k _y is in is selected corresponding to the selection tree having a plurality of said nodes interconnected branches, and a different of R _x and k _y Control means for repeatedly activating the selecting means, and logic means connected to the comparing means and defining the center of gravity from each of all the comparison results by the comparing means. apparatus. 13. The quantization apparatus according to claim 12, wherein a storage capacity of said register means is limited so as to hold no more than ₂ log ₂ N reference groups of said data samples. 14. The quantization according to claim 12, wherein, when r is the number of samples in an input group of data samples, the total number of reference groups of data samples in said register means is 2r or less. apparatus. 15. An input group of data samples having substantially more than N possible values, where N is a predetermined positive integer, the centroid of one of N regions divided from said values. Receiving an input group of the data samples from an electronic input circuit; storing N-1 constants k in a storage means; Holding the reference group μ in the register means, wherein the product μ · k is selected to reach at least two boundaries of the N regions orthogonally to the boundary, and a step of obtaining a respective inner product of the reference group μ input group and all of the data sample, and step of selecting one k _y of one R _x and the constant of the inner product result, the inner product result selected R _x and selected constant _ky A step of comparing the bets, each node, as well as identify and of R _x and k _y for the selection step, indicates take either branch based on the comparison result of the comparing step, and each R _x and k _y are selected to correspond to the selection tree having a plurality of said nodes are interconnected by branches, while selecting and of R _x and k _y for each iteration,
A quantization method comprising: a step of repeating the selection step and the comparison step a plurality of times; and a step of defining the center of gravity by combining all of the comparison results. 16. The method according to claim 15, wherein only one sample in each reference group of data samples has a non-zero value. 17. The method of claim 15, wherein each input group of data samples is a group of pixels. 18. The method of claim 15, wherein each input group of data samples is a sample group from a signal reflected from a radar. 19. The method of claim 15, wherein each input group of data samples is a group of samples from an infrared signal. 20. The method of claim 1, wherein each input group of data samples is a sample group from an acoustic signal.
5. The quantization method according to 5. 21. The method of claim 15, wherein each input group of data samples comprises a mixture of samples from a multi-frequency spectrum. 22. The quantization method according to claim 15, wherein the total number of reference groups of data samples in said storage means is ₂ log ₂ N or less.