CN102760442B - 3D video azimuth parametric quantification method - Google Patents

Abstract

The present invention proposes a method for quantifying horizontal orientation parameters in 3D audio, which considers the perception ability of the human ear in different frequency bands, and introduces the minimum difference JND value that the human ear can perceive; on the basis of quantifying objective distortion, it is compared with the JND value The subjective perceptual distortion is obtained by combining operations, and based on this, the codeword that is most in line with the perceptual characteristics of the human ear is selected as the quantization result. Quantification is performed using this technical solution, and the subjective perception performance of the quantification result is better.

Description

A Quantization Method of Horizontal Orientation Parameters in 3D Audio

technical field

The present invention relates to the technical field of quantization and coding, and more specifically, to a method for quantizing horizontal orientation parameters in 3D audio.

Background technique

Quantization is the main step in compression coding, which achieves data reduction by representing a larger data set with a smaller data set. Smaller data sets are often called codebooks, and the data in the codebooks are called codewords. Larger datasets are collections of values to quantify. After the data is quantized, only the position value of the output codeword in the codebook is used. This positional value is often called an index. There are many types of commonly used quantization techniques, which can be roughly divided into two categories: scalar quantization and vector quantization. Scalar quantization is to quantify the data to be quantized one by one, and the complexity is low; vector quantization is to regard several data as a vector, and perform overall quantization on each vector, preserving the correlation between data. Among them, according to different actual situations, there are many different algorithms for vector quantization, such as trellis vector quantization, product code quantization, and so on.

The basic principle of 3D audio is to make the listener feel that the sound is emitted from the corresponding position of a certain sound source in the space by simulating the sound field generated by a certain sound source in the space. Among them, the judgment of the horizontal orientation information of the sound source is mainly completed by relying on the binaural effect. The horizontal orientation parameter is an important parameter used to express the horizontal orientation information of the sound source in 3D audio, mainly including time difference (ITD, interauraltime difference) and intensity difference (ILD, interaural level difference). Physiologically, the relative auditory importance of ITD and ILD is influenced by factors such as frequency and temporal structure of excitation signals. The human ear has certain rules for the frequency domain resolution of sound. In the low frequency band, the human ear has a high frequency domain resolution, and two sounds with similar frequencies will not mask each other; but in the high frequency band, the human ear's resolution is relatively low. In the actual audio codec, the signal will be divided into several subbands according to this characteristic, and the corresponding ILD and ITD will be extracted for each subband. Therefore, when quantifying the horizontal orientation parameter, it is also necessary to consider the perception ability of the human ear in different frequency bands for quantization processing, so that the quantization result is more in line with the subjective sense of hearing of the human ear. The existing quantization algorithm only quantifies the horizontal orientation parameters, does not consider the influence of frequency band and other related information on the subjective sense of hearing, and takes the best objective performance as the quantification goal, but the objective performance is not completely consistent with the subjective sense of hearing of the human ear, resulting in The subjectively perceived performance of quantizers is not optimal.

In view of the above problems, there is an urgent need to take into account the additional information of the horizontal orientation parameters, and at the same time, the human ear's perception of sound, and a quantizer that conforms to the human ear's subjective sense of hearing, so as to improve the overall subjective performance of encoding.

Contents of the invention

Aiming at the deficiencies of the existing quantization technology, the present invention proposes a quantization technical scheme suitable for human ear perception and applied to the encoding of 3D audio horizontal orientation parameters. During the quantization process, other information related to the horizontal orientation parameter is considered to subjectively perceive distortion. Screen the standard quantization algorithm for the optimal codeword. The purpose is to reduce the perceptual distortion of the human ear and improve the overall quantitative subjective performance by combining the relevant information of the horizontal orientation parameters and selecting codewords according to the evaluation criteria that conform to the perceptual characteristics of the human ear.

The technical solution of the present invention is a method for quantizing horizontal orientation parameters in 3D audio, comprising the following steps:

Step 1.1. Set the total number of sub-bands of 3D audio as N, input the horizontal orientation parameters x ₁ , x ₂ ,…,x _N of N sub-bands, and obtain the corresponding JND values jnd ₁ , jnd ₂ ,…,jnd _N by looking up the table ;

Step 1.2. According to the preset dimension k, divide and obtain multiple k-dimensional horizontal orientation parameter vectors [(x ₁ ,x ₂ ,…,x _k )(x _k+1 ,x _k+2 ,…,x _2k )...(x _N-k+1 ,x _N-k+2 ,...,x _N )], generate the corresponding codebook according to the division result; meanwhile, divide the corresponding JND value into multiple k-dimensional vectors [(jnd ₁ ,jnd ₂ ,...,jnd _k )(jnd _k+1 ,jnd _k+2 ,...,jnd _2k )...(jnd _N-k+1 ,jnd _N-k+2 ,...,jnd _N )];

Step 1.3. Quantize each horizontal orientation parameter vector (x _i+1 , x _i+2 ,...,x _i+k ) to obtain a k-dimensional codeword vector (y _i+1 , y _i+2 ,..., y _i+k ), the value of i is 1,k+1...N-k+1; the implementation method is, for each horizontal orientation parameter vector (xi ₊₁ , _xi+2 ,…,xi _{+ k} ) perform the following sub-steps,

Step 1.3.1. Read a codeword in order from the codebook, and use the read-in codeword as the current quantization of the k-dimensional codeword vector (y _i+1 , y _i+2 ,...,y _i+k ) result;

Step 1.3.2, the current quantization result of the k-dimensional codeword vector (y _i+1 , y _i+2 ,...,y _i+k ) obtained in step 1.3.1 and the corresponding k-dimensional vector (jnd _{i+ 1} , jnd _i+2 ,...,jnd _i+k ) calculate the quantified subjective perceptual distortion

Step 1.3.3, return to step 1.3.1 Read the next codeword in order from the codebook and use it as the current quantization of the k-dimensional codeword vector (y _i+1 , y _i+2 ,...,y _i+k ) As a result, until the codewords in the codebook are traversed, then according to the results of step 1.3.2 each time, the corresponding codeword with the smallest value of the subjective perceptual distortion D _sp is selected as the k-dimensional codeword vector (y _i+1 , y _i+2 ,…,y _i+k ) final quantization result;

Step 1.4, the k-dimensional codeword vector obtained according to step 1.3.3 [(y ₁ ,y ₂ ,…,y _k )(y _k+1 ,y _k+2 ,…,y _2k )…(y _{N-k+ 1} ,y _N-k+2 ,...,y _N )] the final quantization result, and output the quantized horizontal orientation parameters {y ₁ ,y ₂ ,...,y _N } of N subbands and the corresponding index values Index ₁ , Index ₂ ,…,Index _N , output the index value to the code stream.

Moreover, the preset dimension k is 2,

In step 1.2, multiple two-dimensional horizontal orientation parameter vectors [(x ₁ ,x ₂ )(x ₃ ,x ₄ )…(x _N-1 ,x _N )] are obtained by division, and corresponding codebooks are generated according to the division results ; At the same time, divide the corresponding JND values into multiple two-dimensional vectors [(jnd ₁ ,jnd ₂ )(jnd ₃ ,jnd ₄ )...(jnd _N-1 ,jnd _N )];

In step 1.3, each horizontal orientation parameter vector (xi _, xi ₊₁ ) is quantized to obtain a two-dimensional code word vector (y _i ,y _i+1 ), and the value of i is 1,3...N- 1; the implementation method is to execute the following sub-steps for each horizontal orientation parameter vector ( _xi ,xi ₊₁ ),

Step 1.3.1, read in a codeword in sequence from the codebook, and use the read-in codeword as the current component of the two-dimensional codeword vector (y _i , y _i+1 );

Step 1.3.2, calculate the subjective perception of quantization according to the current component of the two-dimensional code word vector (y _i , y _i+1 ) obtained in step 1.3.1 and the corresponding two-dimensional vector (jnd _i , jnd _i+1 ) obtained in step 1.2 distortion

Step 1.3.3, return to step 1.3.1 Read the next codeword in order from the codebook and use it as the current component of the two-dimensional codeword vector (y _i , y _i+1 ), until the codes in the codebook are traversed word, and then select the corresponding codeword in which the value of the subjective perceptual distortion _Dsp is the smallest according to the result of step 1.3.2 each time as the final result of the two-dimensional codeword vector (y _i , y _i+1 );

In step 1.4, according to the two-dimensional code word vector obtained in step 1.3.3 [(y ₁ ,y ₂ )(y ₃ ,y ₄ )…(y _N-1 ,y _N )], output the quantized horizontal orientation of N subbands The final quantization results of the parameters {y ₁ , y ₂ ,...,y _N } and the corresponding index values Index ₁ , Index ₂ ,...,Index _N , output the index values to the code stream.

The horizontal orientation parameter quantization coding technical solution provided by the present invention can be applied to various quantization methods. On the basis of the original quantization algorithm using statistical distortion to screen the optimal codeword, combined with the relevant information of the horizontal orientation parameter, the subjective perception distortion of quantization is calculated. , not only retains the original characteristics of various quantization devices, but also makes the quantization results more in line with the subjective perception characteristics of the human ear. In order to calculate the subjective distortion of the quantization of the horizontal orientation parameter, the present invention introduces the minimum difference JND value perceivable by the human ear, which is a quantity that changes with the size of the horizontal orientation parameter and the frequency band in which it is located. On the basis of calculating the quantitative objective distortion, it is combined with the JND value (perceptible difference value) to obtain the subjective perception distortion. The calculated subjective distortion can well reflect the subjective hearing perception of the human ear, and is helpful to improve the subjective perception performance of the whole quantization device.

Description of drawings

Fig. 1 is the encoding flowchart of the embodiment of the present invention;

Fig. 2 is a decoding flowchart of an embodiment of the present invention.

Detailed ways

The invention provides a method for quantizing horizontal orientation parameters in 3D audio. The main idea is to input the horizontal orientation parameters to be quantized, and obtain the JND value corresponding to the frequency band by looking up the table; Quantize this horizontal orientation parameter; calculate the statistical distortion after quantization using the current codeword, and calculate the subjective perception distortion using the current codeword quantization by combining the statistical distortion and the JND value of the horizontal orientation parameter to be quantized; until the entire codebook is traversed , select the codeword that minimizes the subjective quantization distortion and output its index value. The corresponding decoding process includes receiving the index value input from the encoder; searching the codebook according to the index value to obtain the corresponding quantization horizontal orientation parameter.

During specific implementation, computer software means can be used by those skilled in the art to realize automatic quantization coding according to the technical solution provided. Since in encoding applications, the encoding software method is often solidified to form an encoding device, the encoding device implemented according to the method provided by the present invention is also within the scope of protection.

The present invention groups the horizontal orientation parameters according to the preset dimension k, obtains multiple k-dimensional horizontal orientation parameter vectors, and then calculates the quantified subjective perception distortion, and can accumulate errors caused by each component. Therefore, the value range of k is 1, 2, ... N, when it is 1, it is equivalent to each horizontal orientation parameter as a vector, when it is N, it is equivalent to all horizontal orientation parameters as a vector. During specific implementation, those skilled in the art can set the value by themselves according to specific situations. Generally, N is an integer multiple of k. If it is not an integer multiple, the last divided horizontal orientation parameter vector is less than k-dimensional, and it can be processed according to the actual dimension. The present invention suggests that the value of k is 2 or 3.

In the embodiment, the preset dimension k is 2, split vector quantization in vector quantization is selected as the quantization algorithm, and the statistical distortion adopted is the mean square error. The technical solutions of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

As shown in accompanying drawing 1, the quantization encoding step of embodiment is as follows:

Step 1.1. Set the total number of sub-bands of 3D audio as N (N is an even number), input the horizontal orientation parameters x ₁ , x ₂ ,…,x _N of N sub-bands, and write it as the set X={x ₁ ,x ₂ ,… ,x _N }, and look up the JND table according to the size of the horizontal orientation parameter and the frequency band where it is located, and obtain the corresponding JND values jnd ₁ ,jnd ₂ ,...,jnd _N by looking up the table, which is recorded as the set JND={jnd ₁ ,jnd ₂ , ..., jnd _N }.

In step 1.2, divide according to the preset dimension k=2 to obtain multiple two-dimensional horizontal orientation parameter vectors [(x ₁ ,x ₂ )(x ₃ ,x ₄ )…(x _N-1 ,x _N ) ], generate the corresponding codebook according to the division result; at the same time, divide the corresponding JND values into multiple two-dimensional vectors [(jnd ₁ ,jnd ₂ )(jnd ₃ ,jnd ₄ )…(jnd _N-1 , jnd _N )].

During specific implementation, after the dimension k is determined, the codebook can be trained according to the dimension requirement, and the trained codebook can be put into the quantizer as a component of the quantizer. The specific training method is the prior art, and will not be described in detail in the present invention. The number of codewords in the codebook can be set according to the accuracy requirement.

Step 1.3. Quantize each horizontal orientation parameter vector (xi _, xi ₊₁ ) to obtain a two-dimensional code word vector (y _i ,y _i+1 ), and the value of i is 1,3...N-1 .

In this step, vector quantization is performed. The codebook is a set of codewords trained in advance, and quantization is to select the codeword closest to the vector to be quantized from the codebook.

Among them, the "closest" evaluation standard is the calculation of distortion (error). Therefore, in quantization, each codeword is read in order to calculate the distortion, and the codeword with the smallest distortion is selected as the quantization result through iteration. The embodiment is to select The corresponding two codewords corresponding to the minimum value of the subjective perceptual distortion D _sp are used as the two-dimensional codeword vector (y _i , y _i+1 ).

Therefore, the embodiment performs the following sub-steps for each horizontal orientation parameter vector ( _xi ,xi ₊₁ ),

Step 1.3.1. Read in a codeword quantized horizontal orientation parameter vector (x _i , x _i+1 ) in sequence from the codebook, and use the read-in codeword as a two-dimensional codeword vector (y _i , y _{i+ 1} ) The current quantization result. For example, for the horizontal orientation parameter vector (x ₁ , x ₂ ), when step 1.3.1 is executed for the first time, the first codeword (y ₁ ¹ , y ₂ ¹ ) is read from the codebook, and step 1.3 is executed again. 1, read the second codeword (y ₁ ² , y ₂ ² ) from the codebook...and so on until the last codeword (y ₁ ^m ,y ₂ ^m ) in the codebook is read in, where m is the code The number of codewords in the book.

Step 1.3.2, according to the current quantization result of the two-dimensional codeword vector (y _i , y _i+1 ) obtained in step 1.3.1 and the corresponding two-dimensional vector (jnd _i , jnd _i+1 ) obtained in step 1.2, calculate the subjective quantization Perceptual distortion

In ordinary vector quantization, the calculation of quantization objective distortion uses the mean square error formula However, in the embodiment, the subjective perceptual distortion is calculated on the basis of the mean square error, that is, the subjective perceptual distortion D _sp is calculated by combining the modified mean square error formula and the JND value (jnd _i , jnd _i+1 ).

Step 1.3.3, return to step 1.3.1 Read the codewords in order from the codebook and use them as the current quantization result of the two-dimensional codeword vector (y _i , y _i+1 ), until the codewords in the codebook are traversed , and then select the corresponding codeword with the smallest value of subjective perceptual distortion D _sp as the final quantization result of the two-dimensional codeword vector (y _i , y _i+1 ), that is, for the horizontal orientation parameter vector (xi _, x _i+1 ) The vector quantization result of the two-dimensional horizontal orientation parameters found. For example, for the horizontal orientation parameter vector (x ₁ ,x ₂ ), it is to select the subjective perception distortion from (y ₁ ¹ ,y ₂ ¹ ), (y ₁ ² ,y ₂ ² )...(y ₁ ^m ,y ₂ ^m ) The one with the smallest value of D _sp is taken as the final quantization result.

In step 1.4, according to the two-dimensional code word vector [(y ₁ ,y ₂ )(y ₃ ,y ₄ )…(y _N-1 ,y _N )] obtained in step 1.3.3, output the quantized horizontal orientation of N subbands The final quantization results of the parameters {y ₁ , y ₂ ,...,y _N } and the corresponding index values Index ₁ , Index ₂ ,...,Index _N , output the index values to the code stream.

As shown in accompanying drawing 1, the decoding step of embodiment can be as follows:

Step 2.1. Read the index values Index ₁ , Index ₂ ,..., Index _N from the code stream.

Step 2.2: Search the codebook according to the index values Index ₁ , Index ₂ , ..., Index _N to obtain the quantized horizontal orientation parameters {y ₁ , y ₂ , ..., y _N }.

In order to illustrate the effect of the present invention, the technical solution provided by the embodiment is used for a specific quantization example, and the input N sub-band horizontal orientation parameters are 6.

（a）6维输入待量化水平方位参数X={23.709299，-20.163263，0.947107，2.862993，34.412148，22.736864}，查表得到对应的JND值为JND={0.815555417，0.66236129，0.601276375，0.732569875，0.619971，0.722825075 }.

(b) Compile the input data into three two-dimensional vectors (23.709299, -20.163263), (0.947107, 2.862993), (34.412148, 22.736864), and corresponding JND values also form three two-dimensional vectors (0.815555417, 0.66236129), (0.601276375, 0.732569875), (0.619971, 0.722825075).

(c) Quantize the three vectors in X respectively. The JND value corresponding to the vector (23.709299, -20.163263) is (0.619971, 0.722825075), calculate the statistical distortion one by one with the codewords in the codebook, and calculate the subjective perception distortion, The minimum subjective perceptual distortion is obtained as 0.236205, the corresponding codeword is (25.689648, -20.168681), and the output index is 5.

(d) Do the same operation on the remaining two vectors, and the corresponding codewords are (1.542842, 2.684604), (34.535961, 22.689129), which is the quantization result with the best subjective perception effect, and the output corresponding quantization indexes are 3 , 7.

(d) Get the quantized horizontal orientation parameter X′=(25.689648, -20.168681, 1.542842, 2.684604, 34.535961, 22.689129), and output the corresponding quantization indexes as 5, 3, and 7 respectively.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (2)

1. a level orientation parameter quantification method in 3D audio frequency, is characterized in that, comprises the following steps:

Step 1.1, the sub-band sum of establishing 3D audio frequency are N, the level orientation parameter x of N subband of input ₁, x ₂..., x _n, by tabling look-up, obtain corresponding JND value jnd ₁, jnd ₂..., jnd _n;

Step 1.2, according to default dimension k, divide level orientation the parameter vector [(x obtain a plurality of k dimensions ₁, x ₂..., x _k) (x _k+1, x _k+2..., x _2k) ... (x _n-k+1, x _n-k+2..., x _n)], according to division result, generate corresponding code book; Corresponding JND value is divided into a plurality of k n dimensional vector n [(jnd simultaneously ₁, jnd ₂..., jnd _k) (jnd _k+1, jnd _k+2..., jnd _2k) ... (jnd _n-k+1, jnd _n-k+2..., jnd _n)];

Step 1.3, to each level orientation parameter vector (x _i+1, x _i+2..., x _i+k) quantize, obtain k dimension code word vector (y _i+1, y _i+2..., y _i+k), the value of i is 1, k+1 ... N-k+1; Implementation is, to each level orientation parameter vector (x _i+1, x _i+2..., x _i+k) carry out following sub-step,

Step 1.3.1, from code book, read in order a code word, the dimension code word vector (y using the code word of reading in as k _i+1, y _i+2..., y _i+k) current quantized result;

Step 1.3.2, according to 1.3.1 gained k dimension code word vector (y _i+1, y _i+2..., y _i+k) current quantized result and the corresponding k n dimensional vector n of step 1.2 gained (jnd _i+1, jnd _i+2..., jnd _i+k) the subjective perception distortion that calculate to quantize

Step 1.3.3, return to that step 1.3.1 reads in order next code word from code book and as k dimension code word vector (y _i+1, y _i+2..., y _i+k) current quantized result, until traveled through the code word in code book, then according to the result of each execution step 1.3.2, select wherein subjective perception distortion D _spthe corresponding codewords of value minimum as k dimension code word vector (y _i+1, y _i+2..., y _i+k) final quantized result;

Step 1.4, according to step 1.3.3 gained k dimension code word vector [(y ₁, y ₂..., y _k) (y _k+1, y _k+2..., y _2k) ... (y _n-k+1, y _n-k+2..., y _n)] final quantized result, output N subband quantization level direction parameter { y ₁, y ₂..., y _nand respective index value Index ₁, Index ₂..., Index _n, export index value to code stream.

2. level orientation parameter quantification method in 3D audio frequency as claimed in claim 1, is characterized in that: default dimension k is 2,

In step 1.2, divide level orientation the parameter vector [(x that obtains a plurality of two dimensions ₁, x ₂) (x ₃, x ₄) ... (x _n-1, x _n)], according to division result, generate corresponding code book; Two one group of corresponding JND value is divided into a plurality of two-dimensional vector [(jnd simultaneously ₁, jnd ₂) (jnd ₃, jnd ₄) ... (jnd _n-1, jnd _n)];

In step 1.3, to each level orientation parameter vector (x _i, x _i+1) quantize, obtain type of two-dimension codeword vector (y _i, y _i+1), the value of i is 1,3 ... N-1; Implementation is, to each level orientation parameter vector (x _i, x _i+1) carry out following sub-step,

Step 1.3.1, from code book, read in order a code word, using the code word of reading in as type of two-dimension codeword vector (y _i, y _i+1) current component;

Step 1.3.2, according to 1.3.1 gained type of two-dimension codeword vector (y _i, y _i+1) current component and the corresponding two-dimensional vector (jnd of step 1.2 gained _i, jnd _i+1) the subjective perception distortion that calculate to quantize

Step 1.3.3, return to that step 1.3.1 reads in order next code word from code book and as type of two-dimension codeword vector (y _i, y _i+1) current component, until traveled through the code word in code book, then according to the result of each execution step 1.3.2, select wherein subjective perception distortion D _spthe corresponding codewords of value minimum as type of two-dimension codeword vector (y _i, y _i+1) net result;

In step 1.4, according to step 1.3.3 gained type of two-dimension codeword vector [(y ₁, y ₂) (y ₃, y ₄) ... (y _n-1, y _n)], the quantization level direction parameter { y of N subband of output ₁, y ₂..., y _nfinal quantized result and respective index value Index ₁, Index ₂..., Index _n, export index value to code stream.