Low-coherence interference signal denoising method and system based on attention mechanism

Abstract

The present invention discloses a method and system for denoising low-coherence interference signals based on an attention mechanism. The method comprises: S1, preprocessing the low-coherence signal to construct a low-coherence signal training set containing noise; S2, constructing a deep one-dimensional convolutional denoising neural network consisting of a residual module, a context information extraction module, and a self-attention module; S3, inputting the training data into the trained network to eliminate noise from the signal; and S4, outputting a clean low-coherence signal. The system includes a low-coherence measurement instrument, a data acquisition card, a computing server, an interactive control device, and a display screen. The proposed method and system for denoising low-coherence signals based on context information and self-attention have the advantages of simple parameter adjustment and effective signal noise elimination.

Description

Technical Field

[0001] The present invention belongs to the technical field of low-coherence signal processing, and in particular relates to a low-coherence signal denoising method and system based on contextual information and self-attention. Background Art

[0002] Low-coherence signals are collected by low-coherence interferometers. Low-coherence interferometers are commonly used in optical coherence tomography, surface topography, roughness measurement, mirror flatness, and lens spacing measurement. The presence of noise in low-coherence signals can significantly impact the measurement accuracy and imaging quality of low-coherence interferometers. As precision instruments, low-coherence interferometers are subject to external factors such as ambient air vibration and temperature fluctuations, generating random noise. This noise is characterized by low frequency, high energy, and strong randomness.

[0003] Therefore, to obtain high-quality images and high-precision measurements, the noise in low-coherence signals must first be denoised, while minimizing the risk of damaging the original signal. Traditional low-coherence signal denoising methods include wavelet transform, Fourier transform, and polynomial fitting. However, these methods require complex parameter adjustment and extensive engineering experience. Furthermore, due to the high randomness of the noise, the noise removal effect of these methods is not ideal. Summary of the Invention

[0004] In order to overcome the limitations of the above-mentioned traditional denoising methods, the present invention provides a low-coherence interference signal denoising method and system based on the attention mechanism, which has the advantages of simple parameter adjustment and good signal noise elimination effect, and can realize the denoising of low-coherence signals.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a low-coherence interference signal denoising method based on an attention mechanism, comprising:

[0006] S1. Normalize the low-coherence signal; construct a low-coherence signal dataset containing noise;

[0007] S2. Construct a deep one-dimensional convolutional denoising neural network;

[0008] S3. Use the training set data to train the deep one-dimensional convolutional denoising neural network, the validation set data to verify whether the network converges, and the test set data to test the noise signal elimination effect;

[0009] S4. Use the trained one-dimensional convolutional denoising neural network to process the noisy signal to output a low-coherence signal.

[0010] The preprocessing in S1 specifically includes:

[0011] S11. Collect and normalize low-coherence signals: Use a low-coherence interferometer to collect a data set X containing M low-coherence signal sequences. m ,…,X M}, where the signal sequence X m ={x m1 ,…,x mN}, for each low coherence signal sequence X m Normalize according to equation (1):

[0012]

[0013] In the formula For each low coherence signal X m Sample data x mi The maximum value of For each low coherence signal X m Sample data x mi The minimum value of , where i≤i≤N, N is the number of low coherence signal sampling points.

[0014] The noise added in S1 includes:

[0015] S12. Constructing noisy low-coherence signal samples: Constructing Gaussian distributed white noise n; adding noise of different proportions to the M low-coherence signals in the data set X according to equation (2):

[0016] y mi =z mi +αn (2)

[0017] Get the low coherence signal sequence with white noise added: Y m ={y m1 ,…y mi …,y mN};

[0018] Get a data set of M low-coherence signal sequences with white noise added: Y = {Y1, ...Y i …,Y N};

[0019] Among them, α is the parameter that controls the proportion of noise added, y mi is a low coherent signal with white noise added.

[0020] The data sets divided in S1 include:

[0021] S13. Divide the neural network into training set, validation set, and test set: evenly mix the dataset Y of the low coherent signal sequence after adding Gaussian white noise and the dataset X of the low coherent signal sequence without adding noise, and divide them into training set, validation set, and test set according to the ratio of 5:3:2.

[0022] S2 specifically includes:

[0023] S21. Establish a network model, including: a feature extraction module, a context feature enhancement module, a self-attention module, and a residual module; the feature extraction module combines with the context feature enhancement module to process the original low-coherence signal it receives and extract nonlinear signal features;

[0024] S25, the self-attention module is used to further enhance the noise extraction capability of the network for nonlinear features and output the extracted noise signal;

[0025] S26, the residual module is used to subtract the extracted noise signal from the original low-coherence signal to obtain a clean low-coherence signal.

[0026] The feature extraction module processes the original low-coherence signal received by the feature extraction module in combination with the context feature enhancement module, including:

[0027] S22. The feature extraction module includes the 1st, 2nd, 3rd, 4th, 5th, 6th, 7th, 8th, 9th, 10th and 11th layers of convolutional neural networks; wherein the operations of the 1st, 3rd, 5th, 6th, 8th and 10th layers include convolution, batch normalization and activation; wherein the convolution adopts one-dimensional convolution, the convolution kernel size is 1×3, and the activation adopts a linear rectification function with a step size of 2; wherein the operations of the 2nd, 7th and 9th layers include dilated convolution, batch normalization and activation; wherein the dilation factor of the dilated convolution is 2, the convolution kernel size is 1×3, and the activation function is a linear rectification function with a step size of 2; the 11th layer has only convolution operations, and the convolution adopts one-dimensional convolution with a convolution kernel size of 1×3.

[0028] S23. The features output by the 5th layer network are sent to the context feature enhancement module to enhance the network's ability to obtain global context information. After feature enhancement, the context feature enhancement module inputs the enhanced features into the 6th layer network, and they are processed by the subsequent 7th, 8th, 9th, 10th, and 11th layer networks.

[0029] The features output by the S24 and 11 layers of the network are fused with the original noisy low-coherence signal through a cascade operation to further enhance the network's feature extraction capability. The fused features are then converted into nonlinear features using the hyperbolic tangent (Tanh) activation function and normalized at the same time.

[0030] S3 specifically includes:

[0031] S31. Use the training set to train a deep one-dimensional convolutional denoising neural network, use the mean square error as the loss function, and use the back propagation algorithm to train the network parameters;

[0032] S32. During the training process, use the validation set to check whether the model has converged;

[0033] S33. After the model converges, the test set is used to evaluate the network's ability to eliminate noise signals and the training is ended.

[0034] S4 specifically includes:

[0035] S41. Input the noisy low-coherence signal into the deep one-dimensional convolutional denoising neural network trained in S3 to obtain a clean low-coherence signal.

[0036] A system for the low-coherence signal denoising method based on self-supervised learning according to claim 1, the system comprising:

[0037] Signal acquisition equipment, collecting low-coherence signals;

[0038] A data acquisition card, used for receiving low-coherence signals collected by a low-coherence measuring instrument;

[0039] A computing server equipped with a program module, which, when executed, processes the collected low-coherence signal according to the method and steps described above, implements low-coherence signal denoising, and outputs the processing result to a display screen;

[0040] An interactive control device for inputting instructions to the computing server;

[0041] A display screen is used to visually display the low-coherence signal before and after denoising, as well as a graphical interface during the processing.

[0042] The signal acquisition device is a low-coherence measuring instrument.

[0043] The present invention has the following beneficial effects and advantages:

[0044] 1. The deep one-dimensional convolutional denoising neural network designed in the present invention includes modules such as convolutional layers connected in series, context feature enhancement, self-attention noise extraction enhancement, and residual noise elimination, which has the advantage of strong effect in removing low-coherence signal noise.

[0045] 2. The deep one-dimensional convolutional denoising neural network designed in the present invention has a simple parameter adjustment process. After training is completed, it can denoise a variety of low-coherence signals without the need to adjust the parameters again for different low-coherence signals.

[0046] 3. This invention provides a new approach to noise processing of low-coherence signals using deep learning technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flowchart of the denoising method of the present invention;

[0048] Figure 2 The structure diagram of the deep one-dimensional convolutional denoising neural network of the present invention;

[0049] Figure 3 This is a block diagram of the system modules of the present invention. DETAILED DESCRIPTION

[0050] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, the following detailed description of the specific implementation of the present invention is provided with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the invention. Therefore, the present invention is not limited to the specific implementation disclosed below. Specific embodiment:

[0052] Please refer to Figure 1 The present invention provides a low-coherence interference signal denoising method based on an attention mechanism, comprising:

[0053] S1: Preprocess the low-coherence signal and construct a low-coherence signal dataset containing noise.

[0054] S11. Use the Soreb low coherence interferometer TEL211C1 to collect a data set X={X1,...X1} containing M=10000 low coherence signal sequences. m ,…,X 10000}, where the low coherence signal sequence X m ={x m1 ,…x mi …,x mN}, for each low coherence signal sequence X m Normalize according to equation (1):

[0055]

[0056] In the formula For each low coherence signal X m Sample data x mi The maximum value of For each low coherence signal X m Sample data x mi The minimum value of , where i≤i≤N, N is the number of low coherence signal sampling points, and the number of sampling points N is 100000.

[0057] S12. Construct noisy low-coherence signal samples: construct white noise n with Gaussian distribution; add noise of different proportions to the 10,000 low-coherence signals in the data set X according to equation (2):

[0058] y mi =z mi +αn (2)

[0059] Get the low coherence signal sequence with white noise added: Y m ={y m1 ,…y mi …,y mN};

[0060] Get a data set of M low-coherence signal sequences with white noise added: Y = {Y1, ...Y i …,Y N};

[0061] The white noise obeys the Gaussian distribution N(0,1), the value range of α is 0.00-0.2, the value interval is 0.01, and y mi is a low-coherence signal with white noise added. The final datasets X and Y contain 200,000 low-coherence signals with different signal-to-noise ratios.

[0062] S13. Divide the neural network into training set, validation set, and test set: uniformly mix the dataset Y of the low-coherence signal sequence after adding Gaussian white noise and the dataset X of the low-coherence signal sequence without adding noise, and divide the mixed signal into training set, validation set, and test set in the ratio of 50%, 30%, and 20%.

[0063] S2, please refer to Figure 2 , construct an 11-layer deep one-dimensional convolutional denoising neural network consisting of a residual module, a context information extraction module, and a self-attention module.

[0064] S22, the feature extraction module includes the 1st layer, the 2nd layer, the 3rd layer, the 4th layer, the 5th layer, the 6th layer, the 7th layer, the 8th layer, the 9th layer, the 10th layer, and the 11th layer of convolutional neural networks; wherein the operations of the 1st layer, the 3rd layer, the 5th layer, the 6th layer, the 8th layer, and the 10th layer include convolution, batch normalization (bath norm) and activation, wherein the convolution adopts one-dimensional convolution, the convolution kernel size is 1×3, and the activation adopts a linear rectification function (linear rectification function, ReLU) with a step size of 2; the operations of the 2nd layer, the 7th layer, and the 9th layer include dilated convolution, batch normalization and activation, wherein the dilation factor of the dilated convolution is 2, the convolution kernel size is 1×3, and the activation function ReLU is adopted with a step size of 2; the 11th layer has only convolution operations, and the convolution adopts one-dimensional convolution with a convolution kernel size of 1×3.

[0065] S23. The features output by the 5th layer network are sent to the context feature enhancement module to enhance the network's ability to obtain global context information. After feature enhancement, the context feature enhancement module inputs the enhanced features into the 6th layer network, and they are processed by the subsequent 7th, 8th, 9th, 10th, and 11th layer networks.

[0066] The features output by the S24 and 11 layers of the network are fused with the original low-coherence signal containing noise through a cascade operation (Cat) to further enhance the network's feature extraction capability. The fused features are then converted into nonlinear features using the hyperbolic tangent (Tanh) activation function and normalized at the same time.

[0067] S25. The normalized nonlinear features are fed into the self-attention module to further enhance the noise extraction capability of the network; the self-attention module outputs the noise signal extracted by the network.

[0068] S26. Finally, the residual module is used to subtract the extracted noise signal from the original low-coherence signal to obtain a clean low-coherence signal.

[0069] S3. Use the training set to train the deep one-dimensional convolutional denoising neural network, the validation set to verify whether the network converges, and the test set to test the noise signal elimination effect.

[0070] S31. Use the training set to train the deep one-dimensional convolutional denoising neural network, and use the validation set to verify the network's ability to eliminate noise during training.

[0071] S32. Use mean square error as the loss function and use the back propagation algorithm to train the network parameters. The specific definition of the loss function is:

[0072]

[0073] Where θ is the parameter of the deep one-dimensional convolutional denoising neural network. During the training process, the validation set is used to test whether the model has converged. After the model converges, the test set is used to evaluate the network's ability to eliminate noise signals and end the training.

[0074] S4. Use the trained one-dimensional convolutional denoising neural network to output a clean low-coherence signal.

[0075] S41. Input the noisy low-coherence signal into the deep one-dimensional convolutional denoising neural network trained in S3 to obtain a clean low-coherence signal.

[0076] Low-coherence interference signal denoising system based on attention mechanism, including hardware and software; such as Figure 3 As shown, the hardware includes a low-coherence measurement instrument ATR206C1, a data acquisition card ATS9350, a high-performance computing server Inspur NF5468M5, an interactive control device (mouse, keyboard, etc.), and a high-definition display screen T3252U; the low-coherence signal collected by the low-coherence measurement instrument ATR206C1 is input into the high-performance computing server Inspur NF5468M5 through the data acquisition card ATS9350, and the high-performance computing server Inspur NF5468M5 is deployed with the Ubuntu 16.04 operating system; the high-performance computing server Inspur NF5468M5 controls the low-coherence measurement instrument ATR206C1 and the data acquisition card ATS9350 through the interactive control device (keyboard, mouse, etc.); the high-definition display screen T3252U is used to display a graphical interface;

[0077] The software includes a graphical interface, the deep learning framework TensorFlow 1.12.0, an Ubuntu 16.04 operating system, and a denoising module. The deep learning framework TensorFlow 1.12.0 is deployed on the Ubuntu 16.04 operating system, and the denoising module is implemented within this framework. Users operate the graphical interface through an interactive control device and control the high-performance Inspur NF5468M5 server through the Ubuntu 16.04 operating system. The denoising module contains the denoising method and program steps for signal denoising.

[0078] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A low-coherence interference signal denoising method based on attention mechanism, characterized by: include: S1, normalize the low coherence signal; Construct a low-coherence signal dataset containing noise; S2. Construct a deep one-dimensional convolutional denoising neural network; including: S21. Establish a network model, including: a feature extraction module, a context feature enhancement module, a self-attention module, and a residual module; the feature extraction module combines with the context feature enhancement module to process the original low-coherence signal it receives and extract nonlinear signal features; S22, the feature extraction module includes a 1st layer, a 2nd layer, a 3rd layer, a 4th layer, a 5th layer, a 6th layer, a 7th layer, an 8th layer, a 9th layer, a 10th layer, and an 11th layer of convolutional neural networks; wherein the operations of the 1st layer, the 3rd layer, the 5th layer, the 6th layer, the 8th layer, and the 10th layer include convolution, batch normalization, and activation; the operations of the 2nd layer, the 7th layer, and the 9th layer include dilated convolution, batch normalization, and activation; and the 11th layer has only convolution operations; S23: The features output by the 5th layer network are fed into the context feature enhancement module to enhance the network's ability to obtain global context information. After feature enhancement, the context feature enhancement module feeds the enhanced features into the 6th layer network, which is then processed by the subsequent 7th, 8th, 9th, 10th, and 11th layers of the network. The features output by the S24 and 11 layers of the network are fused with the original low-coherence signal containing noise through a cascade operation to further enhance the network's feature extraction capability. The fused features are then converted into nonlinear features using the hyperbolic tangent (Tanh) activation function and normalized at the same time. S25, the self-attention module is used to further enhance the noise extraction capability of the network for nonlinear features and output the extracted noise signal; S26, the residual module is used to subtract the extracted noise signal from the original low-coherence signal to obtain a clean low-coherence signal; S3. Use the training set data to train the deep one-dimensional convolutional denoising neural network, the validation set data to verify whether the network converges, and the test set data to test the noise signal elimination effect; including: S31. Use the training set to train a deep one-dimensional convolutional denoising neural network, use the mean square error as the loss function, and use the back propagation algorithm to train the network parameters; S32. During the training process, use the validation set to check whether the model has converged; S33. After the model converges, the test set is used to evaluate the network's ability to eliminate noise signals and the training ends. S4. Use the trained one-dimensional convolutional denoising neural network to process the noisy signal to be processed and output a low-coherence signal.

2. The low-coherence interference signal denoising method based on the attention mechanism according to claim 1, characterized in that: The preprocessing in S1 specifically includes: S11. Collect and normalize low-coherence signals: Use a low-coherence interferometer to collect a data set X containing M low-coherence signal sequences. m ,…,X M }, where the signal sequence X m ={x m1 ,…,x mN }, for each low coherence signal sequence X m Normalize according to equation (1): In the formula For each low coherence signal X m Sample data x mi The maximum value of For each low coherence signal X m Sample data x mi The minimum value of , where i≤i≤N, N is the number of low coherence signal sampling points.

3. The low-coherence interference signal denoising method based on the attention mechanism according to claim 1, characterized in that: The noise added in S1 includes: S12. Constructing noisy low-coherence signal samples: Constructing Gaussian distributed white noise n; adding noise of different proportions to the M low-coherence signals in the data set X according to equation (2): y mi =z mi +αn (2) Get the low coherence signal sequence with white noise added: Y m ={y m1 ,…y mi …,y mN }; Get a data set of M low-coherence signal sequences with white noise added: Y = {Y1, ...Y i …,Y N }; Among them, α is the parameter that controls the proportion of noise added, y mi is a low coherent signal with white noise added.

4. The low-coherence interference signal denoising method based on the attention mechanism according to claim 1, characterized in that: The data sets divided in S1 include: S13. Divide the neural network into training set, validation set, and test set: evenly mix the dataset Y of the low coherent signal sequence after adding Gaussian white noise and the dataset X of the low coherent signal sequence without adding noise, and divide them into training set, validation set, and test set according to the ratio of 5:3:

2.

5. The low-coherence interference signal denoising method based on the attention mechanism according to claim 1 is characterized in that S4 specifically includes: S41. Input the noisy low-coherence signal into the deep one-dimensional convolutional denoising neural network trained in S3 to obtain a clean low-coherence signal.

6. A system for the low-coherence signal denoising method based on self-supervised learning according to claim 1, characterized in that: The system comprises: Signal acquisition equipment, collecting low-coherence signals; A data acquisition card, used for receiving low-coherence signals collected by a low-coherence measuring instrument; A computing server equipped with a program module, which, when the program is executed, is used to process the collected low-coherence signal according to the method steps described in any one of claims 1 to 5, to implement low-coherence signal denoising, and output the processing results to a display screen; An interactive control device for inputting instructions to the computing server; A display screen is used to visually display the low-coherence signal before and after denoising, as well as a graphical interface during the processing.

7. A system for the low-coherence signal denoising method based on self-supervised learning according to claim 6, characterized in that: The signal acquisition device is a low-coherence measuring instrument.

Images