Noise reduction neural network training method and device, equipment and storage medium

Abstract

The invention belongs to the technical field of signal noise reduction, and discloses a noise reduction neural network training method, device and equipment and a storage medium, and the method comprises the following steps: a training step: obtaining a noise adding signal, and inputting the noise adding signal into an initial neural network to obtain a sampling sequence; performing Gaussian sampling according to the sampling sequence to obtain a noise signal; obtaining an original signal according to the noise adding signal and the noise signal; calculating an error of the initial neural network according to the original signal; judging whether the error is smaller than a preset error threshold; if not, returning to the training step; and if yes, ending training, and taking the initial neural network as a noise reduction neural network. According to the method, the noise reduction neural network can be obtained without collecting pure signals, unsupervised learning is realized, and the method has extremely high applicability in engineering application.

Description

Technical Field

[0001] The present application relates to the technical field of signal noise reduction, and in particular to a training method, apparatus, device and storage medium for a noise reduction neural network. Background Art

[0002] Currently, the algorithms for denoising audio signals include frequency domain denoising algorithm, time domain denoising algorithm, wavelet domain denoising algorithm, etc. These algorithms all require the use of denoising autoencoders. The working principle of the autoencoder requires the input of a set of pure signals and noisy signals, and the noise function is obtained by comparing the encoded and decoded signals with the original pure signals using the noisy signals.

[0003] However, in actual applications, truly pure audio signals without noise are difficult to obtain. Generally, some noise is mixed in the audio signals, which leads to low accuracy of the constructed noise function and poor noise reduction effect of the algorithm. Summary of the invention

[0004] The present application provides a training method, apparatus, device and storage medium for a denoising neural network, which can obtain a denoising neural network without collecting pure signals, realizes unsupervised learning, and has extremely high applicability in engineering applications.

[0005] In a first aspect, an embodiment of the present application provides a method for training a denoising neural network, comprising:

[0006] Training steps: Get the noisy signal and input it into the initial neural network to get the sampling sequence;

[0007] Gaussian sampling is performed according to the sampling sequence to obtain a noise signal;

[0008] The original signal is obtained according to the noise-added signal and the noise signal;

[0009] Calculate the error of the initial neural network based on the original signal;

[0010] Determine whether the error is less than a preset error threshold; if not, return to the training step; if so, end the training and use the initial neural network as the denoising neural network.

[0011] Furthermore, the method also includes:

[0012] After getting the original signal, add 1 to the number of training times;

[0013] Determine whether the number of training times reaches the preset training threshold; if not, return to the training step; if so, end the training and use the initial neural network as the denoising neural network.

[0014] Furthermore, the above-mentioned obtaining the original signal according to the noise-added signal and the noise signal includes:

[0015] Subtract the noise signal from the noisy signal to get the original signal.

[0016] Furthermore, the above calculation of the error of the initial neural network based on the original signal includes:

[0017] Calculate the adjacent difference sequence of the original signal;

[0018] Compute the variance of a series of adjacent differences; use the variance as the error.

[0019] Furthermore, the sampling sequence includes a mean value and a variance of a noise signal in a noise-added signal; and the noise-added signal is a noise-added frequency signal.

[0020] In a second aspect, an embodiment of the present application provides a training device for a denoising neural network, comprising:

[0021] A training module is used to obtain a noise-added signal and input it into an initial neural network to obtain a sampling sequence;

[0022] A sampling module, used for performing Gaussian sampling according to a sampling sequence to obtain a noise signal;

[0023] A noise reduction module, used for obtaining an original signal according to a noise-added signal and a noise signal;

[0024] An error calculation module, used to calculate the error of the initial neural network based on the original signal;

[0025] The judgment module is used to return to the training module when the error is greater than or equal to the preset error threshold; when the error is less than the preset error threshold, the initial neural network is used as the denoising neural network.

[0026] Furthermore, the noise reduction module is specifically used to subtract the noise signal from the noise-added signal to obtain the original signal.

[0027] Furthermore, the error calculation module is specifically used to calculate the adjacent difference sequence of the original signal and the variance of the adjacent difference sequence; and the variance is used as the error.

[0028] In a third aspect, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a denoising neural network training method as described in any of the above embodiments are performed.

[0029] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for training a denoising neural network as in any of the above embodiments.

[0030] In summary, compared with the prior art, the technical solution provided in the embodiment of the present application has at least the following beneficial effects:

[0031] The embodiment of the present application provides a method for training a denoising neural network. During training, it is only necessary to obtain a noise signal, and then use the output of the neural network to sample a Gaussian signal to obtain a noise signal, and then obtain the original signal based on the noise signal and the noise signal. Since the output of the initial neural network may not be accurate during the training process, it is also necessary to calculate the error based on the original signal. If the error is less than the preset error threshold, it means that the output accuracy of the initial neural network has reached the requirement of noise reduction, and the initial neural network can be denoised. The above method does not need to collect pure signals during the training of the denoising neural network, realizes unsupervised learning, and has extremely high applicability in engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a method for training a denoising neural network is provided as an exemplary embodiment of the present application.

[0033] Figure 2 A structural diagram of a denoising neural network training device provided as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0035] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0036] Existing noise reduction algorithms include frequency domain noise reduction algorithm, time domain noise reduction algorithm, wavelet domain noise reduction algorithm and so on.

[0037] The frequency domain noise reduction algorithm is based on the frequency characteristics of the signal. By analyzing the difference between the noise and the signal in the signal spectrum, the noise component is separated from the signal to achieve noise elimination. In the field of audio processing, the commonly used frequency domain noise reduction algorithms include FIR filters based on fast Fourier transform and Kalman filters.

[0038] The time domain noise reduction algorithm mainly separates and eliminates noise by analyzing the time domain waveform characteristics of noise signals and non-noise signals. The time domain noise reduction algorithm usually includes adaptive noise reduction algorithm and deep learning-based noise reduction algorithm.

[0039] Wavelet transform is a signal decomposition and reconstruction technology that decomposes the signal into multiple frequency components and processes each component accordingly. In the field of noise reduction, the wavelet domain noise reduction algorithm can be analyzed and processed based on both the frequency domain and the time domain.

[0040] These noise reduction algorithms all require the use of denoising autoencoders. The working principle of the autoencoder requires the input of a set of pure signals and noisy signals, and the noise function is obtained by comparing the encoded and decoded signals with the original pure signals.

[0041] However, in actual applications, truly pure audio signals without noise are difficult to obtain. Generally, some noise is mixed in them, which results in low accuracy of the constructed noise function and poor noise reduction effect.

[0042] See also Figure 1 In order to solve the problem that it is difficult to obtain a pure signal without noise in existing projects, the present application embodiment provides a method for training a denoising neural network, which specifically includes:

[0043] Step S1, training step: obtain a noisy signal and input it into an initial neural network to obtain a sampling sequence.

[0044] The noise-added signal may be a noise-added frequency signal.

[0045] The initial neural network can adopt convolutional neural network or deep residual shrinkage network.

[0046] The sampling sequence is a two-dimensional sequence, which represents the mean and variance of the noise signal in the noise-added signal.

[0047] Step S2, performing Gaussian sampling according to the sampling sequence to obtain a noise signal.

[0048] The Gaussian signal used may be an existing one-dimensional Gaussian signal.

[0049] Step S3, obtaining the original signal according to the noise-added signal and the noise signal.

[0050] Specifically, the noise signal is subtracted from the noise-added signal to obtain the original signal.

[0051] Step S4, calculating the error of the initial neural network based on the original signal.

[0052] Specifically, the adjacent difference sequence of the original signal is calculated; the variance of the adjacent difference sequence is calculated; and the variance is used as the error.

[0053] Assuming that the sampling sequence is [X1, X2, X3, X4], the adjacent differential sequence of the original signal here is [Y1 = X2-X1, Y2 = X3-X2, Y3 = X4-X3], and the variance of the adjacent differential sequence is taken as the error.

[0054] Step S5, determine whether the error is less than a preset error threshold; if not, return to the training step; if yes, end the training and use the initial neural network as the denoising neural network. Specifically, in addition to judging based on the error, the end of training can also be achieved by limiting the number of training times: after obtaining the original signal, add 1 to the number of training times; determine whether the number of training times reaches the preset training threshold; if not, return to the training step; if yes, end the training and use the initial neural network as the denoising neural network.

[0055] In the specific implementation process, after obtaining the trained denoising neural network, the audio signal to be denoised can be input into the denoising neural network. After obtaining the sampling sequence, Gaussian sampling is performed according to the sampling sequence. The sampling result is used as the noise signal, and the noise signal is subtracted from the audio signal to be denoised to obtain the denoised, pure original audio signal.

[0056] The above embodiment provides a method for training a denoising neural network. During training, it is only necessary to obtain a noise signal, and then use the output of the neural network to sample the Gaussian signal to obtain a noise signal, and then obtain the original signal based on the noise signal and the noise signal. Since the output of the initial neural network may not be accurate during the training process, it is also necessary to calculate the error based on the original signal. If the error is less than the preset error threshold, it means that the output accuracy of the initial neural network has reached the noise reduction requirement, and the initial neural network can be denoised. The above method does not need to collect pure signals during the training of the denoising neural network, realizes unsupervised learning, and has extremely high applicability in engineering applications.

[0057] See also Figure 2 Another embodiment of the present application provides a training device for a denoising neural network, comprising:

[0058] The training module 101 is used to obtain a noisy signal and input it into an initial neural network to obtain a sampling sequence.

[0059] The sampling module 102 is used to perform Gaussian sampling according to the sampling sequence to obtain a noise signal.

[0060] The noise reduction module 103 is used to obtain the original signal according to the noise-added signal and the noise signal.

[0061] The error calculation module 104 is used to calculate the error of the initial neural network according to the original signal.

[0062] The judgment module 105 is used to return to the training module when the error is greater than or equal to the preset error threshold; when the error is less than the preset error threshold, use the initial neural network as the denoising neural network.

[0063] Furthermore, the noise reduction module 103 is specifically configured to subtract the noise signal from the noise-added signal to obtain the original signal.

[0064] Furthermore, the error calculation module 104 is specifically used to calculate the adjacent difference sequence of the original signal and the variance of the adjacent difference sequence; and the variance is used as the error.

[0065] For the specific limitations of a denoising neural network training device provided in this embodiment, please refer to the embodiment of a denoising neural network training method described above, which will not be repeated here. Each module in the above-mentioned denoising neural network training device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0066] An embodiment of the present application provides a computer device, which may include a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the processor executes the steps of a training method for a denoising neural network as in any of the above embodiments.

[0067] The working process, working details and technical effects of the computer device provided in this embodiment can be found in the above embodiment of a training method for a denoising neural network, which will not be described in detail here.

[0068] An embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a training method for a denoising neural network as in any of the above embodiments are implemented. Wherein, the computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive and / or a memory stick (Memory Stick), etc., and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The working process, working details and technical effects of the computer-readable storage medium provided in this embodiment can be found in the above embodiment of a training method for a denoising neural network, which will not be repeated here.

[0069] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0070] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for training a denoising neural network, characterized in that: include: Training steps: Get the noisy signal and input it into the initial neural network to get the sampling sequence; Perform Gaussian sampling according to the sampling sequence to obtain a noise signal; Obtaining an original signal according to the noise-added signal and the noise signal; Calculating the error of the initial neural network according to the original signal; Determine whether the error is less than a preset error threshold; if not, return to the training step; if so, end the training and use the initial neural network as a denoising neural network.

2. The method for training a denoising neural network according to claim 1, characterized in that: Also includes: After obtaining the original signal, the number of training times is increased by 1; Determining whether the number of training times reaches a preset training threshold; If not, return to the training step; If yes, the training is terminated and the initial neural network is used as the denoising neural network.

3. The method for training a denoising neural network according to claim 1, characterized in that: The obtaining the original signal according to the noise-added signal and the noise signal comprises: The noise signal is subtracted from the noise-added signal to obtain the original signal.

4. The method for training a denoising neural network according to claim 1, characterized in that: The calculating the error of the initial neural network according to the original signal comprises: Calculating adjacent difference sequences of the original signal; Calculate the variance of the adjacent difference sequence; and use the variance as the error.

5. The method for training a denoising neural network according to claim 1, characterized in that: The sampling sequence includes the mean and variance of the noise signal in the noise-added signal; the noise-added signal is a noise-added frequency signal.

6. A training device for a denoising neural network, characterized in that: include: A training module is used to obtain a noise-added signal and input it into an initial neural network to obtain a sampling sequence; A sampling module, used for performing Gaussian sampling according to the sampling sequence to obtain a noise signal; A noise reduction module, used for obtaining an original signal according to the noise-added signal and the noise signal; An error calculation module, used for calculating the error of the initial neural network according to the original signal; A judgment module, used for returning to the training module when the error is greater than or equal to a preset error threshold; When the error is less than a preset error threshold, the initial neural network is used as a denoising neural network.

7. The denoising neural network training device according to claim 6, characterized in that: The noise reduction module is specifically configured to subtract the noise signal from the noise-added signal to obtain the original signal.

8. The denoising neural network training device according to claim 6, characterized in that: The error calculation module is specifically used to calculate the adjacent difference sequence of the original signal and the variance of the adjacent difference sequence; and use the variance as the error.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the training method of the denoising neural network as claimed in any one of claims 1 to 5 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the training method of the denoising neural network as claimed in any one of claims 1 to 5 are implemented.

Images