JP2005181354A - Device and method for decoding - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To shorten the time needed for decoding processing by reducing the calculation loads of inverse quantization processing and inverse normalization processing after the decoding processing of an entropy-encoded code series. <P>SOLUTION: A decoding section 12 cuts variable-length codes out of the code series supplied from a shift register 20 by a maximum word length from the position that a counter 21 indicates, and an index corresponding to the variable-length codes is supplied from an index ROM 22 to a spectrum coefficient ROM 24 by using its output value. The spectrum coefficient ROM 24 stores conversion tables showing correspondence relation between quantization spectrum coefficients x_quant and spectrum coefficients spec corresponding to indexes by values of scale factors sf. The value at an index position is read out of a conversion table selected based upon the value of a scale factor sf supplied from a subordinate information decoding part 11 and outputted as a spectrum coefficient spec. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a decoding apparatus for decoding a code string obtained by encoding by assigning any of a plurality of variable length codes to each symbol that has been subjected to normalization processing and quantization processing on a plurality of coefficients. And a method.

  In data compression such as video and audio based on a perceptual information model, high-efficiency compression is often realized by a two-stage process in which data compression is first performed by information modeling and then entropy coding is performed. Among these, in the first stage process, for example, a higher compression ratio can be realized by allowing a non-perceptible loss. Further, in the second stage processing, a higher compression ratio can be realized by assigning a short code to a symbol having a high appearance frequency and shortening the average code length of the encoding result. As a technique for performing such two-stage data compression, standard techniques such as MPEG (Moving Picture Experts Group), which is a video / audio compression standard standardized by ISO, exist and are widely used.

  Here, since human perception has a characteristic of exhibiting non-linear characteristics, normalization using a logarithm or non-linear quantization is often performed as the first stage processing. At this time, as a logarithmic value used for normalization, a coefficient that is a power of 2 is often used for the purpose of efficient implementation by a computer using a binary number.

  When decoding an encoded result that has been data-compressed in a two-stage process, the process is performed in the reverse direction at the time of compression. That is, the entropy code is first decoded, and then the data depending on the modeling is decoded. That is, after decoding the entropy code, some processing is performed on the decoding result to obtain a final output result.

  By the way, various methods have been proposed so far for efficiently decoding an entropy-coded code string and obtaining a symbol value (see, for example, Patent Document 1).

Japanese Patent No. 3229690

  Here, when the two-stage data compression is performed as described above, since the symbol value itself is not the final decoding result, a process corresponding to the first-stage process is required thereafter. However, when this processing cost is high, there is a problem that sufficient performance cannot be obtained as a whole even if the decoding process of the entropy-coded code string is sufficiently efficient.

  As an example of processing corresponding to the first stage processing, an extract of decoding processing in the MPEG1 Layer 3 (hereinafter referred to as MP3) standard as the above-described MPEG audio compression is shown below. As for the detailed contents of the standard, the ISO / MPEG standard text “Encoding of movies and audio for digital storage media of about 1.5 Mbit / s-CD11172-3 (Audio Part 3)” (ISO / IECJTCI / SC29 , 1992).

  In MP3, a scale factor sf is calculated from information such as a global gain included in an area in which information unique to each frame, which is called sub-information, is called a sub-information, and the following equation (1) is calculated from the scale factor sf. According to the above, gain gain is calculated.

  Similarly, a quantized spectral coefficient x_quant is obtained from entropy-encoded frequency domain data included in the code string, and normalized by inverse nonlinear quantization as shown in the following equation (2) from the quantized spectral coefficient x_quant. Obtain the normalized spectral coefficient x_invquant.

  Then, using both of them, a spectral coefficient spec is calculated according to the following equation (3), and a filter bank process or the like is further performed on the spectral coefficient spec to obtain a final decoding result (speech waveform or the like). Yes.

  At this time, the gain gain and the value of the scale factor sf that is the basis of the gain gain are shared by a plurality of normalized spectral coefficients x_invquant. Note that frequency bands sharing the same scale factor sf value are referred to as scale factor bands.

  As described above, in MP3, in addition to decoding the quantized spectral coefficient x_quant which is an entropy-coded symbol, the gain gain shown in Expression (1) is calculated for each scale factor band, and then Expression (2) It can be seen that the nonlinear quantization decoding shown in (1) and the multiplication shown in Equation (3) need to be performed.

  Therefore, when the processing cost of the processes shown in the equations (1) to (3) is high, even if the decoding process of the entropy-encoded code string is sufficiently efficient, sufficient performance as a whole can be obtained. There will be no.

  The present invention has been proposed in view of such a conventional situation, and reduces the calculation load in the inverse quantization process and the inverse normalization process after the decoding process of the entropy-encoded code string. An object of the present invention is to provide a decoding apparatus and method capable of reducing the time required for the above.

  In order to achieve the above-described object, the decoding apparatus according to the present invention assigns one of a plurality of variable length codes to each symbol obtained by performing normalization processing and quantization processing on a plurality of coefficients. In the decoding apparatus that decodes the code string obtained by encoding, the plurality of variable length codes and values obtained by performing inverse quantization processing and inverse normalization processing on symbols corresponding to the variable length codes in advance. A code string portion that matches any of the plurality of variable length codes in the conversion table is searched from the storage means storing the associated conversion table and the code string, and matches the searched code string portion. Decoding means for restoring a coefficient corresponding to the searched code string portion by reading a value associated with the variable-length code is provided.

  In order to achieve the above-described object, the decoding method according to the present invention assigns one of a plurality of variable length codes to each symbol obtained by performing normalization processing and quantization processing on a plurality of coefficients. In the decoding method for decoding a code string obtained by encoding, a value obtained by performing inverse quantization processing and inverse normalization processing on the plurality of variable length codes and symbols corresponding to the variable length codes in advance. Is searched for a code string portion that matches any of the plurality of variable length codes in the conversion table from the code string, and a variable length that matches the searched code string portion. It has a decoding step of restoring the coefficient corresponding to the searched code string portion by reading the value associated with the code.

  In such a decoding apparatus and its method, instead of performing the inverse quantization process and the inverse normalization process on the symbol obtained by decoding the code string, the inverse quantization process and the symbol corresponding to each variable length code are performed in advance. A value associated with a variable-length code that matches a code string portion cut out from a code string is read using a conversion table associated with a value subjected to denormalization processing.

  According to the decoding apparatus and method of the present invention, instead of performing inverse quantization processing and inverse normalization processing on symbols obtained by decoding a code string, inverse quantization is performed in advance on symbols corresponding to each variable length code. In order to read out the value associated with the variable length code that matches the code string portion cut out from the code string using the conversion table in which the values subjected to the normalization process and the denormalization process are associated with each other, the code string It is possible to reduce the calculation load in the inverse quantization process and the inverse normalization process after the decoding process, and to shorten the time required for the decoding process. Further, as the calculation load is reduced, the circuit scale necessary for the calculation can be reduced, so that the cost of parts can be reduced.

  Hereinafter, specific embodiments to which the present invention is applied will be described in detail with reference to the drawings. In this embodiment, the present invention is applied to a decoder (audio decoder) in MP3 which is one of high-efficiency speech coding techniques. Of course, the present invention can be applied not only to audio data compression but also to video data compression.

  This audio decoder decodes a code string in which a variable length code is assigned to a value (quantized spectrum coefficient x_quant) obtained by normalizing and nonlinearly quantizing frequency information, and for the decoded quantized spectrum coefficient x_quant. Then, the processing corresponding to the above-described equations (1) to (3) is performed to decode the spectral coefficient spec.

  According to the MP3 standard, about 360 types of scale factors sf and about 8200 types of quantized spectral coefficients x_quant can be used. However, for simplicity, eight values are used as scale factors sf below. Assume that only four types of values (0, 1, 2, 3) can be selected as (0, 1,..., 7) and the quantized spectral coefficient x_quant. In addition, as shown in Table 1 below, it is assumed that 1-3 bit variable length codes are assigned to each quantized spectral coefficient x_quant.

  FIG. 1 shows a schematic configuration of the audio decoder in the present embodiment. In the audio decoder 1 shown in FIG. 1, the signal separation unit 10 divides the input compressed data (MP3 bit stream) in units of frames, and the frame information is divided into a code string including the quantized spectral coefficient x_quant and other sub-codes. Separated into information. The signal separation unit 10 supplies the separated sub information to the sub information decoding unit 11, and also supplies the code sequence including the quantized spectral coefficient x_quant to the decoding unit 12.

  The sub information decoding unit 11 decodes the value of the scale factor sf in the sub information and supplies the scale factor sf to the decoding unit 12. The actual sub-information includes values other than the scale factor sf, but are omitted here for simplicity.

  The decoding unit 12 decodes the quantized spectral coefficient x_quant from the code string including the quantized spectral coefficient x_quant, and performs the inverse nonlinear quantization process and the inverse normalization process on the quantized spectral coefficient x_quant to thereby obtain the spectral coefficient. Restore spec.

  The orthogonal transform unit 13 transforms the signal on the frequency axis constituted by the restored spectral coefficient spec into a signal on the time axis by orthogonal transform, and the time signal processing unit 14 performs the processing on the signal on the time axis. Conversion based on the sub information is performed, and decoded data (PCM data) is output.

  FIG. 2 shows an internal configuration of the decoding unit 12 in the audio decoder 1. In FIG. 2, the signal separation unit 10 and the sub information decoding unit 11 are also shown. As shown in FIG. 2, the decoding unit 12 stores a shift register 20 that performs a shift operation every time a code string is input, and a location where decoding is to be performed next out of signals supplied from the outside, and from there externally A counter 21 for controlling the shift register 20 so as to cut out the number of bits corresponding to a given maximum word length, an index ROM 22 for outputting an index described later from the value of the shift register 20, and a code length from the index Code length ROM 23, and a spectral coefficient ROM 24 for obtaining a spectral coefficient spec from the index and scale factor sf and supplying the spectral coefficient spec to the orthogonal transform unit 13 (FIG. 1).

  Hereinafter, the operation of each block until the spectral coefficient spec is output in the decoding unit 12 will be described.

  First, the shift register 20 has the number of bits (maximum word length) of the code string supplied from the signal separation unit 10 having the maximum length as the variable length code from the position indicated by the counter 21, specifically, Cuts out 3 bits. For simplicity, it is assumed that the maximum word length is always 3 in the present embodiment, but when this value is variable, information of the maximum word length is given from, for example, the sub information decoding unit 11 Is also possible.

  Next, the index ROM 22 is accessed using the output value of the shift register 20. The contents of the index ROM 22 are shown in Table 2 below.

  As shown in Table 2, for each of the four types of variable length codes, a code string of 3 bits (maximum code length) including the variable length code as it is and the first bit of the variable length code as the most significant bit is taken. The same index value is assigned to all the obtained values. Thereby, an index corresponding to the variable length code can be acquired from the shift register 20. Here, for convenience, the same value as the quantized spectral coefficient x_quant is used as the index value, but this value can be freely selected within a range where consistency with the spectral coefficient ROM 24 can be obtained.

  Subsequently, the code length ROM 23 is accessed using the output of the index ROM 22. The contents of the code length ROM 23 are shown in Table 3 below.

  As shown in Table 3, the actual code length is associated with the index values corresponding to the four types of variable length codes. By increasing the value of the counter 21 using this code length, it becomes possible to correctly match the read start position of the next code string with the start position of the next variable length code.

  The spectrum coefficient ROM 24 is composed of eight conversion tables created for each value of the scale factor sf. Tables 4 to 7 below show conversion tables when the values of the scale factor sf are 0, 1, 2, and 3.

  As can be seen from Tables 4 to 7, each conversion table shows the correspondence between the quantized spectral coefficient x_quant corresponding to the index value and the spectral coefficient spec when the scale factor sf is a certain value. Has been. Of the eight conversion tables, the one actually used is selected using the value of the scale factor sf supplied from the sub information decoding unit 11. Then, the value recorded at the index position is read from the selected table, and the value is output as the spectrum coefficient spec.

  When the selection and access of the spectral coefficient ROM 24 is implemented by software, it may be performed by accessing a two-dimensional array in which the first dimension is the scale factor sf and the second dimension is the output value of the index ROM 22.

  The decoding unit 12 assigns a variable-length code to a value (quantized spectrum coefficient x_quant) obtained by normalizing and nonlinearly quantizing the frequency information by repeating the above operation until the input from the signal separating unit 10 is completed. The spectral coefficient spec can be obtained by decoding the obtained code string.

  In the actual high-efficiency speech coding technique, as described above, since more types of values can be selected as the value of the scale factor sf, the size of the spectrum coefficient ROM 24 becomes very large.

  Therefore, in practice, as described below, the size of the spectrum coefficient ROM 24 can be suppressed to some extent by using the periodicity of the normalization coefficient.

  As described above, since a value that is a power of 2 is normally used as the normalization coefficient, the gain gain value changes with periodicity when viewed in the binary number domain.

  That is, the relationship between the gain gain and the scale factor sf is as shown in Equation (3) described above. At this time, when the remainder obtained by dividing the scale factor sf by 4 is expressed as sf% 4, the following equation (4) is obtained. Can be deformed.

  Here, since (sf−sf% 4) in the expression (4) is always a multiple of 4, 2 ^ ((sf−sf% 4) / 4) is always an integer power of 2. In other words, multiplying this value is equivalent to multiplying the value of 2 to the power of an integer, and if it is a computer using a binary number, it is a calculation that can be easily realized only by operation of the exponent part in the shift or floating point expression. is there.

  Therefore, in FIG. 2, the value of the scale factor sf that is the decoding result of the sub information is not used for selecting the spectral coefficient ROM 24, but is selected using sf% 4, and the value read from the spectral coefficient ROM 24 is used. If the process of multiplying the output by 2 ^ ((sf-sf% 4) / 4) using the shift operation or the like is performed without shifting the output as it is, the calculation amount of the spectrum coefficient ROM 24 is not significantly increased. The size can be kept small.

  As described above, according to the audio decoder 1 of the present embodiment, when the spectral coefficient spec is obtained, multiplication and exponent calculation with the largest load can be substantially eliminated, so that the calculation load during execution is greatly increased. The time required for the decoding process can be shortened. Specifically, in the software implementation by the present inventors, the processing time of the part for obtaining the spectral coefficient spec from the quantized spectral coefficient x_quant is shortened to about 時間 time compared to the conventional method. I was able to.

  In addition, since the circuit scale required for calculation can be reduced as the calculation load is reduced, for example, in an embedded device that requires execution with a limited memory size, decoding processing is performed with a smaller memory size. This makes it possible to reduce the cost of parts.

  It should be noted that the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the scope of the present invention.

  For example, in the above-described embodiment, an example is described in which the present invention is applied to an audio decoder that uses parameters only during denormalization processing, such as MP3, but is not limited to MP3. The present invention can also be applied to other audio decoders that use parameters in the denormalization process.

  As an example, an audio decoder that performs the following decoding process is assumed. That is, first, the gain gain is calculated from the scale factor sf according to the following equation (5).

  Also, a quantized spectral coefficient x_quant is obtained from entropy-encoded frequency domain data, and a normalized spectral coefficient x_invquant is obtained from the quantized spectral coefficient x_quant by inverse nonlinear quantization as shown in the following equation (6). Here, max_quant in equation (6) is a value obtained from the quantization step for each quantization spectral coefficient x_quant.

  Then, using both of them, a spectrum coefficient spec is calculated according to the following equation (7), and a filter bank process or the like is further performed on the spectrum coefficient spec to obtain a final decoding result (speech waveform or the like).

  Therefore, in an audio decoder in which two parameters such as the scale factor sf and max_quant exist in this way, these two conversion tables indicating the correspondence between the quantized spectral coefficient x_quant corresponding to the index value and the spectral coefficient spec are displayed. What is necessary is just to prepare in the spectrum coefficient ROM 24 for every one parameter.

  Further, as in the above-described embodiment, the periodicity of the normalization coefficient is used, and the value of the scale factor sf that is the decoding result of the sub information in FIG. 2 is not used for selection of the spectrum coefficient ROM 24. The value read out from the spectral coefficient ROM 24 is not output as it is, but the process of multiplying the output by 2 ^ ((sf-sf% 3) / 3) is performed as a shift operation. If used, the size of the spectrum coefficient ROM 24 can be kept small without significantly increasing the amount of calculation.

It is a figure which shows schematic structure of the audio decoder in this Embodiment. It is a figure which shows the internal structure in the decoding part of the audio decoder.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 Audio decoder, 10 Signal separation part, 11 Sub information decoding part, 12 Decoding part, 13 Orthogonal transformation part, 14 Time signal processing part, 20 Shift register, 21 Counter, 22 Index ROM, 23 Code length ROM, 24 Spectral coefficient ROM

Claims (8)

In a decoding device that decodes a code string obtained by encoding by assigning any of a plurality of variable length codes to each symbol subjected to normalization processing and quantization processing on a plurality of coefficients,
Storage means for storing a conversion table in which the plurality of variable length codes and the values corresponding to the symbols corresponding to the variable length codes are previously subjected to inverse quantization processing and inverse normalization processing;
From the code string, searching for a code string portion that matches any of the plurality of variable length codes in the conversion table, and reading a value associated with the variable length code that matches the searched code string part, A decoding device comprising: decoding means for restoring a coefficient corresponding to the searched code string portion.   The decoding apparatus according to claim 1, wherein the conversion table is created for each parameter used for the denormalization process, or the dequantization process and the denormalization process. The normalization coefficient used in the denormalization process is obtained according to the following formula 1, where the normalization coefficient is gain and the scale factor used in the denormalization process is sf:

The decoding apparatus according to claim 1, wherein the conversion table is created for each value of n according to at least the periodicity of the normalization coefficient. In the conversion table, a plurality of variable length codes and a remainder value obtained by dividing the scale factor by the value of n are regarded as the scale factor, and a symbol corresponding to each variable length code is dequantized in advance. And a value subjected to denormalization processing are associated with each other,
The decoding means reads a value associated with the variable-length code that matches the searched code string portion, sets the scale factor to sf, and divides the scale factor by the value of n as sf% n. 4. The coefficient corresponding to the searched code string portion is restored by multiplying the read value by 2 ^ ((sf−sf% n) / n). Decoding device. In a decoding method for decoding a code string obtained by encoding by assigning any of a plurality of variable length codes to each symbol subjected to normalization processing and quantization processing on a plurality of coefficients,
From the code string, using the conversion table in which the plurality of variable-length codes and the values corresponding to the symbols corresponding to the variable-length codes are previously subjected to inverse quantization processing and inverse normalization processing, The searched code string is searched by searching for a code string portion that matches any of the plurality of variable length codes in the conversion table and reading a value associated with the variable length code that matches the searched code string portion. A decoding method comprising: a decoding step of restoring a coefficient corresponding to the portion.   6. The decoding method according to claim 5, wherein the conversion table is created for each parameter used for the denormalization process or the dequantization process and the denormalization process. The normalization coefficient used in the denormalization process is obtained according to the following formula 2 when the normalization coefficient is gain and the scale factor used in the denormalization process is sf.

The decoding method according to claim 5, wherein the conversion table is created for each value of n according to at least the periodicity of the normalization coefficient. In the conversion table, a plurality of variable length codes and a remainder value obtained by dividing the scale factor by the value of n are regarded as the scale factor, and a symbol corresponding to each variable length code is dequantized in advance. And a value subjected to denormalization processing are associated with each other,
The decoding means reads a value associated with the variable length code that matches the searched code string part, sets the scale factor to sf, and divides the scale factor by the value of n as sf% n. The coefficient corresponding to the searched code string portion is restored by multiplying the read value by 2 ^ ((sf-sf% n) / n). Decryption method.

Images