JP3010638B2 - Signal processing device and signal processing method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a signal processing device and a signal processing method for compressing and quantizing input data.

[Summary of the Invention]

The present invention performs variable-length encoding on a vector-quantized output obtained by vector-quantizing input data, monitors the vector-quantized output for each predetermined unit of the input data, and performs vector quantization according to the bit amount in the output. The present invention provides a signal processing device and a signal processing method capable of efficiently quantizing input data by controlling the parameters (1) and (2).

[Conventional technology]

For example, as a method for quantizing an audio signal or the like, a vector (a series of sample words) divided into an appropriate size is collected into one block, and the one block is quantized as one vector. There is a quantization method. In this vector quantization, an index corresponding to a code vector most similar to the input vector (for example, closest to the distance) is obtained, so that the number of bits can be reduced. Further, since the correlation between each word of the input data can be removed, the quantization efficiency is high.

[Problems to be solved by the invention]

By the way, in the above-described vector quantization, the occurrence probability of the index is not constant and the probability distribution is biased, which is a disadvantage in improving the efficiency of the quantization. In this case, the uneven distribution of the index occurrence probability can be reduced by increasing the dimension of the vector (that is, increasing the number of words in the divided block). But,
Increasing the dimension of the vector increases the amount of calculation and complicates the configuration. Therefore, in practice, the dimension of the input vector cannot be sufficiently increased.

Therefore, the present invention has been proposed in view of the above-described situation, and can perform highly efficient quantization even when the occurrence probability of an index is biased due to low-dimensional vector quantization. It is an object of the present invention to provide a signal processing device and a signal processing method capable of keeping the bit amount of a quantized output within a predetermined amount.

[Means for solving the problem]

A signal processing apparatus according to the present invention has been proposed to achieve the above-described object. Long encoding means and control means for monitoring the output of the vector quantization means for each predetermined unit of the input data and controlling a parameter for quantization of the vector quantization means according to the bit amount at the output. It is provided.

The signal processing method of the present invention has been proposed in order to achieve the above-mentioned object, and performs vector quantization of input data, variable-length encoding of vector quantization output, and vector quantization output of input data. Monitor every predetermined unit,
The parameter for vector quantization is controlled according to the bit amount in the output.

Here, as the monitoring of the vector quantization output, for example, an output in which n input vectors (n blocks) are made into one large block is stored in, for example, an output buffer.
It is monitored whether or not the number of bits per large block is within a predetermined bit number range.

[Action]

According to the present invention, since the output (index, identification code) of the vector unit of the vector quantization means is variable-length coded, the bias of the probability of occurrence of the index from the vector quantization means is eliminated to improve the coding efficiency. In addition, by controlling a parameter for vector quantization in accordance with the bit amount of each predetermined unit, it is possible to prevent occurrence of overflow in variable-length encoding.

〔Example〕

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 shows a schematic configuration example of the vector quantizer of the present embodiment.

In FIG. 1, input data such as an audio signal via an input terminal 1 is converted into an input vector having a plurality of words (a plurality of samples) as one block, and is vector-quantized by a vector quantizer 11. That is, the vector quantizer 11 has a code book in a memory in which a code vector (representative vector) associated with an index (identification code) is recorded, and is stored in the input vector and the code book. The code vector that is present is compared, for example, by distance calculation or the like, the code vector that is closest (similar) is found, and the index corresponding to the code vector is read. As described above, since the vector of the input data is vector-quantized, the number of bits can be reduced, and the correlation between each word of the input data is removed so that the quantization can be performed efficiently.

Further, in the present embodiment, the vector quantizer 11
Is supplied to the entropy encoder 12, which is the variable-length encoding means, and so-called variable-length encoding is performed, whereby a certain code vector is used more frequently than other code vectors. The resulting bias in the occurrence probability of the index is eliminated, and the quantization efficiency can be further increased.

By the way, in the case of entropy coding as described above,
As described later, there is a possibility that data to be encoded may overflow. Therefore, in this embodiment,
Monitoring the output of the vector quantizer 11 stored in the output buffer 12a in the entropy encoder 12, and controlling the resolution or the quantization step size of the vector quantizer 12 according to the output. And eliminates the overflow. That is, the output of the vector quantizer 11 monitored by the entropy encoder 12 is, for example, n input vectors (n blocks), as described later.
The output is one large block. By storing the large block in the output buffer 12a, it is monitored whether the number of bits per large block is within a predetermined bit number range. At this time, when overflow occurs, that is, when the number of bits exceeds a predetermined number,
The resolution or the quantization step size of the vector quantizer 11 is controlled. The data encoded in this manner is output from the output terminal 2.

Here, the cause of occurrence of the above-mentioned overflow and the control of the resolution or the quantization step size as a countermeasure thereof will be described below together with the description of entropy coding.

In the variable-length encoding performed by the entropy encoder 12, generally, when there is a bias in the occurrence probability of a code such as an amplitude distribution such as an audio signal, that is, vector quantization is performed as in the present embodiment. When there is a bias in the probability of occurrence of the index from the unit 11, encoding is performed by changing each bit length according to the frequency of appearance of the quantization result, thereby reducing the number of bits of the signal. That is, if a short code is used for a code with a high appearance frequency and a long code is used for a code with a low frequency, the average code length can be shortened. At this time, statistically, the bit length is preferably determined according to the entropy, and this code is called an entropy code. A specific algorithm of the entropy code is a so-called Shannon-
There are Fano codes, Huffman codes, and the like. These entropy codes have a large effect when combined with uniform quantization, and the highest S / N can be obtained. By the way, in the present embodiment, as the vector unit output of the vector quantizer 11, for example, n input vectors (n blocks) are one large block, and the transmission rate for each large block is constant. Here, the output from the entropy encoder 12 has a variable rate because it has an unequal length, but if a code error occurs at this time, the transmission rate for each large block is made constant. If there is no synchronization between the large blocks, error propagation may occur forever.
Therefore, reset is performed for each large block. However, if the entropy of a signal of a certain large block is large, there is a possibility that an overflow occurs in the large block.
That is, the index cannot be encoded at a predetermined rate.

For this reason, in the apparatus according to the present invention, as described above, the output buffer in the entropy encoder 12 is used.
A feedback loop is formed between 12a and the vector quantizer 11, and the output buffer 12a in the entropy encoder 12 monitors a large block of vector unit output of the vector quantizer 11 and outputs the result. According to the above vector quantizer 11
The overflow is prevented by controlling the resolution or the quantization step size. That is, if an overflow occurs in the output (large block) of the vector quantizer 11, the resolution (bit / vector) of the vector quantizer 11 in the large block is reduced. Alternatively, increase the quantization step size. At this time, the resolution of the vector quantizer 11 or the step size of the quantization is fixed in one large block, and the amount of change in the resolution or the step size is:
By making it proportional to the amount of overflow, it is possible to quickly converge. When underflow occurs, the resolution can be increased or the step size can be reduced.

Next, a structure for making the resolution of the vector quantizer 11 variable will be described. In this case, the code book of the vector quantizer 11 has, for example, a binary tree structure, that is, a tree structure, so that the resolution can be easily changed.

In the vector quantization using the tree structure codebook, the index of the code vector is obtained by storing the index of the codebook in a binary tree structure in advance and tracing the binary tree.
For example, if the size of the codebook of the binary tree is N = 2 ^kr (N,
When k and r are integers of 0 or more), log ₂ N = kr binary decisions,
That is, the N-dimensional space is sequentially divided, and the input vector is compared with the two code vectors only at the final decision stage. Here, for example, in the case of a four-level binary tree having nodes t, t ₀ , to t ₁₁₁ at each stage as shown in FIG. 2, the input vector is, for example, a node at the first stage of the binary tree. At t, it is compared with the code vectors of the nodes t ₀ and t ₁ in the second stage (for example, distance calculation). At this time, a node having a code vector most similar to the input vector (a code vector with a short distance) is selected. In the example of FIG. 2, the index corresponding to the code vector in the second stage is represented by 1 bit / vector. At this time, for example, if the code vector of the contact point t ₀ is selected,
Next, at the third-stage nodes t ₀₀ and t ₀₁ , the input vector is compared with the code vectors of these third-stage nodes t ₀₀ and t ₀₁ . In the example of FIG. 2, the index corresponding to the third-stage code vector is represented by 2 bits / vector. Similarly, for example, when a node t ₀₁ was chosen, this time node t _010, t code vector and the input vector ₀₁₁ of the fourth stage of the last stage are compared, (close in distance) the most similar code vector Is selected, and its index is obtained. In the example of FIG. 2, the index corresponding to the last-stage code vector is 3 bits /
It is represented by a vector.

That is, at the final stage of the binary tree coding, an index of 3 bits (3 bits / vector) is obtained for the input vector, and this index is entropy-coded and the number of output bits for each block is variable. Become.

As for the output of such entropy coding, as described above, it is necessary that the number of bits of a large number of blocks, for example, n blocks, be constant. It is assumed that an overflow occurs in the large block exceeding the number of bits. In such a case, the least significant bit of the index is cut into 2 bits / vector, and the index represented by the 2 bits / vector is entropy-encoded. At this time, even if the least significant bit of the index is cut to 2 bits / vector as described above, since the codebook is defined in a tree structure, the similarity between the centroid (representative point) and the index is similar. Matched. The mapping of the entropy encoder may or may not be changed depending on the data length of the index (2 bits / vector or 3 bits / vector, etc.). For example, when changing the mapping of the entropy encoder, it is necessary to design and prepare the entropy encoder according to each pit rate (pit / vector). When the entropy encoder is not changed, for example, the least significant bit or the most significant bit of the data rounded to 2 bits / vector is set to 0, and the data is made to appear as 3 bits / vector data, and entropy encoding is performed. . In this case, it is necessary to shift to a data side with a higher probability of occurrence by rounding the index from 3 bits / vector to 2 bits / vector. That is, otherwise, the result of entropy encoding will not be shortened. That is, for example, when the 3-bit index “001” is rounded to 2 bits to be “00”, the code vector of the index “000” is associated with the 2-bit index “00”. The occurrence probability of “000” is required to be higher than the index “001”.

Further, an example of a structure for making the quantization step size of the vector quantizer 11 variable will be described. That is, the configuration with the variable step size can be realized as shown in FIG. FIG. 3 shows that the gain variable means 10 is provided in front of the vector quantizer 11 of FIG. 1 described above, and the output from the output buffer 12a in the entropy encoder 12 is sent to the gain variable means 10. The feedback is used to control the step size. Since other configurations and operations are the same as those in FIG. 1, parts corresponding to those in FIG.

In FIG. 3, an overflow is prevented by changing the step size while keeping the resolution (bit / vector) of the vector quantizer 11 constant.
Here, when an overflow occurs, the gain varying means 10 is controlled in a direction in which the step size increases, that is, in a direction in which the gain decreases. By reducing the gain as described above, a code having a higher occurrence probability, that is, a code having a shorter code length is assigned by the entropy encoder 12, and a code vector having a smaller norm is frequently selected. On the other hand, when an underflow occurs, the gain varying means 10 is controlled in the direction in which the step size is small, that is, the gain is large. As described above, when the gain is increased, a code having a low occurrence probability, that is, a code having a long code length is assigned. The index indicating the magnitude of the gain is transmitted separately.

As described above, Huffman coding or the like can be used for entropy coding.

The Huffman coding will be briefly described. For example, assuming that there are eight levels and their appearance probabilities are known, first, level numbers are assigned in descending order of probability, and encoding is performed in a total of seven steps. That is, numbers (0, 1) are assigned to the two levels with the lowest probabilities, the probabilities of the two are summed, the summed probabilities are regarded as one level, and they are rearranged again in descending order of the probabilities. The coding can be performed by following the line connecting each level in reverse) and picking up the number (0,1).

If such Huffman coding is applied to the above-described entropy encoder 12, the same effect as described above can be obtained.

〔The invention's effect〕

In the present invention, the output of the vector quantization means is changed to a variable length coding means and is subjected to variable length coding, and the output of the vector quantization means in variable units is subjected to variable length coding. The uneven distribution of the index occurrence probability distribution can be eliminated, and therefore, the input data can be efficiently quantized. In addition, since the vector quantization output is monitored for each predetermined unit of input data, and a parameter for vector quantization is controlled according to the bit amount in the output, an overflow occurs during variable-length encoding. There is no danger.

[Brief description of the drawings]

FIG. 1 is a block circuit diagram showing a schematic configuration of an apparatus according to an embodiment of the present invention, FIG. 2 is a conceptual diagram for explaining binary tree coding,
FIG. 3 is a block circuit diagram showing a schematic configuration of the apparatus of the embodiment when the step size is variable. 10 Gain variable means 11 Vector quantizer 12 Entropy encoder 12a Output buffer

Continued on the front page (72) Inventor Yoshihito Fujiwara 6-7-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Inside Sony Corporation (72) Inventor Makoto Akune 6-35-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Stock In-house (56) References JP-A-64-29121 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) H03M 7/30

Claims (2)

(57) [Claims] 1. Vector quantization means for vector-quantizing input data, variable-length coding means for performing variable-length coding on the output from the vector quantization means, and output from the vector quantization means as input data. And a control unit for monitoring a parameter for quantization of the vector quantization unit according to a bit amount in an output of the signal processing unit. 2. Vector quantization of input data, variable length coding of the vector quantization output, monitoring of the vector quantization output for each predetermined unit of the input data, and A signal processing method comprising controlling parameters for vector quantization.