Underwater sound echo noise reduction method based on wavelet convolution U-Net network

Abstract

The invention relates to the technical field of underwater acoustic signal processing, in particular to an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network. According to the underwater acoustic echo noise reduction method provided by the invention, the underwater acoustic echo can be subjected to noise reduction more efficiently and accurately; an existing technical scheme is not flexible enough to process frequency characteristics of underwater acoustic echoes. The wavelet convolution and reattention mechanism endows the model with multi-resolution analysis capability, can better adapt to the time-varying characteristics of the underwater acoustic echo at different frequencies, and can more meticulously analyze the underwater acoustic echo at different frequencies and time scales. The U-Net architecture is excellent in performance in the aspect of dense prediction tasks, and the method combines U-Net with wavelet convolution. Compared with a traditional underwater acoustic echo method based on a single structure, the scheme provided by the invention provides a more effective means for extracting and fusing features at different levels, and useful information in underwater acoustic echoes can be better mined.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic signal processing, and in particular to an underwater acoustic echo denoising method based on a wavelet convolution U-Net network. Background Art

[0002] In underwater environments, the propagation and application of acoustic signals are extremely extensive, covering many key areas such as marine resource exploration, underwater target detection, marine ecological monitoring, and submarine communications. However, underwater acoustic echoes are often interfered by a variety of complex factors, resulting in a large amount of noise in the acquired acoustic signals, which seriously affects the subsequent accurate extraction of effective information in the acoustic signals and the reliability and accuracy of related applications.

[0003] For example, in underwater target detection scenarios, the noise in the echo may mask the true characteristics of the target, leading to misjudgment of the target's position, shape or attributes; in marine ecological monitoring, noise interference will cause deviations in the analysis of the emitted acoustic signals, thereby affecting the correct assessment of the ecological status. Therefore, how to effectively reduce the noise of underwater acoustic echoes has become an important issue that needs to be solved in the field of underwater acoustics.

[0004] Over the years, many traditional noise reduction methods have been applied to underwater acoustic echo processing. Common methods include filtering-based methods, such as high-pass filters, low-pass filters, or band-pass filters, etc., which try to filter out noise components by setting a suitable frequency range. However, such methods are often only effective for noise of a specific frequency, and it is difficult to achieve ideal results for noise with complex frequency distribution and overlapping with the target signal spectrum. In addition, there are some noise reduction methods based on statistical models, such as the minimum mean square error algorithm (LMS) and adaptive filtering algorithms. Although they can suppress noise to a certain extent based on the statistical characteristics of signals and noise, they usually rely on accurate grasp of the prior knowledge of signals and noise. In the actual complex and changeable underwater environment, it is difficult to meet the needs of accurate noise reduction. Therefore, there is an urgent need for a more efficient and accurate underwater acoustic echo noise reduction technology solution. Summary of the invention

[0005] The main purpose of the present invention is to provide an underwater acoustic echo denoising method based on a wavelet convolution U-Net network, aiming to solve the problem that a more efficient and accurate underwater acoustic echo denoising technical solution is currently needed.

[0006] The technical solution proposed by the present invention is:

[0007] A method for underwater acoustic echo denoising based on a wavelet convolution U-Net network, comprising:

[0008] Acquire an underwater acoustic echo signal, and extract the time domain features of the underwater acoustic echo signal;

[0009] Construct a U-net underwater acoustic echo noise reduction network with a feature channel attention mechanism. Among them, the basic structure of the U-net is composed of one-dimensional wavelet convolution blocks. The encoder and decoder have a symmetric structure of twelve layers, which are responsible for downsampling and upsampling the underwater acoustic echo features respectively;

[0010] Construct a one-dimensional wavelet convolution block of the U-net underwater acoustic echo noise reduction network. Among them, the wavelet convolution block includes a frequency band separation layer, a convolution scaling layer, a deconvolution reconstruction layer, a weight processing layer, and a convolution processing output layer;

[0011] Construct a cross-channel cross-attention module of the U-net underwater acoustic echo noise reduction network. Among them, the cross-channel cross-attention module includes multi-layer perceptron layers, an average pooling layer, a linear transformation layer, an activation function, and a feature fusion layer with two branches;

[0012] Perform noise reduction processing on the underwater acoustic echo signal through the U-net underwater acoustic echo noise reduction network and the time-domain features of the underwater acoustic echo signal.

[0013] Preferably, the obtaining of the underwater acoustic echo signal and the extraction of the time-domain features of the underwater acoustic echo signal include:

[0014] Determine the linear sum of the mixed speech signal:

[0015] S(x) = v(x) + n(x),

[0016] where S(x) is the linear sum of the mixed speech signal; v(x) represents the time-domain vector of the clean speech; n(x) represents the time-domain vector of the noise; S(x), v(x), and n(x) are all vectors of R×1, and R represents the number of templates.

[0017] Preferably, after determining the linear sum of the mixed speech signal, it further includes:

[0018] Use the neural network model G to directly estimate the approximate clean speech v as Among them, the mapping relationship of the neural network model G is:

[0019]

[0020] In the frame-level processing process, divide the noise signal S into a matrix F composed of overlapping frames.

[0021] Preferably, after dividing the noise signal S into a matrix F composed of overlapping frames in the frame-level processing process, it further includes:

[0022] Through the vector f t represent t th frame, where f tThe frame expression is:

[0023]

[0024] In the formula, F is an N×L matrix; f is an L×1 vector, and N is the number of frames filled with zeros, L is the frame length, and J is the frame shift; The definition of is: The definition of is:

[0025]

[0026] Preferably, the one-dimensional wavelet convolution block for constructing the U-net underwater acoustic echo noise reduction network includes:

[0027] The one-dimensional wavelet convolution outputs data that combines the characteristics of wavelet transform and traditional convolution features, including:

[0028] Perform initialization: Assign the input signal to To represent the low-frequency part of the signal:

[0029]

[0030] Perform convolution: Convolve the signal with the kernel w (0) to obtain the initial low-frequency output

[0031] Preferably, after performing the convolution, it further includes:

[0032] Perform iteration: Decompose the signal using wavelet decomposition to generate a low-frequency component and a high-frequency component

[0033]

[0034]

[0035] Perform wavelet synthesis on the convolved low-frequency component and high-frequency component using the inverse wavelet to synthesize the output signal:

[0036]

[0037] Add the initial low-frequency convolution result to the result of the inverse wavelet transform to obtain the final output:

[0038]

[0039] Preferably, the cross-channel cross-attention module for constructing the U-net underwater acoustic echo noise reduction network includes:

[0040] Input preparation: Use the feature map E of the i-th level encoder i ∈R C×H×W and the feature map D of the i-th level decoder i ∈R C×H×W as the input for cross-channel attention;

[0041] Spatially compress E i and D i respectively through global average pooling layers to generate a vector G(E i )∈R C×1×1 , where the k-th channel of G(X) is:

[0042]

[0043] Preferably, the spatially compressing E i and D i respectively through global average pooling layers to generate a vector G(E i )∈R C×1×1 is followed by:

[0044] Attention mask generation: Calculate the attention mask M i =L 1 ·G(E i )+L 2 ·G(D i ), where L 1 =R C×C and L 2 =R C×C are the weights of two linear layers;

[0045] Construct a channel attention map: Use a single linear layer and an activation function to construct a channel attention map.

[0046] Preferably, after constructing the channel attention map, it is followed by:

[0047] Re-calibrate or stimulate features: Use the obtained vector to re-calibrate or stimulate E i to Z i =σ(M i )·E i , where σ(M i ) represents the importance of the channel.

[0048] Preferably, after re-calibrating or stimulating features, it is followed by:

[0049] Connect the mask with the encoder features of the i-th level of the decoder to complete the feature fusion process.

[0050] Through the above technical solutions, the following beneficial effects can be achieved:

[0051] The underwater acoustic echo noise reduction method proposed by the present invention can more efficiently and accurately reduce the noise of underwater acoustic echoes; the existing technical solutions are not flexible enough in dealing with the frequency characteristics of underwater acoustic echoes. The wavelet convolution and attention mechanism proposed by the present invention endow the model with multi-resolution analysis ability, can better adapt to the time-varying characteristics of underwater acoustic echoes at different frequencies, and can more carefully analyze the underwater acoustic echoes at different frequencies and time scales. The U-Net architecture performs excellently in dense prediction tasks, and the present invention combines U-Net with wavelet transform. Compared with the traditional underwater acoustic echo method based on a single structure, the solution proposed by the present invention provides a more effective means for extracting and fusing features at different hierarchical levels, and can better mine the useful information in underwater acoustic echoes. Description of the Drawings

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0053] Figure 1 It is a flowchart of the steps of the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network proposed by the present invention;

[0054] Figure 2 It is a structural block diagram of an underwater acoustic echo noise reduction system of the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network proposed by the present invention;

[0055] Figure 3 It is a schematic diagram of a one-dimensional wavelet convolution module of the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network proposed by the present invention;

[0056] Figure 4 It is a schematic diagram of a cross-channel cross-attention module of the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network proposed by the present invention;

[0057] Figure 5 It is a schematic diagram of noisy signal data of the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network proposed by the present invention;

[0058] Figure 6 It is a schematic diagram of the output signal data after noise reduction processing of the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolution U-Net network proposed by the present invention. Detailed Embodiments

[0059] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] The present invention proposes an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network.

[0061] As shown in the Figure 1 accompanying drawings, in the first embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, this embodiment includes the following steps:

[0062] Step S110: Obtain an underwater acoustic echo signal and extract the time-domain characteristics of the underwater acoustic echo signal.

[0063] Step S120: Construct a U-net underwater acoustic echo noise reduction network with a feature channel attention mechanism. Among them, the basic structure of the U-net is composed of one-dimensional wavelet convolutional blocks, and the encoder and decoder have a symmetric structure of twelve layers, which are responsible for downsampling and upsampling of underwater acoustic echo features respectively.

[0064] Specifically, the adopted wavelet convolutional block decomposes the features of each layer of the network into different frequency levels and divides the overall features of each layer of the network into local features; U-Net is an architecture based on convolutional neural network (CNN); U-Net can significantly improve the performance of the model on small-sample data sets by introducing skip connections and a symmetric encoder-decoder structure. Currently, U-Net and its variants have become one of the preferred methods for image segmentation in computer vision tasks.

[0065] Step S130: Construct a one-dimensional wavelet convolutional block of the U-net underwater acoustic echo noise reduction network. Among them, the wavelet convolutional block includes a band separation layer, a convolutional scaling layer, a deconvolution reconstruction layer, a weight processing layer, and a convolutional processing output layer.

[0066] Specifically, as shown in the Figure 3 accompanying drawings, by designing a one-dimensional wavelet convolutional block, which includes a band separation layer, a convolutional scaling layer, a deconvolution reconstruction layer, a weight processing layer, and a convolutional processing output layer, the forward propagation process of the entire wavelet convolutional module is completed, and feature data that combines the characteristics of wavelet transform and the feature extraction ability of traditional convolution is output.

[0067] Step S140: Construct a cross-channel cross-attention module of the U-net underwater acoustic echo noise reduction network. Among them, the cross-channel cross-attention module includes multi-layer perceptron layers, an average pooling layer, a linear transformation layer, an activation function, and a feature fusion layer with two branches.

[0068] Specifically, as shown in the Figure 4As shown, the cross-channel cross-attention module is used to adaptively adjust the importance of features in different channels, enabling the network to better utilize the multi-scale information extracted by the encoder during the process of decoding and restoring features, improving the feature representation ability of the entire network for input data and the final task processing performance.

[0069] Step S150: Denoise the underwater acoustic echo signal through the U-net underwater acoustic echo denoising network and the time-domain features of the underwater acoustic echo signal.

[0070] The underwater acoustic echo denoising method proposed by the present invention can more efficiently and accurately denoise underwater acoustic echoes; the existing technical solutions are not flexible enough in dealing with the frequency characteristics of underwater acoustic echoes. The wavelet transform proposed by the present invention endows the model with multi-resolution analysis ability, can better adapt to the time-varying characteristics of underwater acoustic echoes at different frequencies, and can more carefully analyze underwater acoustic echoes at different frequencies and time scales. The U-Net architecture performs well in dense prediction tasks, and the present invention combines U-Net with wavelet transform. Compared with the traditional underwater acoustic echo method based on a single structure, the solution proposed by the present invention provides a more effective means for extracting and fusing features at different hierarchical levels, and can better mine useful information in underwater acoustic echoes.

[0071] In addition, the existing models cannot effectively distinguish the importance of different channels in underwater acoustic echoes. The channel attention mechanism proposed by the present invention can adaptively recalibrate the feature responses in the channel direction, highlight important channels, and suppress less relevant channels. Enable the model to focus on the most informative part of the underwater acoustic echo, improve the ability to capture key features of the underwater acoustic echo, and thus enhance the enhancement effect. Compared with the prior art, the feature attention mechanism of the present invention can capture long-range dependencies in the time-frequency domain of underwater acoustic echoes. Enable the model to focus on specific regions or patterns crucial for underwater acoustic echo denoising, suitable for the denoising task of underwater acoustic echoes in non-stationary environments.

[0072] As shown in the appendix Figure 2 As shown, the present invention also proposes an underwater acoustic echo denoising system based on a wavelet convolutional UNet network; as shown in the appendix Figure 5 and the appendix Figure 6 As shown, by using a wavelet convolutional layer with a channel feature attention mechanism, fusing the multi-scale wavelet component features of different channels, and effectively connecting the reconstructed features to the decoder, the present invention effectively improves the underwater acoustic echo denoising effect.

[0073] In the second embodiment of an underwater acoustic echo denoising method based on a wavelet convolutional U-Net network proposed by the present invention, based on the first embodiment, step S110 includes the following steps:

[0074] Step S210: Determine the linear sum of the mixed speech signals:

[0075] S(x) = v(x) + n(x),

[0076] wherein, S(x) is the linear sum of the mixed speech signals; v(x) represents the time-domain vector of the clean speech; n(x) represents the time-domain vector of the noise; S(x), v(x), and n(x) are all vectors of R×1, and R represents the number of templates.

[0077] In the third embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the second embodiment, after step S210, the following steps are further included:

[0078] Step S310: Use the neural network model G to directly estimate the approximate clean speech v as wherein, the mapping relationship of the neural network model G is:

[0079]

[0080] Specifically, different from the masking-based inverse transform and feature map-based deconvolution methods for signal estimation, this embodiment uses the neural network model G to directly estimate the approximate clean speech v as

[0081] Step S320: During the frame-level processing, divide the noise signal S into a matrix F composed of overlapping frames.

[0082] In the fourth embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the third embodiment, after step S320, the following steps are further included:

[0083] Step S410: Represent the t t frame by the vector f th wherein, the frame expression of f t is:

[0084]

[0085] wherein, F is an N×L matrix; f is an L×1 vector, N is the number of frames filled with zeros, L is the frame length, and J is the frame shift; The definition of is:

[0086]

[0087] Specifically, this embodiment shows that is predicted from n 1 frames before the current frame to n 2 frames after it.

[0088] In the fifth embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the fourth embodiment, step S130 includes the following steps:

[0089] Step S510: Output data that fuses the characteristics of wavelet transform and traditional convolution features through one-dimensional wavelet convolution, including the following steps:

[0090] Step S511: Perform initialization: Assign the input signal to to represent the low-frequency part of the signal (initially the original signal):

[0091]

[0092] Step S512: Perform convolution: Convolve the signal with the kernel w (0) to obtain the initial low-frequency output

[0093]

[0094] In the sixth embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the fifth embodiment, after step S512, the following steps are further included:

[0095] Step S610: Perform iteration: Decompose the signal by wavelet to generate a low-frequency component and a high-frequency component

[0096]

[0097] Step S620: Use inverse wavelet to perform wavelet synthesis on the convolved low-frequency component and high-frequency component to synthesize the output signal:

[0098]

[0099] Step S630: Add the initial low-frequency convolution result to the result of the inverse wavelet transform to obtain the final output:

[0100]

[0101] In the seventh embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the sixth embodiment, step S140 includes the following steps:

[0102] Step S710: Input preparation: The i-th level encoder feature map E i ∈R C×H×W and the i-th level decoder feature map Di ∈R C×H×W As the input of the cross-attention between channels.

[0103] Step S720: Respectively perform spatial compression on E i and D i to generate a vector G(E i ) ∈ R C×1×1 , where the k-th channel of G(X) is:

[0104]

[0105] Specifically, this operation is used to embed global spatial information.

[0106] In the eighth embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the seventh embodiment, after step S720, the following steps are further included:

[0107] Step S810: Attention mask generation: Calculate the attention mask M i = L 1 · G(E i ) + L 2 · G(D i ), where L 1 = R C×C and L 2 = R C×C are the weights of two linear layers.

[0108] Step S820: Construct a channel attention map: Use a single linear layer and an activation function to construct a channel attention map.

[0109] In the ninth embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the eighth embodiment, after step S820, the following steps are further included:

[0110] Step S910: Re-calibrate or excite features: Use the obtained vector to re-calibrate or excite E i to Z i = σ(M i ) · E i , where σ(M i ) represents the importance of the channel.

[0111] In the tenth embodiment of an underwater acoustic echo noise reduction method based on a wavelet convolutional U-Net network proposed by the present invention, based on the ninth embodiment, after step S910, the following steps are further included:

[0112] Step S1010: Connect the mask with the encoder features of the i-th level of the decoder to complete the feature fusion process, thereby effectively reducing semantic ambiguity and improving the segmentation performance of the model.

[0113] The serial numbers of the embodiments of the present invention above are only for description and do not represent the advantages or disadvantages of the embodiments.

[0114] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims. All of these fall within the protection scope of the present invention.

Claims

1. A method for underwater acoustic echo denoising based on wavelet convolution U-Net network, characterized in that: include: Acquire an underwater acoustic echo signal, and extract the time domain features of the underwater acoustic echo signal; Construct a U-net underwater acoustic echo denoising network with wavelet convolution and feature channel attention mechanism. The basic structure of U-net is composed of one-dimensional wavelet convolution blocks. The encoder and decoder have a twelve-layer symmetrical structure, which are responsible for downsampling and upsampling of underwater acoustic echo features respectively. Construct a one-dimensional wavelet convolution block of the U-net underwater acoustic echo denoising network, wherein the wavelet convolution block includes a frequency band separation layer, a convolution scaling layer, a deconvolution reconstruction layer, a weight processing layer and a convolution processing output layer; Construct a cross-channel cross-attention module of the U-net underwater acoustic echo denoising network, where the cross-channel cross-attention module includes a two-branch multi-layer perceptron layer, an average pooling layer, a linear transformation layer, an activation function and a feature fusion layer; The underwater acoustic echo signal is denoised through the U-net underwater acoustic echo denoising network and the time domain characteristics of the underwater acoustic echo signal.

2. According to claim 1, a method for underwater acoustic echo denoising based on wavelet convolution U-Net network is characterized in that: The step of obtaining the underwater acoustic echo signal and extracting the time domain features of the underwater acoustic echo signal includes: Determine the linear sum of mixed speech signals: S(x)=v(x)+n(x), Where S(x) is the linear sum of the mixed speech signals; v(x) represents the time domain vector of the clean speech; n(x) represents the time domain vector of the noise; S(x), v(x) and n(x) are all R×1 vectors, where R represents the number of samples.

3. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 2 is characterized in that: The step of determining the linear sum of the mixed speech signals further comprises: The neural network model G is used to directly estimate the approximate clean speech v as Among them, the mapping relationship of the neural network model G is: During frame-level processing, the noise signal S is divided into a matrix F consisting of overlapping frames.

4. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 3 is characterized in that: In the frame-level processing, the noise signal S is divided into a matrix F consisting of overlapping frames, and then further includes: Through the vector f t Indicates t th frame, where f t The frame expression is: Where F is an N×L matrix; f is an L×1 vector, and N is The number of frames filled with zeros, L is the frame length, and J is the frame shift; is defined as:

5. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 4 is characterized in that: The one-dimensional wavelet convolution block for constructing the U-net underwater acoustic echo denoising network includes: The one-dimensional wavelet convolution output combines the wavelet transform characteristics with the traditional convolution feature data, including: Initialize: assign the input signal to To represent the low-frequency part of the signal: Perform convolution: transform the signal With the core w (0) Convolution is performed to obtain the initial low-frequency output 6. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 5 is characterized in that: The performing of convolution further comprises: Iterate: The signal Perform wavelet decomposition to generate low-frequency components and high frequency components The low-frequency component and high-frequency component after convolution are synthesized using inverse wavelet to synthesize the output signal: The initial low-frequency convolution result is added to the result of the inverse wavelet transform to get the final output:

7. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 6 is characterized in that: The cross-channel cross-attention module for constructing the U-net underwater acoustic echo denoising network includes: Input preparation: The i-th level encoder feature map E i ∈R C×H×W and the i-th level decoder feature map D i ∈R C×H×W As input for cross-channel attention; Through the global average pooling layer, E i and D i Perform space compression to generate vector G(E i )∈R C×1×1 , where the kth channel of G(X) is:

8. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 7 is characterized in that: The global average pooling layer is used to i and D i Perform space compression to generate vector G(E i )∈R C×1×1 , and later: Attention mask generation: Calculate the attention mask M i =L1·G(E i )+L2·G(D i ), where L1 = R C×C and L2=R C×C are the weights of the two linear layers; Constructing channel attention map: Construct a channel attention map using a single linear layer and activation function.

9. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 8 is characterized in that: The constructing of the channel attention map further includes: Recalibrate or stimulate features: The resulting vector is used to recalibrate or stimulate E i to Z i =σ(M i )·E i , where σ(M i ) indicates the importance of the channel.

10. The underwater acoustic echo denoising method based on wavelet convolution U-Net network according to claim 9 is characterized in that: The recalibration or stimulation feature then further comprises: The mask is concatenated with the encoder features at the i-th level of the decoder to complete the feature fusion process.

Images