JPH0369451B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an adaptive vector quantizer that blocks a plurality of signal sequences and quantizes the blocks in a multidimensional signal space. [Prior Art] First, the principle of vector quantization will be briefly explained. Now, K input signal sequences are put together into an input vector X = {x ₁ , x ₂ , . . . x _K }. At this time, N representative points (i.e., output vectors) of the K-dimensional Euclidean signal space R ^K ( X ∈R ^K ) y _i =y _i1 ,
y _i2 ,...y _iK } is set as Y=[ y ₁ , y ₂ ,...,
_yN ]. From the set of output vectors, the vector quantizer determines the output vector y _i that is the shortest distance from the input vector (resulting in the least distortion) as follows, and searches for this. _if , d( X , yi ) _< d( X , yj ) for all j X → yi , where d( X , yi ₎ is the distance (distortion) between the input and output vectors _. At this time, the input vector X is transmitted or recorded according to the output vector index i, and is replaced by the output vector y _i during playback. The set Y of output vectors y _i becomes the training model. It can be obtained by clustering using a signal sequence (selection of representative points and quantization of each representative point of the training model are repeated until the total sum of distortion is minimized).
Furthermore, in order to improve the efficiency of vector quantization and the versatility of the set of output vectors, the average value of the vector is separated and vector quantized after being normalized by amplitude. A conventional vector quantizer will be described below with reference to a specific example of its configuration. FIG. 4 shows an example of an encoding section. In this configuration example, the absolute value distortion is expressed by equation (1) as an example of the definition of d( X , yi ₎ . d ( X , y _i ) _{= K} 〓 _j ⁼¹ |
is an average value separation normalization circuit, 3 is a normalized input vector, 5 is an average value component of input vector 1, 4 is an amplitude component, 6 is a normalized input vector register, 7 is a code table address counter, and 8 is a code table address , 9 is a normalized output vector code table memory, 10 is a normalized output vector register, 11 is a parallel subtracter, 12 is a parallel absolute value calculator, 13 is an absolute value distortion calculation circuit (accumulator), 14 is a minimum distortion normalization output vector detection circuit,
16 is a strobe signal, 17 is an index latch, 18 is an encoder for collectively encoding the vector quantization index and the average value component and the amplitude component, and 22 is an output signal from the encoding section. Also,
FIG. 5 is an example of a decoding section. In the figure, 23 is a decoder that decodes the vector quantization index, the average value component, and the amplitude component, 24 is a normalized output vector, 25 is a mean value separation/normalization restoration circuit, and 26 is an output vector. Next, the operation will be explained. First, an input vector 1, S = {S ₁ , S ₂ , ...S _K }, which is a block of multiple input signal sequences, is divided into an average value component 5 and an amplitude component by an average value separation and normalization circuit 2. Lucky ingredient 4
and the normalized input vector 3, X = {x ₁ ,
x ₂ ,...x _K } is formed. The average value component 5 is m,
If the transfer component 4 is σ, then, for example, when absolute value transfer is used, it is expressed by equation (2). The average value separation and normalization process brings the input vector closer to a constant distribution in the signal space, thereby increasing the efficiency of vector quantization. Normalized input vector 3 is latched into normalized input vector register 6. The code table address counter 7 sequentially reads the normalized output vector y _i from the normalized output vector code table memory 9 and latches it into the normalized output vector register 10 . The absolute value distortion calculation circuit 13 obtains the distortion of X and y _i from the parallel subtracter 11 and the parallel absolute value calculator 12 as follows. _{d i} = d ( X , y _i ) = _K 〓 j ₌ ¹ _| Find the minimum value of d _i . In other words, the minimum distortion is expressed by equation (4). d=min id _i (4) When the normalized output vector resulting in the minimum distortion is detected, the strobe signal 16 is sent to the index latch 17, and the code table address 8, which is the address of the vector, is taken in. The encoder 21 determines the index of the normalized output vector resulting in the minimum distortion;
The average value component 5 and amplitude component 4 of the input vector are encoded together and outputted as an encoder output signal 22. In the decoding section, the decoder 23 decodes the vector quantization index, the average value component, and the amplitude component,
The normalized output vector y _i is read from the normalized output vector code table 9 and latched into the register 10. Furthermore, the average value separation normalization restoration circuit 25
Accordingly, the output vector 26, S' = {S' ₁ , S' ₂ , . . . S' _K }, is decoded using the average value component 5 and the amplitude component 4. That is, S′ _j =σ・y _ij +m ...(5) (j=1, 2, ..., K) [Problem to be solved by the invention] The conventional vector quantizer is configured as described above. Therefore, there was a problem of mismatching of the quantizer. In other words, it is impossible to obtain a general-purpose set of output vectors that can correspond to all input signal sequences, so several adaptive control methods have been attempted, but they tend to have the problem of increasing the equipment scale. Ta. This invention was made to solve the above-mentioned conventional problems, and by correcting the amplitude when reproducing the normalized output vector using quantization distortion, the above-mentioned quantum The purpose of this paper is to obtain an adaptive vector quantizer that eliminates mismatching of the quantizer. [Means for solving the problem] The adaptive vector quantizer according to the present invention is used when normalizing an input vector using quantization distortion calculated in the process of vector quantization. This is provided with an amplitude correction circuit that corrects the encoded and transmitted amplitude value. [Operation] In this invention, after the input vector is average-separated and normalized, vector quantization is performed, the amplitude value is corrected using quantization distortion, and the corrected amplitude value and vector quantization index are , and the average value are encoded and transmitted. [Example] Hereinafter, an example of the present invention will be described with reference to the drawings. In FIGS. 1 and 2, the same parts as in FIGS. 4 and 5 are designated by the same reference numerals. In FIGS. Quantization distortion, 18 is a vector quantization index, 19 is an amplitude correction circuit, and 20 is a corrected amplitude value. Also, the second
The figure shows an example of the configuration of a decoding section in an adaptive vector quantizer according to the present invention. Next, the operation will be explained. A signal sequence input as a K-dimensional input vector 1, S = {s ₁ , s ₂ , ..., s _K } is normalized as an input by the mean value separation and normalization circuit 2 as shown in equation (6). It is separated into a vector 3, X , an amplitude component 4, σ, and an average value component 5, m. Here, we took the absolute value amplitude as an example, but σ=[K ^-1 _K 〓 ^j=1 (s _j −m) ² ] ^1/2 (standard deviation) σ=1/2 [ max j (s _j −m)−min j(s _j −m)] (peak-to-peak value), etc. may of course be used. The normalized input vector 3, X is latched into the normalized input vector register 6. The code table address counter 7 sequentially reads the normalized output vector y _i from the normalized output vector code table 9 according to the code table address 8 and latches it into the normalized output vector register 10 . The absolute value distortion calculation circuit 13 includes a parallel subtracter 11 and a parallel absolute value calculator 1
2, the absolute value d _i is calculated by equation (7). _{d i =} ^K _{〓 j} ⁼ _{1 |} −y _ij ) ² (square distortion) d _i = max j | x _j −y _ij | (maximum element distortion), etc. may of course be used. The minimum distortion detection circuit 14 emits a strobe signal 16 when the quantum device distortion is minimum, and the code table address 8
is taken into the index latch 17, and the minimum quantization distortion 15 min d _i is output. When the code table address counter 7 completes one rotation, the code table address stored in the index latch 17 becomes the vector quantization index 18. The amplitude correction circuit 19 corrects the amplitude component 4 using the minimum quantization distortion 15 and outputs a corrected amplitude value 20. Amplitude correction is performed to minimize distortion during signal reproduction, and corresponds to applying a coefficient to the amplitude so that, for example, when quantization distortion is large, the amplitude is reduced in reproduction.
This not only makes it possible to reduce distortion due to quantizer mismatch, but also has the effect of reducing noise components contained in the input signal. The amplitude correction circuit 19 has the minimum quantization distortion 15 and the amplitude component 4.
It can be configured with a ROM that takes address input as address input, and the increase in device scale is extremely small.
The correction characteristics can be determined in advance by statistically calculating the relationship between the quantization distortion min d _i and a coefficient that minimizes the distortion during reproduction by multiplying the amplitude.
Further, when standard deviation is used for normalization and square distortion is used as the definition of distortion, the correction characteristics can also be determined using a mathematical formula. FIG. 3 is an explanatory diagram showing the principle of correction for an embodiment of the present invention using standard deviation and square distortion. In the figure, X is a normalized input vector, and y _i is a normalized output vector obtained by vector quantizing X. Since both X and y _i are average-separated and normalized, they lie on a supercircle with radius √ in the K-dimensional space. The angle θ formed by _the normalized _input vector As is clear from the figure, cos θ is the optimal coefficient to obtain the best approximation of X by multiplying y _i by a coefficient. The above relationship is preserved in a similar form even if the amplitude is multiplied during playback, so the amplitude is expressed as a coefficient.
By multiplying by cos θ, distortion during reproduction can be reduced corresponding to the quantization distortion d _i . Since d _i is calculated at the time of vector quantization, the coefficient cos θ is obtained as cos θ=1−d _i /2K (8). Therefore, the corrected amplitude value 20σ' is expressed as σ'=σ(1-d _i /2K) (9). The encoder 21 uses the corrected amplitude value 20σ', the average value component 5m, and the vector quantization index 18.
is encoded and output as the encoder output signal 22. In the decoding section, the decoder 23 decodes the vector quantization index 18, the average value component 5, and the corrected amplitude value 20, reads out the normalized output vector y _i from the normalized output vector code table 9, and normalizes it. Latch into output vector register 10. Furthermore, the average value separation normalization restoration circuit 25 uses the average value component 5 and the corrected amplitude value 20 to generate an output vector 26S'={s ₁ ', s ₂ ', . . . , s _K '}
decrypt. That is, s _j ′=σ・y _ij +m ...(10) (j=1, 2, ..., K) The operation and configuration of the decoding section are the same as that of the conventional vector quantizer decoding section. . [Effects of the Invention] As described above, according to the present invention, even if mismatching occurs in the quantizer, the amplitude during reproduction is corrected according to the quantization distortion. This has the effect of reducing quantization noise of quantization and noise included in the input signal.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an example of the configuration of the encoding section of an adaptive vector quantizer according to an embodiment of the present invention, and FIG. 2 is a decoding section according to an embodiment of the adaptive vector quantizer according to the present invention. 3 is an explanatory diagram illustrating the principle of correction in an embodiment of the present invention. FIG. 4 is a block diagram illustrating an example configuration of an encoding section of a conventional vector quantizer.
FIG. 5 is a block diagram showing an example of the configuration of a decoding section of a conventional vector quantizer. In the figure, 1 is an input vector, 2 is an average value separation normalization circuit, 4 is an amplitude component, 5 is an average value component, 6 is a regular input vector register, 7 is a code table address counter, and 9 is a normalized output vector code table. Memory, 10 is a normalized output vector register, 11 is a parallel subtracter, 12 is a parallel absolute value calculator, 13 is an absolute value distortion calculation circuit, 14 is a minimum distortion normalized output vector detection circuit, 15 is a minimum quantization distortion , 17 is an index latch, 19 is an amplitude correction circuit, 20 is a corrected amplitude value, 21 is an encoder, 22 is an encoder output signal, 23 is a decoder,
25 is a mean value separation, normalization and restoration circuit, and 26 is an output vector.

Claims (1)

[Claims] 1. A mean value separation and normalization circuit that obtains a normalized input vector by separating the mean value component from an input vector obtained by blocking the input signal every K samples and normalizing the amplitude; a normalized output vector code table memory that stores a set of average value-separated normalized output vectors obtained in advance so as to cause minimum distortion for the normalized input vector; and a memory for the normalized vector and the normalized output vector code table. an absolute value distortion calculation circuit that calculates the distortion between the average value and the normalized output vector read out sequentially from the normalized output vector; A distortion normalized output vector detection circuit performs amplitude reproduction of the normalized output vector giving the minimum distortion, and performs the above processing with reference to the minimum distortion so that the distortion between the output vector added with an average value and the input vector is minimized. an amplitude correction circuit that corrects the normalization constant calculated by the average value separation normalization circuit; and a code that encodes the average value component of the input vector, the corrected amplitude value, and a vector quantization index that provides the minimum distortion. a decoder that decodes the output of the encoding circuit to obtain an index of a normalized output vector that provides minimum distortion, the average value component, and the corrected amplitude value, and the normalized output vector code. For the average-separated normalized output vector read out from the table memory according to the decoded index,
An adaptive vector quantizer comprising a mean value separation, normalization and restoration circuit that performs amplitude reproduction and mean value addition using the decoded mean value component and the corrected amplitude value to obtain a reproduced output vector.