CN105551503B - Based on the preselected Audio Matching method for tracing of atom and system - Google Patents

Abstract

The invention discloses a kind of Audio Matching method for tracing and system preselected based on atom, the invention firstly uses correlations existing between signal energy and Auditory Perception, pretreatment based on energy is carried out to original signal, extracts the higher part signal of its Energy distribution；Matched jamming is carried out for the part signal again, obtains sparse coefficient；Signal reconstruction is carried out by sparse coefficient and original dictionary.Computation complexity and calculating speed can be greatly reduced while guaranteeing sound quality without decline in the present invention.

Description

Audio matching tracking method and system based on atomic preselection

technical field

The invention belongs to the technical field of audio coding, and in particular relates to an audio matching tracking method and system based on atomic pre-selection.

Background technique

Sparse representation generally refers to using as few basis functions as possible to accurately represent the original signal, so as to capture the main features of the signal, thereby substantially reducing the cost of signal processing. Matching pursuit (MP, Matching pursuit) is one of the widely used sparse representation algorithms. Its basic idea is to select the optimal atom from the overcomplete dictionary in turn in the iterative process, so that the approximation of the signal is more optimal. Since the overcomplete dictionary base used by the MP algorithm to represent the signal can be adaptively and flexibly selected according to the characteristics of the signal itself; and it adopts a greedy algorithm of repeated iterative approximation in the process of atom selection, which ensures that the final atomic The number of coefficients is small, and the MP algorithm is widely used in various fields of signal analysis, such as image processing, biomedical signal processing, audio processing, etc.

With the improvement of people's requirements for streaming media quality and the continuous increase of the number of mobile terminal users, the requirements for audio and video coding efficiency are also increasing day by day. The traditional matching pursuit algorithm is not suitable for real-time processing due to its high computational complexity. At present, a variety of fast matching pursuit algorithms have been proposed, such as the joint dictionary method in Reference 1 and the algorithm improvement optimization method in Reference 2. However, these algorithms involve time-consuming optimization, or sacrifice the efficiency of sparse representation for compensation, and the calculation speed is also difficult to meet. The Need for Large-Scale Problems, Reference 3 et al. proposed a traversal algorithm based on short-term Gabor atoms, which uses incomplete fixed-length atoms to traverse from the start to the end of the signal, and selects the optimal matching atom for multiple iterations to obtain the final sparse coefficient. The data volume of this algorithm dictionary is very small, which effectively reduces the computational burden on storage while reducing the computational complexity.

Although this method is slightly less computationally complex than other sparse representation algorithms, it is still difficult to use in real-time applications. One of the main ways to reduce the computational complexity of the matching pursuit algorithm is to reduce the number of iterations. When the sparse dictionary used is a short-term dictionary, the time-consuming of MP algorithm for local long-term signals will be much less than that of the ergodic MP algorithm.

The following references are included in the text:

[1]Ravelli E,Richard G,Daudet L.Union of MDCT bases for audio coding[J].Audio,Speech,and Language Processing,IEEE Transactions on,2008,16(8):1361-1372.

[2] Gharavi-Alkhansari M, Huang T S.A fast orthogonal matching pursuitalgorithm[C]//Acoustics,Speech and Signal Processing,1998.Proceedings of the1998IEEE International Conference on.IEEE,1998,3:1389-1392.

[3] S,Gribonval R.MPTK:Matching pursuit made tractable[C]//Acoustics,Speech and Signal Processing,2006.ICASSP 2006Proceedings.2006IEEEInternational Conference on.IEEE,2006,3:III-III.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the prior art, the present invention provides an audio matching and tracking method and system based on atomic pre-selection according to the influence of energy on auditory perception. The present invention performs matching and tracking on the part with higher signal energy value through successive iterations. , which shortens the time spent in the traversal process without affecting the final reconstruction effect of the signal.

The technical scheme adopted in the present invention is as follows:

An audio matching tracking method based on atomic preselection, comprising:

Signal decomposition and signal reconstruction, wherein the signal decomposition includes the steps:

S1 selects a short-term dictionary according to the original signal type, and uses the short-term dictionary as a sparse dictionary;

S2 calculates the energy of consecutive samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal one by one, i takes 1, 2,...length(S)-N+1 in turn, and the extracted energy is the highest The continuous samples of , denoted as S _maxenergy ; N is the atomic length of the short-term dictionary; length(S) is the length of the original signal;

S3 obtains the atomic weight of each atom in the sparse dictionary on S _maxenergy , and the maximum value of the absolute value of the atomic weight is

S4 calculates the signal residual for the corresponding atom; at the same time, the Recorded in the i _opt max row and the j _opt max column of the current sparse coefficient matrix, i _opt max is The atomic label of , j _opt max is The atomic center position of , the initial value of the current sparse coefficient matrix is zero matrix;

S5, when the signal residual S' _later reaches the target SNR or the number of iterations reaches a preset value, the signal decomposition is ended, and the current sparse coefficient matrix is output; otherwise, steps 2 to 5 are repeated with the current signal residual S' _later as the original signal;

Signal reconstruction includes:

S7 extracts the atomic weights and their corresponding row numbers and column numbers in the current sparse coefficient matrix;

S8 Multiplies the atom weights with the corresponding atoms to obtain the restored signal, assigns each restored signal to the zero vector M _i with the same length as the original signal in step 1, and takes the j _opt max point of the zero vector M _i as the restored signal The center point of , j _opt max is the column number of the atomic weight corresponding to the current restored signal; the assigned vectors are sequentially accumulated to obtain the reconstructed signal.

In step S2, the energy of the continuous samples {S _i , S _i+1 , . . . S _i+N-1 } in the original signal is the sum of the squares of the amplitudes of all the samples in the continuous samples.

In step S2, the energy of the continuous samples {S _i , S _i+1 , . . . S _i+N-1 } in the original signal is the sum of the absolute values of the amplitudes of all samples in the continuous samples.

In step S2, the energy of the continuous samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal is the maximum value of all sample amplitudes in the continuous samples.

The above-mentioned system corresponding to the audio matching tracking method based on atomic pre-selection includes:

A signal decomposition unit and a signal reconstruction unit, wherein the signal decomposition unit further includes:

The dictionary building module 101 is used to select a short-term dictionary according to the original signal type, and use the short-term dictionary as a sparse dictionary;

The preprocessing module 102 is used to calculate the energy of successive samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal one by one, i is 1, 2,...length(S)-N in turn +1, extract the continuous sample with the highest energy, denoted as S _maxenergy ; N is the atomic length of the short-term dictionary; length(S) is the length of the original signal;

The weight comparison module 103 is used to obtain the atomic weight of each atom of the sparse dictionary on S _maxenergy , and the maximum value of the absolute value of the atomic weight is

The residual error calculation module 104 is used to calculate the signal residual error for the corresponding atom; at the same time, the Recorded in the i _opt max row and the j _opt max column of the current sparse coefficient matrix, i _opt max is The atomic label of , j _opt max is The atomic center position of , the initial value of the current sparse coefficient matrix is zero matrix;

The threshold value control module 105 is used to end the signal decomposition and output the current sparse coefficient matrix when the signal residual S' _later reaches the target SNR or the number of iterations reaches a preset value; otherwise, the current signal residual S' _later is input as the original signal preprocessing module 102;

The signal reconstruction unit further includes:

The reconstruction coefficient extraction module 201 is used to extract the atomic weights and their corresponding row numbers and column numbers in the current sparse coefficient matrix;

The signal synthesis module 202 is used to multiply the atomic weights with the corresponding atoms to obtain the recovered signal, assign each recovered signal to a zero vector M _i with the same length as the original signal, and use the zero vector M _i as the j _opt max point is the center point of the restored signal, and j _opt max is the column number of the atomic weight corresponding to the current restored signal; the assigned vectors are sequentially accumulated to obtain the reconstructed signal.

Compared with the prior art, the present invention has the following characteristics:

The invention reduces the number of traversal calculations and reduces the computational complexity by performing the MP algorithm of the incomplete dictionary on the part with high short-term energy in the signal. In the dictionary construction, the frequency span of atoms is increased, and the constraints on frequency components of the dictionary are reduced. The sparse expression algorithm is not limited by the length of the signal to be processed, and the amount of dictionary data is small. Compared with other fast matching pursuit algorithms (such as method) can obtain a faster calculation speed without degrading the sound quality.

Description of drawings

Fig. 1 is the concrete flow chart of the signal decomposition part of the embodiment of the present invention;

Fig. 2 is the specific flow chart of the signal reconstruction part of the embodiment of the present invention;

3 is a structural block diagram of a signal decomposition subsystem according to an embodiment of the present invention;

4 is a structural block diagram of a signal reconstruction subsystem according to an embodiment of the present invention;

Figure 5 is a schematic diagram of the position of the atomic center.

Detailed ways

In order to facilitate understanding and implementation by those skilled in the art, the technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, but not to limit the present invention. .

Figures 1 to 2 show the specific flow of the method of the present invention, including two parts: signal decomposition and signal reconstruction.

The specific implementation of signal decomposition includes the following steps:

Step 1, select a short-term dictionary according to the original signal type.

This step is a conventional step in the audio matching tracking method. For a speech processing system, a short-term dictionary with speech characteristics is selected; for a transient signal processing system, a relatively transient short-term dictionary is selected. For some systems whose features are not obvious or need to deal with multiple types of signals at the same time, a short-term dictionary with strong universality is selected.

In this embodiment, the test samples include speech signals, music signals, etc., and the short-term dictionary selects the Gabor dictionary with strong flexibility. Atoms in the Gabor dictionary are constructed as follows:

In formula (1), w represents the frequency scale; μ represents the time offset; σ represents the time scale; λ _{w, μ, σ} represents the atomic energy under w, μ, and σ; _gw,μ,σ (n) represents the atomic amplitude at time domain sample n.

When the traditional Gabor dictionary-based matching pursuit method takes the value of the time offset μ, the time offset μ of various scales will be obtained as much as possible within the range allowed by the number of atoms in the dictionary. In this embodiment, μ=0, so that all atoms in the dictionary correspond to the part of the signal with higher energy value selected in the preprocessing, and the energy is located at the center position thereof. Assuming that the variation range of n is from 1 to N, and there are M combinations of frequency scale w, time scale σ, and atomic energy λ, the size of the dictionary is M×N. In this embodiment, M is 20, and N is 1001.

Step 2, the original signal is preprocessed, and the energy of successive samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal is calculated one by one, and the continuous sample with the highest energy value is recorded as S _maxenergy , The length of consecutive samples is the atomic length N of the short-term dictionary selected in step 1.

Several energy calculation methods are provided below.

(1) Calculate the energy value Energy of continuous samples according to the energy definition, as follows:

In formula (1), S _i is the ith sample of the original signal S, which is also used to represent the amplitude of the ith sample of the original signal S; m is the scale shift, and m takes 0, 1, ... length(S) in turn -N,length(S) is the length of the original signal S.

(2) Since there is an analogous proportional relationship between the sum of the squares of the amplitudes of the samples and the sum of the absolute values of the amplitudes of the signals, in addition, the calculation amount of the sum of the absolute values of the sample amplitudes is much smaller than that of the sum of the squares of the amplitudes of the samples. Therefore, formula (3) can be used to approximate the energy value Energy of continuous samples:

In formula (3), S _i is the ith sample of the original signal S, and is also used to represent the amplitude of the ith sample of the original signal S; m is the scale shift, and m takes 0, 1, ... length(S) in turn -N,length(S) is the length of the original signal S.

(3) Different energy calculation methods can be selected according to the original signal characteristics. If most of the original signals are signals with relatively continuous amplitudes, the maximum value of the amplitudes of all samples of the continuous sample is used as the energy of the continuous sample. Compared with (1) and (2), this method further reduces the computational complexity.

Step 3, take the short-term dictionary selected in step 1 as the sparse dictionary, make each atom in the sparse dictionary Do inner product with S _maxenergy in turn to get each atom The atomic weight on S _maxenergy , the maximum value of the absolute value of the atomic weight is recorded as

The calculation formula is as follows:

In formula (4), i _opt represents the atomic label in the sparse dictionary, i _opt =1,2,...M, M is the number of atoms in the sparse dictionary; That is, the i _opt -th atom in the sparse dictionary; express and the absolute value of the inner product of S'.

Step 4, calculate the component of S _maxenergy at the largest atom of the sparse dictionary The signal residual S′ _later is S _maxenergy and The vector difference value of , see formula (5); at the same time, update the current sparse coefficient matrix.

in, for The corresponding atom, the largest atom.

The current sparse coefficient matrix is updated as follows:

The initial value of the sparse coefficient matrix is a zero matrix, the row number represents the atomic label, the column number represents the atomic center position, and the element is the atomic weight. The atomic center position is the position of the central sample of the continuous sample S _maxenergy relative to the initial point of the original signal, as shown in Figure 5. If the initial position of the original signal is 0, the atomic center position is m.

is the sparse coefficient matrix before the update, is the updated sparse coefficient matrix, and the weight matrix c has the same size as the sparse coefficient matrix. The way to obtain the weight matrix c is: Assigned to the weight matrix c, i _opt max row, j _opt max column, i _opt max is the largest atom The label of , j _opt max is The atomic center position of , which is the center sample position of S _maxenergy

Step 5, when the signal residual S' _later reaches the target SNR or the number of iterations reaches a preset value, the signal decomposition is ended, and the current sparse coefficient matrix is output; otherwise, the signal residual S' _later is used as the original signal in step 2. Repeat the steps 2 to 5.

The matching pursuit method processes the signal by accumulating iteratively, and expresses the original signal as the sum of the superposition of the multiplication of the atomic weight and the corresponding atom and the sum of the signal residual. The signal residual S′ _later can be obtained from step 4. When S′ _later reaches the target SNR or the number of iterations reaches a preset value, the iteration is terminated, and the current sparse coefficient matrix is output. The target SNR and the preset number of iterations are artificially set according to experience and actual needs.

The signal-to-noise ratio SNR is defined as follows:

In formula (7), S represents the original signal, and S′ is the signal after the sparse restoration.

In this embodiment, for a segment signal whose sampling frequency is 48 kHz, its length is 500,000 sample points (10 s), the preset number of iterations is 20,000 times, and the target SNR is 20 dB.

A signal reconstruction method, the specific implementation of which includes the following steps:

Step 6: Extract from the current sparse coefficient matrix the atomic weight to be used in the reconstructed signal and the atomic label and atomic center position corresponding to the atomic weight.

Step 7, put the atomic weights their corresponding atoms Do the product to get the recovered signal of length N Each recovered signal They are respectively assigned to the zero vector M _i with the same length as the original signal in step 1, and the j _opt max point of the zero vector M _i is used as the recovered signal during assignment. The center point of , j _opt max is the atomic weight The column number in the current sparse coefficient matrix; the assigned vectors _Mi are successively accumulated to obtain the reconstructed signal S'.

The reconstructed signal synthesis formula is as follows:

where k is the number of atomic weights in the current sparse coefficient matrix.

Referring to Figures 3-4, the present invention also provides an audio matching tracking system based on atomic pre-selection, including a signal decomposition unit and a signal reconstruction unit. The signal decomposition unit further includes a dictionary building module 101, a preprocessing module 102, a weight comparison module 103, a residual calculation module 104, and a threshold control module 105; the signal reconstruction unit further includes a reconstruction coefficient extraction module 201 and a signal synthesis module 202. in:

The dictionary building module 101 is used to select a short-term dictionary according to the original signal type, and use the short-term dictionary as a sparse dictionary.

The preprocessing module 102 is used to calculate the energy of successive samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal one by one, i is 1, 2,...length(S)-N in turn +1, extract the continuous sample with the highest energy, denoted as S _maxenergy ; N is the atomic length of the short-term dictionary; length(S) is the length of the original signal.

In the preprocessing module 102, the continuous sample energy can be calculated in the following manner:

(1) The square sum of all sample amplitudes in the continuous samples {S _i , S _i+1 ,...S _i+N-1 } is taken as the energy of the continuous sample, see formula (2).

(2) The sum of the absolute values of all sample amplitudes in the continuous samples {S _i , S _i+1 ,...S _i+N-1 } is taken as the energy of the continuous sample, see formula (3).

(3) The maximum value of all sample amplitudes in the continuous samples {S _i , S _i+1 ,...S _i+N-1 } is taken as the energy of the continuous sample.

The weight comparison module 103 is used to obtain the atomic weight of each atom of the sparse dictionary on S _maxenergy , and the maximum value of the absolute value of the atomic weight is

The residual error calculation module 104 is used to calculate the signal residual error for the corresponding atom; at the same time, the Recorded in the i _opt max row and the j _opt max column of the current sparse coefficient matrix, i _opt max is The atomic label of , j _opt max is The atomic center position of , the initial value of the current sparse coefficient matrix is zero matrix.

The threshold value control module 105 is used to end the signal decomposition and output the current sparse coefficient matrix when the signal residual S' _later reaches the target SNR or the number of iterations reaches a preset value; otherwise, the current signal residual S' _later is input as the original signal Preprocessing module 102 .

The reconstruction coefficient extraction module 201 is used to extract the atomic weights and their corresponding row numbers and column numbers in the current sparse coefficient matrix.

The signal synthesis module 202 is used to multiply the atomic weights with the corresponding atoms to obtain the recovered signal, assign each recovered signal to a zero vector M _i with the same length as the original signal, and use the zero vector M _i as the j _opt max point is the center point of the restored signal, and j _opt max is the column number of the atomic weight corresponding to the current restored signal; the assigned vectors are sequentially accumulated to obtain the reconstructed signal.

It should be understood that the parts not described in detail in this specification belong to the prior art.

It should be understood that the above description of the preferred embodiments is relatively detailed, and therefore should not be considered as a limitation on the protection scope of the patent of the present invention. In the case of the protection scope, substitutions or deformations can also be made, which all fall within the protection scope of the present invention, and the claimed protection scope of the present invention shall be subject to the appended claims.

Claims (5)

1. an audio matching tracking method based on atomic pre-selection, is characterized in that, comprises: Signal decomposition and signal reconstruction, wherein the signal decomposition includes the steps: S1 selects a short-term dictionary according to the original signal type, and uses the short-term dictionary as a sparse dictionary; S2 calculates the energy of consecutive samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal one by one, i takes 1, 2,...length(S)-N+1 in turn, and the extracted energy is the highest The continuous samples of , denoted as S _maxenergy ; N is the atomic length of the short-term dictionary; length(S) is the length of the original signal; S3 obtains the atomic weight of each atom in the sparse dictionary on S _maxenergy , and the maximum value of the absolute value of the atomic weight is S4 calculates the signal residual for the corresponding atom; at the same time, the Recorded in the i _opt max row and the j _opt max column of the current sparse coefficient matrix, i _opt max is The atomic label of , j _opt max is The atomic center position of , the initial value of the current sparse coefficient matrix is zero matrix; S5, when the signal residual S' _later reaches the target SNR or the number of iterations reaches a preset value, the signal decomposition is ended, and the current sparse coefficient matrix is output; otherwise, steps 2 to 5 are repeated with the current signal residual S' _later as the original signal; Signal reconstruction includes: S7 extracts the atomic weights and their corresponding row numbers and column numbers in the current sparse coefficient matrix; S8 Multiplies the atom weights with the corresponding atoms to obtain the restored signal, assigns each restored signal to a zero vector M _i with the same length as the original signal in step 1, and takes the j _opt max point of the zero vector M _i as the restored signal The center point of , j _opt max is the column number of the atomic weight corresponding to the current restored signal; the assigned vectors are sequentially accumulated to obtain the reconstructed signal. 2. the audio frequency matching tracking method based on atomic pre-selection as claimed in claim 1, is characterized in that: In step S2, the energy of the continuous samples {S _i , S _i+1 , . . . S _i+N-1 } in the original signal is the sum of the squares of the amplitudes of all the samples in the continuous samples. 3. the audio matching tracking method based on atomic pre-selection as claimed in claim 1, is characterized in that: In step S2, the energy of the continuous samples {S _i , S _i+1 , . . . S _i+N-1 } in the original signal is the sum of the absolute values of the amplitudes of all samples in the continuous samples. 4. the audio matching tracking method based on atomic pre-selection as claimed in claim 1, is characterized in that: In step S2, the energy of the continuous samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal is the maximum value of all sample amplitudes in the continuous samples. 5. An audio matching tracking system based on atomic pre-selection, is characterized in that, comprises: A signal decomposition unit and a signal reconstruction unit, wherein the signal decomposition unit further includes: The dictionary building module 101 is used to select a short-term dictionary according to the original signal type, and use the short-term dictionary as a sparse dictionary; The preprocessing module 102 is used to calculate the energy of successive samples {S _i , S _i+1 ,...S _i+N-1 } in the original signal one by one, i is 1, 2,...length(S)-N in turn +1, extract the continuous sample with the highest energy, denoted as S _maxenergy ; N is the atomic length of the short-term dictionary; length(S) is the length of the original signal; The weight comparison module 103 is used to obtain the atomic weight of each atom of the sparse dictionary on S _maxenergy , and the maximum value of the absolute value of the atomic weight is The residual error calculation module 104 is used to calculate the signal residual error for the corresponding atom; at the same time, the Recorded in the i _opt max row and the j _opt max column of the current sparse coefficient matrix, i _opt max is The atomic label of , j _opt max is The atomic center position of , the initial value of the current sparse coefficient matrix is zero matrix; The threshold control module 105 is used to end the signal decomposition and output the current sparse coefficient matrix when the signal residual S' _later reaches the target SNR or the number of iterations reaches a preset value; otherwise, the current signal residual S' _later is input as the original signal preprocessing module 102; The signal reconstruction unit further includes: The reconstruction coefficient extraction module 201 is used to extract the atomic weights and their corresponding row numbers and column numbers in the current sparse coefficient matrix; The signal synthesis module 202 is used to multiply the atomic weights with the corresponding atoms to obtain the recovered signal, assign each recovered signal to a zero vector M _i with the same length as the original signal, and use the zero vector M _i as the j _opt max point is the center point of the restored signal, and j _opt max is the column number of the atomic weight corresponding to the current restored signal; the assigned vectors are sequentially accumulated to obtain the reconstructed signal.