JPH0695754B2 - Dynamic vector encoding / decoding method, dynamic vector encoder and decoder - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a vector quantizer that collects a plurality of input signal sequences into blocks, and quantizes each block in a multidimensional space.

[Conventional technology]

Figures 4 and 5 show, for example, Technical Report IT85-61 of the Institute of Electronics and Communication Engineers.
(1985) is a block diagram showing a configuration example of a vector quantization encoder and a decoder using the conventional vector quantization technique described in "Image High Efficiency Coding by Vector Quantization" and the like. In FIG. 4, (1) is an input vector, (2) is an average value separation circuit that separates the in-vector average value component from this input vector, (3) is the average value in the vector, and (4) is the average value encoder. , (5) is an average value separation input vector, (6) is an inner product calculation unit for obtaining an inner product value of the average value separation input vector and a normalized output vector described later,
(7) is a codebook ROM in which a plurality of normalized output vectors are described, (8) is an address signal, (9) is a normalized output vector, (10) is an address counter, and (11) is an inner product computing unit. The inner product value calculated in (6), (12) the maximum inner product detection unit that finds the maximum of a plurality of inner product values, (13) the correction amplitude defined by the maximum inner product value,
(14) is a strobe signal, (15) is an amplitude encoder, (1
6) is an index latch that takes in the address signal,
(17) is average value encoded data, (18) is amplitude encoded data, (19) is an index, (20) is an encoded data multiplexing unit, and (21) is encoder output data.

In the configuration example of the vector quantization decoder of FIG. 5,
(22) is a coded data demultiplexer, (23) is an average value decoder, (24) is an amplitude decoder, and (28) is a codebook RO.
M, (25) is the average decoded value, (26) is the amplitude decoded value, (27)
Is the decoding vector.

Next, the vector quantization coding operation will be described. The input vector (1) (hereinafter appropriately referred to as S ), which is a block of the input signal series for every k (k is an integer of 2 or more), is processed by the following equation in the average value separation circuit (2). The average value (3) in the vector (hereinafter appropriately referred to as μ) is separated and converted into an average value separation input vector (5) (hereinafter appropriately expressed as X ). .

The input vector S 1 , the average value μ in the vector, and the average value separation input vector X are represented by the respective equations. S = [S ₁ , S ₂ , ..., S _k ] x _j = S _j −μ X = [X ₁ , X ₂ , ..., X _k ] The average value μ is high-efficiency coded by the average value encoder (4), and the average value encoded data (27 ) Is converted to. The average value separation input vector X is input to the inner product calculating section (6) and converted into a corrected amplitude (13) and an index (19) described below through the following inner product vector quantization coding processing. First, let the average value separation input vector X be the size of the vector. Normalized output vector of N that is generated based on the normalization plurality normalized input vector X ^* statistical properties of the at (N is a natural number) (9) (hereinafter, y _{i (i} = 1,2, , N)) is written in the codebook ROM (7). When the average value separation input vector X is input to the inner product calculating unit (6), the address counter (10) is reset and the counter operation of the cycle N is started. next,
The normalized output vector y _i on the address indicated by the address signal (8) output from the address counter (10) is sequentially input from the codebook ROM (7) to the inner product calculating unit (6), An inner product (11) of the average value separation input vector X and the N normalized output vectors y _i (hereinafter expressed as P ( x , y _i ), i = 1, 2, ..., N) is expressed by the following equation. Is calculated and output according to.

In the maximum inner product detection unit (12), after detecting the maximum inner product value P _max of the N inner products P ( x , y _i ) obtained by the above calculation, the normalized output vector giving the maximum inner product value. y _i
The address signal (8) indicating the address in the codebook ROM (7) is fetched by the index latch (16) at the timing synchronized with the strobe signal (14). The fetched address signal (8) is a predetermined normalized output vector y _i.
Is sent to the encoded data multiplexing unit (20) as an index (19) for identifying the. On the other hand, the maximum inner product value P
_max is given as a correction amplitude (13) (hereinafter referred to as g) that approximates the magnitude | x | of the average value separation input vector x as shown in the following equation.

P _max = P ( x , y _i ) = | x ||| y _i | .cos θ _i = | x | cos θ _i = g The corrected amplitude g is high-efficiency coded by the amplitude encoder (15). , Amplitude encoded data (18). In the coded data multiplexing unit (20), the average value coded data (1
7), the amplitude coded data (18) and the index (19) are multiplexed and sent as encoder output data (21) according to a predetermined format.

Next, the decoding operation will be described. The encoder output data (21) is supplied to the encoded data demultiplexer (22),
The average value encoded data (17), the amplitude encoded data (18), and the index (19) are separated according to a predetermined format. The average value encoded data (17) is decoded by the average value decoder (23), and the average value decoded value (25) (hereinafter It is written as. ) Is converted to. Similarly, the amplitude decoder (24) outputs the amplitude decoded value (26) (hereinafter It is written as. ) Is output. Also, codebook ROM
(17) a normalized output vector y _i on the address indicated is read by Indetsukusu (19) from the decoded vector (27) for the input vector s through the processing of the following equation (hereinafter It is written as. ) Is obtained.

[Problems to be Solved by the Invention] Since the conventional vector quantizer is configured as described above, it is impossible to update the contents of the codebook during the encoding operation, and the peculiar input having different properties For a vector,
There is a problem that the degree of approximation of the decoded vector may decrease.

The present invention has been made to solve the above problems, and an object thereof is to obtain a vector quantizer capable of obtaining a decoded vector having a high degree of approximation even for a singular input vector.

[Means for solving problems]

In the dynamic vector coding / decoding method according to the present invention, a plurality of output vectors are fixed codebook based on a plurality of input vectors formed by dividing an input signal sequence into k units (k is an integer of 2 or more). A plurality of outputs that are not stored in the fixed codebook are stored in a dynamic codebook in an updatable manner; the degree of approximation between the output vector and the input vector stored in the fixed codebook and the dynamic codebook. Is calculated by the approximation degree calculation means; when the difference between the output vector giving the maximum approximation degree among the outputs of the approximation degree calculation means and the input vector is equal to or more than a set allowable value, the dynamic codebook update operation is performed. Output vector stored in the dynamic codebook When updated, the code equivalent to the update operation is sent, and when not updated, the code corresponding to the selected output vector is sent by the coded data multiplexing unit to perform coding, and the sent code The output data is separated into an identification code by a coded data demultiplexing unit in a predetermined format; a plurality of output vectors corresponding to predetermined code inputs similar to the above-mentioned encoding are stored in a fixed codebook; A plurality of outputs which are not stored in the codebook and are renewably stored in a dynamic codebook; the separated identification codes correspond to output vectors stored in the fixed codebook and the dynamic codebook. Read the corresponding output vector and obtain the decoded vector corresponding to the transmitted input vector; When serial separated identification code instructs the update of the dynamic codebook, the encoded output by replacing the output vector of said dynamic codebook in relation receiving input vector to perform decoding of data.

Further, the dynamic vector encoder and decoder according to the present invention include a plurality of output vectors based on a plurality of input vectors formed by blocking an input signal sequence into k units (k is an integer of 2 or more). A fixed codebook, a dynamic codebook that renewably stores a plurality of outputs that are not stored in the fixed codebook, an output vector stored in the fixed codebook and the dynamic codebook, and the input vector. And an approximation degree calculating means for determining the approximation degree of the dynamic codebook and controlling the update operation of the dynamic codebook when the difference between the output vector and the input vector that gives the maximum approximation degree among the outputs Dynamic codebook control unit and output stored in the dynamic codebook Vector is when the updated sends an identification code corresponding on updating operation, if not updated and a coded data multiplexing unit for sending the code corresponding related to the output vectors selected.

Also, the dynamic vector decoder is compatible with a coded data demultiplexing unit that separates the transmitted coded output data into an identification code in a predetermined format, and a predetermined code input similar to the coding unit. A fixed codebook that stores a plurality of output vectors, a dynamic codebook that stores up to a plurality of outputs that are not stored in the fixed codebook similar to the encoding unit, and a code that separates the identification code is a fixed code. When the output vector corresponding to the output vector stored in the book and the dynamic codebook is read, the corresponding output vector is read out, the means for obtaining the decoded vector corresponding to the transmitted input vector, and the separated identification code update the dynamic codebook. When instructing,
Means for replacing the output vector in the dynamic codebook with the associated received input vector.

In addition, if multiple output vectors in the fixed codebook are 2 ⁿ
(N is a natural number) Arranged in a tree-like manner, the inner product value with the input vector from the 2 ⁿ normalized output vectors for each branch from the upper hierarchy to the lower hierarchy of the tree structure A tree search maximum approximation degree detection unit that sequentially selects the output vector having the maximum?

Further, a normalization circuit that separates and normalizes the average value of the input vector is provided, and the average value separated input vector is normalized by the value given by the square root of the sum of squares of the sample values in the vector.

Further, the fixed codebook stores a plurality of output vectors whose mean is zero and whose magnitude is 1 which is determined by the statistical properties of the plurality of input vectors.

Also, the degree of approximation between the input vector and the decoded vector after vector quantization of the input vector is obtained by the difference between the square sum of the normalization coefficient and the square value of the maximum inner product value.

Further, when the identification code indicates the normalized output vector, the dynamic vector decoder multiplies the normalized output vector by the amplitude decoding and adds the product and the average value decoded value.

Further, when the dynamic vector decoder detects that the identification code indicates updating, the predetermined operation is to multiply the normalized input vector by the amplitude decoded value and add the product and the average decoded value. did.

[Action]

The content of the dynamic code book according to the present invention is adaptively updated by the dynamic code book control unit to temporarily store a small number of output vector groups having different characteristics from the output vector group stored in the fixed code book. To do.

Example of Invention

An embodiment of the present invention will be described below with reference to the drawings. First
In the figure, (28) is a dynamic code book RAM capable of writing and reading at any time, (29) is a normalization circuit for normalizing the average value separation input vector x by the magnitude | x | of the vector, (30) Is a normalization coefficient, and is denoted by σ below.
(31) is the normalized input vector, and is denoted as x ^* below.
(32) is a dynamic code book control unit for controlling the updating procedure of the dynamic code book RAM, (33)
Is a write request signal, (34) is a normalized input vector with index, (35a) (35b) is a selector, (36) is vector data, and in FIG. 2, (37) is a vector data decoding unit.

Next, the vector quantization coding operation in one embodiment of the present invention will be described with reference to FIG. The input vector s indicated by reference numeral (1) in FIG. 1 is converted into the average value separated input vector x indicated by reference numeral (5) in the same figure by the average value separation circuit (2), and the inner product calculation unit (6 ) And the normalization circuit (34). The average value is processed and transmitted in the same manner as the conventional method. The normalized output vector y _i indicated by the reference numeral (9) in the figure and read out from the fixed code book ROM (7) and the dynamic code book RAM (28) in the inner product calculating section (6) and the average value separation input vector. After finding the inner product p ( x , y _i ) with x in the same manner as the conventional method, the normalized output vector y _i that gives the maximum inner product value for the above-mentioned average value separation input vector x is detected, and this detected An address signal (8) indicating an address of the normalized output vector y _{i on} the fixed code book ROM or on the dynamic code book RAM is output from the index latch (16) as an index (19). Further, the maximum inner product value P _max is output from the maximum inner product detection unit (12) as a correction amplitude g indicated by reference numeral (13) in FIG. In the above process, the inner product calculating unit (6), the maximum inner product detecting unit (12), the address counter (10) and the index latch (16) perform the same operations as the conventional one.
As shown in FIG. 3, the dynamic code book RAM (28) is composed of a first-in / first-out type FIFO memory in which writing and reading operations are performed asynchronously.

It is updated at any time according to the update procedure controlled by the dynamic code block control section (32). Hereinafter, an example of a procedure for updating the contents of the dynamic code block RAM (28) will be described based on an operation example of the dynamic code block control unit (32). The average value separation input vector x is input to the normalization circuit (29), subjected to the processing of the following equation, and the normalization coefficient σ indicated by reference numeral (30) in FIG. 1 and the reference numeral (31) in the same figure. After being decomposed into a normalized input vector x ^{* represented} by, the two parameters σ and x ^* are supplied to the dynamic code book control unit (32).

^{_{x * j = x j / σ}} x * = ^{_{[x * 1, x * 2,}} ..., x * k ] In Dyna honey click code book control unit (32), using the correction amplitude g and the normalization factor sigma , The average input / output input vector x and the normalized output vector that is reproduced in amplitude during decoding
g. The similarity d ² between y _i calculated as follows, based on the size of the near similarity score d ^2, the Dyna Mitsu click code book RAM (2
8) Control the update procedure adaptively.

_{^{d 2 = (x -g y i}} ) 2 = | x | 2 + g 2 | y i | 2 -2g (x, y i) = σ 2 + g 2 -2g 2 (| y i | 2 = 1, (x , Y _i ) = g) = σ ² −g ² That is, when the approximation degree d ² is larger than the allowable value D ² , the write request signal (33) is set to “1” and the normalized input vector x ^* is set. While writing to the dynamic code block RAM (28), a normalized input vector (34) with an index added with a prefix indicating that the normalized input vector x ^* has been written is output. On the other hand, when the approximation degree d ² is less than or equal to the allowable value D ^{2, the} write request signal (38) is set to “0”.
Is only output as. The write control signal (33) is sent to the first and second selectors (35c) and is used to select the encoded data to be transmitted corresponding to the updating procedure of the dynamic code book RAM (28). To be That is, when the write request signal (33) is '0', the dynamic code block RAM (28) is not updated. Therefore, the first inner selector detects the maximum inner product detector (12) as in the conventional process. ) Is supplied to the amplitude encoder (15), and the second secretor (35b) is supplied with the corrected amplitude g (code (13)) from the index latch (16). ) As vector data (41) to the encoded data multiplexing unit (20). When the write request signal (33) is '1', the dynamic code block RAM is also updated on the receiving side, and at the same time, the normalized output vector y _i at the time of decoding is changed to the normalized input vector x ^*.
(The code (31)) and transmits the normalization coefficient σ (code (30)) obtained in the normalization process.
Selector (35a) of the normalization coefficient σ to the amplitude encoder (1
The second selector (35b) supplies the indexed normalized input vector (34) as vector data (36) to the encoded data multiplexing unit (20). The amplitude encoder (15) operates similarly to the conventional one.

After the above processing, the average value encoded data (17), the amplitude encoded data (18) and the vector data (36) are multiplexed in the encoded data multiplexing unit (20) according to a predetermined format, It is output as encoder output data (21).

Next, referring to FIG. 3, the dynamic code block RAM
The operation example of (28) will be described. The read control unit (28d)
The operation of reading the data on the address designated by the read address pointer, that is, the normalized output vector y ^* is sequentially repeated.

When the write request signal (35) is '1', the write control unit (28b) writes the write data, that is, the normalized input vector x ^* at the address designated by the write address pointer and at the end of writing. The address counter (28a) repeats the operation of counting up the address by 1 and resetting it to 0 when the address exceeds the maximum value. By executing the above operation, it is possible to use a certain number of the latest normalized input vector x ^* determined by the memory capacity as the normalized output vector y _i for vector quantization.

Next, the vector quantization decoding operation in the embodiment of the present invention will be described with reference to FIG. The encoder output data (21) is converted into the average value encoded data (1
7), the amplitude coded data (18) and the vector data (36) are separated.

The average value encoded data (17) and the amplitude encoded data (18) are decommissioned in the same manner as in the conventional method, (Reference numeral (29) in FIG. 2), (Reference numeral (30) in the figure). In the vector data decoding unit, it is determined whether the separated vector data (41) matches the prefix added to the normalized input vector. If they match, the write request signal (33) is set to '1. '
Then, the write request signal (33) and the normalized input vector x ^* (reference numeral (31) in FIG. 2) received after the prefix are output. If they do not match, the write request signal (33) is output. (33) is set to '0' and vector data (36)
Is output as the index (19). Fixed codebook ROM (7) and dynamic codebook RAM (33)
Is the normalized output vector y _i (reference numeral (9) in FIG. 2) corresponding to the index (19) to the third selector (35
Supply to c). Further, when the write request signal (33) is '1', the normalized input vector x ^* (reference numeral (31) in FIG. 2) is written in the dynamic code book RAM (28) to perform the encoding operation. The same dynamic codebook as
The RAM (28) is updated. Third selector (35c)
Then, when the write request signal (33) is '0', the normalized output vector y _i (reference numeral (9) in FIG. 2) is output, and the write request signal (33) is '1'. At this time, the normalized input vector x (reference numeral (25) in FIG. 2) is output. By multiplying the vector output from the third selector (35c) by the amplitude decoded value (26) and adding the average value decoded value (25) (Reference numeral (27) in FIG. 2) is obtained.

In the above embodiment, the mean value separation normalized vector quantizer in which the mean value of the input vector is separated and the output vector stored in the codebook is mean value separated and normalized is described. Is directly vector-quantized, and the same effect can be obtained when updating the dynamic code block RAM based on the vector quantization error.

〔The invention's effect〕

As described above, according to the present invention, the codebook of the vector quantizer is composed of a dynamic codebook which can be written / read at any time and a conventional fixed codebook, and the codebook based on the approximation of the input vector and the decoded vector is used. Since a means for executing vector quantization coding / decoding while adaptively updating the contents of the dynamic codebook is used, for a peculiar input vector having a property different from the output vector group stored in the fixed codebook, The effect of suppressing the decrease in the degree of approximation of the decoded vector is obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an encoding unit of a dynamic vector quantizer according to an embodiment of the present invention, and FIG. 2 is a decoding unit of a dynamic vector quantizer according to an embodiment of the present invention. FIG. 3 is a block diagram showing the configuration, FIG. 3 is a block diagram showing the configuration of the dynamic code block according to one embodiment of the present invention, and FIG. 4 is a block diagram showing the configuration of the encoding unit of the conventional vector quantizer. FIG. 5 is a block diagram showing the configuration of the decoding unit of the conventional vector quantizer. In the figure, (1) is an input vector, (2) is an average value separation circuit, (3) is an average value, (4) is an average value encoder,
(5) is an average value separation input vector, (6) is an inner product computing unit, (7) is a fixed codebook ROM, (8) is a dynamic codebook RAM, (9) is a normalized output vector, and (1
7) is the normalization circuit, (18) is the normalized input vector, (1
9) is the normalization coefficient, (23) is the vector data decoding unit, (1
7) is average value coded data, (18) is amplitude coded data, (19) is index, (20) is coded data multiplexer, (21) is encoder output data, and (32) is dyna. The mix code book control unit (34) is a normalized input vector with an index. In the drawings, the same reference numerals indicate the same or corresponding parts.

Claims (9)

[Claims] 1. K input signal sequences (k is an integer of 2 or more)
A plurality of output vectors are stored in a fixed codebook based on a plurality of input vectors formed by blocking for each; a plurality of outputs that are not stored in the fixed codebook are stored in a dynamic codebook updatable; The approximation degree between the output vector and the input vector stored in the fixed codebook and the dynamic codebook is calculated by the approximation degree calculating means; the output vector and the input vector giving the maximum approximation degree among the outputs of the approximation degree calculating means. When the difference between and is equal to or more than a set allowable value, the dynamic codebook update operation is controlled by the dynamic codebook control unit; when the output vector stored in the dynamic codebook is updated, the identification code related to the update operation The output vector selected if not updated Coding is performed by sending the code corresponding to the relevant code by the coded data multiplexer, and the transmitted coded output data is separated into an identification code by the coded data demultiplexer in a predetermined format; A plurality of output vectors corresponding to a predetermined code input similar to the above are stored in a fixed codebook; a plurality of outputs which are not stored in the codebook similar to the above-described encoding are rewritably stored in a dynamic codebook. When the separated identification code corresponds to the output vector stored in the fixed codebook and the dynamic codebook, the corresponding output vector is read to obtain a decoded vector corresponding to the transmitted input vector; When the separated identification code indicates the update of the dynamic codebook, the associated receive input vector Dynamic vector coding decoding method for performing decoding of the encoded output data by replacing the output vector in serial dynamic codebook. 2. k input signal sequences (k is an integer of 2 or more)
A fixed codebook that stores a plurality of output vectors based on a plurality of input vectors formed by blocking for each, a dynamic codebook that stores a plurality of outputs that are not stored in the fixed codebook in an updatable manner; Approximation degree calculating means for obtaining an approximation degree between the output vector and the input vector stored in the fixed codebook and the dynamic codebook, and an output vector and an input vector giving a maximum approximation degree among the outputs of the approximation degree calculating means. And a dynamic codebook control unit that controls the updating operation of the dynamic codebook when the difference between the values is greater than or equal to a set allowable value, and an identification code related to the updating operation when the output vector stored in the dynamic codebook is updated. Output, and the selected output vector if not updated Dynamic vector encoder with an encoded data multiplexing unit for sending the code corresponding related. 3. A coded data demultiplexing unit for separating the transmitted coded output data into an identification code in a predetermined format,
A fixed codebook that stores a plurality of output vectors corresponding to a predetermined code input similar to the encoding unit, and a plurality of outputs that are not stored in the fixed codebook similar to the encoding unit and can be updated. When the dynamic codebook and the separated identification code correspond to the output vector stored in the fixed codebook and the dynamic codebook, the corresponding output vector is read, and the decoded vector corresponding to the transmitted input vector. And a means for replacing an output vector in the dynamic codebook with an associated received input vector when the separated identification code indicates an update of the dynamic codebook. 4. A plurality of output vectors in a fixed codebook are arranged in a 2 ⁿ (n is a natural number) binary tree shape, and 2 ⁿ pieces are provided for each branch from the upper hierarchy to the lower hierarchy of the tree structure. The tree search maximum approximation degree detection unit that sequentially selects the output vector having the maximum inner product value with the input vector from among the normalized output vectors of A dynamic vector encoder according to item. 5. A normalization circuit that separates and normalizes the average value of the input vector is provided, and the normalization circuit normalizes the average value separation input vector by a value given by the square root of the sum of squares of sample values in the vector. The dynamic vector encoder according to claim 2, characterized in that 6. The fixed codebook stores a plurality of output vectors having a mean of zero and a magnitude of 1, which are determined by the statistical properties of a plurality of input vectors. The dynamic vector encoder according to item 2. 7. An approximation degree between an input vector and a decoded vector after vector quantization of the input vector is obtained by a difference between a square sum of normalization coefficients and a square value of a maximum inner product value. 5. A dynamic vector encoder according to claim 2. 8. When the identification code indicates a normalized output vector, the predetermined operation multiplies the normalized output vector and the amplitude decoding, and adds the product and the average value decoded value. 5. A dynamic vector decoder according to claim 3. 9. When it is detected that the identification code indicates an update, the predetermined operation multiplies the normalized input vector by the amplitude decoded value and adds the product and the average decoded value. A dynamic vector decoder according to claim 3.