Adaptive convolution denoising network near-field microwave signal random noise suppression method

Abstract

The application discloses a kind of based on adaptive convolution denoising network near-field microwave signal random noise suppression method, belong to near-field microwave signal random noise suppression field.The adaptive convolution denoising network model of the present application utilizes noise level estimation and convolution denoising two sub-networks, uses asymmetric learning loss function, enhances the effectiveness and accuracy of denoising result;Network training adopts model noise and actual noise two types of data alternately training, improves the denoising effect and generalization ability of network model;Combining the noise estimation method based on weak signal block carries out noise level estimation, improves the reliability of actual data training, can realize the adaptive denoising of near-field microwave data in the case where unknown noise standard deviation;The application can significantly improve noise suppression capability, adaptively attenuate random noise in data, protect effective signal, and can obtain relatively ideal denoising effect for near-field microwave signal.

Description

Technical Field

[0001] This invention belongs to the field of random noise suppression of near-field microwave signals, and specifically relates to a method for random noise suppression of near-field microwave signals using an adaptive convolutional denoising network. Background Technology

[0002] Random noise is caused by various interference factors. During microwave near-field testing and measurement, the acquired microwave signal is inevitably affected by various environmental factors. The random noise generated by these factors is received by the receiver along with the valid signal. The presence of this random noise interferes with the test results, masks some weak valid signals, and reduces the accuracy of the test results. Therefore, it is necessary to suppress random noise in the acquired signal to improve the accuracy of the test and measurement results.

[0003] With the rapid development of computer hardware, especially graphics processing units, deep neural networks have been a hot topic since 2010, including deep belief networks, stacked autoencoders, and deep convolutional neural networks. Deep convolutional neural networks, by utilizing the correlation of local convolutional filters, avoid using a large number of parameters and fully extract the structural features of data while preserving its local structure. Deep convolutional neural networks have achieved success in solving highly nonlinear computer vision problems in various fields. Residual convolutional neural networks use CNN networks with 17 convolutional layers for image denoising. Utilizing the idea of ​​residual learning, they use noise as output, which can accelerate the training process and improve denoising performance. Advanced feature extraction algorithms have recently achieved excellent results in image segmentation and super-resolution image processing. Deep learning algorithms have deeper levels of data feature decomposition and extraction, capable of calculating the deep structure of data during training. Compared with traditional single-layer decomposition and feature extraction methods such as dictionary learning, they can more effectively preserve the structural features of data.

[0004] Prior to the implementation of this invention, there was no domestic method for denoising near-field microwave signals using random noise suppression based on adaptive convolutional denoising networks. Conventional data noise suppression methods, such as block matching denoising algorithms and sparse dictionary filtering methods, while effective for data with high noise standard deviations, are insensitive to data with low noise standard deviations. Their performance on low-standard-deviation noise data is less than ideal, especially when processing near-field microwave signals. These methods struggle to balance noise suppression with signal preservation, making them unsuitable for practical data processing.

[0005] Traditional random noise attenuation methods are typically based on filtering techniques, generally assuming a Gaussian distribution for noise. Although many methods for random noise suppression have been proposed, two problems remain in practical applications: incorrect noise assumptions and inaccurate parameter settings. The model data is only an approximation of the actual data and cannot accurately represent the actual acquired microwave signal data. For example, sparse transform-based methods assume that a specially designed transform can be used to sparsely represent the actual data, but this assumption is usually invalid for real-world data. Adaptive dictionary learning methods train adaptive sparse transforms from the dataset. Although they do not use pre-designed transforms, dictionary learning still relies on the sparsity assumption. Furthermore, to achieve good denoising results, repeated testing and fine-tuning of denoising parameters based on experience is necessary, making it difficult to guarantee the use of suitable denoising parameters for specific data, thus affecting the denoising effect. Conventional data noise suppression methods, such as block matching denoising algorithms and sparse dictionary filtering methods, are effective in processing data with high noise standard deviations, but they are not sensitive to data with low noise standard deviations. Their performance is not ideal when processing low standard deviation noise data, especially when processing near-field microwave signals. These methods are difficult to balance suppressing noise and protecting the effective signal, and therefore cannot be applied to data processing.

[0006] To address the aforementioned issues, this invention employs a random noise suppression method for near-field microwave signals based on an adaptive convolutional denoising network. This method utilizes a noise standard deviation estimation network and a convolutional denoising network, combined with a noise estimation method based on weak signal blocks, to estimate the noise level. This method can perform random noise denoising processing on the characteristics of microwave near-field detection and acquisition data, providing an effective denoising algorithm while protecting the effective signal from being suppressed. Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, this invention proposes a method for suppressing random noise in near-field microwave signals based on an adaptive convolutional denoising network. This method is rationally designed, overcomes the shortcomings of existing technologies, and achieves good results.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] An adaptive convolutional denoising network method for suppressing random noise in near-field microwave signals, comprising two sub-networks: a noise level estimation network and a convolutional denoising network, includes the following steps:

[0010] Step 1: Use the block matching algorithm to denoise the actual collected data, select the denoised data and noisy data with good denoising effect as training data, and build a data training library using two types of data: model noise and actual noise.

[0011] Step 2: Use a noise estimation method based on weak signal blocks to estimate the noise level of the training data, obtain the noise standard deviation, and then use the obtained data to train the noise level estimation network.

[0012] Step 3: Calculate the noise estimation standard deviation of the noisy data y using a noise level estimation network. The data and parameters are then input into a convolutional denoising network to train the denoising model, and the standard deviation is estimated using the noise. Adjust network training parameters;

[0013] Step 4: Use a convolutional denoising network to extract the noisy data y and estimate the standard deviation of the noise. and noise data As input, an asymmetric loss function is used to eliminate the underfitting and overfitting problems of the convolutional denoising network, and the final denoising model is obtained.

[0014] Step 5: Use the trained denoising model to denoise other data, achieving adaptive intelligent denoising based on the characteristics of microwave near-field data, while protecting the effective signal from being suppressed, and obtaining the final denoising result.

[0015] Preferably, in step 3, the noise level estimation network uses a five-layer convolutional network with a kernel size of 3×3×32, and employs batch normalization to accelerate training and improve the accuracy of the training model in estimating the standard deviation of data noise.

[0016] Preferably, in step 4, the convolutional denoising network uses a 16-layer U-Net network structure and employs residual learning to learn the residual data. In other words, the noise data is removed from the input noisy data to obtain the denoised result.

[0017] Preferably, in step 2, during network training, the noiseless model data is denoised to obtain noisy data, whose noise standard deviation is known. However, the noisy model data and the collected noisy actual data differ in terms of data characteristics. As for the collected actual data, the noise is caused by the collection environment, and it is impossible to obtain accurate actual noiseless data. Only the denoised data that is close to noiseless data can be used as a reference to calculate the noise standard deviation of the actual data in the training library.

[0018] Therefore, a noise estimation method based on weak signal blocks is used to calculate the actual noise standard deviation. This method utilizes probabilistic and statistical thinking, fitting the standard deviation of each segmented data block by statistically analyzing the data segmentation results. The noise standard deviation σ is obtained by decomposing the covariance matrix of the noisy data block to find the minimum eigenvalue.

[0019] β min (Σy )=β min (Σ s )+σ 2 (1);

[0020] In the formula, Σ y Σ represents the covariance matrix of the noisy data block. s Represents the covariance matrix of the noiseless data block, β min (Σ y ) represents the smallest eigenvalue of the covariance matrix of the noisy data block, β min (Σ s β is the smallest eigenvalue of the covariance matrix of the noiseless data block; for actual data, β min (Σ y The value of the weak signal block set is unknown, but it can be replaced by the smallest eigenvalue of the covariance matrix of the weak signal block set. The weak signal block set can be selected by the signal strength of the local gradient matrix of the data and its statistical properties.

[0021] Preferably, in step 4, the convolutional denoising network is asymmetrically sensitive to the error of the noise estimation result. An asymmetric loss function is used to eliminate this asymmetric sensitivity and suppress noise.

[0022] Estimated noise level at point i and actual value σ(y) i The objective loss function for constructing the entire adaptive convolutional denoising network is:

[0023] L = L r +λ a L a +λ T L T (2);

[0024] In the formula, λ a L a +λ T L T The asymmetric loss function for the noise level estimation network; L a Here is the loss function for noise suppression, used to ensure that noise can be sufficiently suppressed:

[0025]

[0026] when In order to fully suppress noise, a denoising weighting factor λ is used. a By setting the weight parameter α: 0 < α < 0.5, the denoising weights of data with different noise standard deviations can be balanced; L T To address the overfitting loss function and ensure sufficient preservation of the effective signal after denoising, a total variation regularization term is introduced to constrain it. Smoothness:

[0027]

[0028] Where, λ T Its weighting factor;

[0029] L r The loss function for the convolutional denoising network is defined using the L2 norm to define the reconstruction error:

[0030]

[0031] The beneficial technical effects of this invention are as follows:

[0032] Compared to traditional random noise suppression algorithms, deep learning-based denoising methods reduce reliance on data models and prior information. This invention utilizes an adaptive convolutional denoising network-based random noise suppression method for near-field microwave signals. Tailored to the characteristics of near-field microwave signals, the denoising model, after training, achieves adaptive intelligent denoising. The adaptive convolutional denoising network model employs two sub-networks: noise level estimation and convolutional denoising, using an asymmetric learning loss function to enhance the effectiveness and accuracy of the denoising results. Network training alternates between model noise and actual noise data, improving the denoising effect and generalization ability of the network model. Combining noise level estimation based on weak signal blocks improves the reliability of training on actual data, enabling adaptive denoising of near-field microwave data even with unknown noise standard deviations. Compared to traditional sparse dictionary filtering methods, this significantly improves noise suppression capabilities, adaptively attenuating random noise in the data, preserving discontinuities and boundary information, protecting the structural characteristics of the data, and achieving a more ideal denoising effect for near-field microwave signals. Attached Figure Description

[0033] Figure 1 Diagram of the adaptive convolutional denoising network structure;

[0034] Figure 2 This is a comparison chart of the data waveforms before and after noise reduction. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0036] The adaptive convolutional denoising network used in this invention comprises two sub-networks: a noise level estimation network and a convolutional denoising network. First, the noise level estimation network calculates the estimated noise standard deviation of the noisy data y. Then, the convolutional denoising network will combine y and The final denoising result is obtained by taking it as input. Noise level estimation networks can estimate the standard deviation of noise. The network training parameters were adjusted, and then the data and parameters were input into the convolutional denoising network for model training. The noise level estimation network used a five-layer convolutional network with a kernel size of 3×3×32, and batch normalization was applied to accelerate training and improve the accuracy of the trained model in estimating the standard deviation of data noise. The convolutional denoising network used a 16-layer U-Net network structure and employed residual learning to learn from residual data. In other words, noisy data can be obtained by removing noisy data from the input noisy data.

[0037] Convolutional denoising networks exhibit asymmetric sensitivity to noise estimation errors. The denoising effect is best when the standard deviation of the input noise matches the standard deviation of the actual noise. When the standard deviation of the input noise is lower than the actual value, noise suppression is incomplete, leaving residual noise in the denoised data. When the standard deviation of the input noise is higher than the actual value, although noise can be effectively suppressed, overfitting to the denoised data may occur. To eliminate this asymmetric sensitivity, an asymmetric loss function can be used for noise suppression.

[0038] Estimated noise level at point i and actual value σ(y) i The objective loss function for constructing the entire adaptive convolutional denoising network is:

[0039] L = L r +λ a L a +λ T L T (6);

[0040] In the formula λ a L a +λ T L T An asymmetric loss function for the noise level estimation network can balance adequate noise suppression and overfitting issues; L a The loss function for noise suppression ensures that noise can be sufficiently suppressed.

[0041]

[0042] when In order to effectively suppress noise, a larger denoising weighting factor λ is used. a By setting the weight parameter α: 0 < α < 0.5, the denoising weights for data with different noise standard deviations can be balanced; L T To address the overfitting loss function and ensure sufficient preservation of the effective signal after denoising, a total variation regularization term is introduced to constrain it. Smoothness:

[0043]

[0044] λ T It is its weighting factor.

[0045] L r The loss function for the convolutional denoising network is defined using the L2 norm to define the reconstruction error:

[0046]

[0047] During network training, noisy model data is processed by adding noise to obtain noisy data. The noise standard deviation of this noisy data is known, but there are differences in data characteristics between the noisy model data and the actual noisy data collected. For the actual collected data, the noise is caused by the collection environment, making it impossible to obtain accurate actual noise-free data. Only denoised data that approximates noise-free data can be used as a reference, and the actual noise standard deviation is also unknown. To improve the performance of the noise level estimation network, it is necessary to calculate the noise standard deviation of the actual data in the training library. Therefore, during model training, a noise estimation method based on weak signal blocks is used to calculate the actual noise standard deviation. This method utilizes probabilistic statistical thinking, fitting the standard deviation of each segmented data block by statistically analyzing the data segmentation results. The noise standard deviation σ is obtained by decomposing the covariance matrix of the noisy data block to find the minimum eigenvalue.

[0048] β min (Σ y )=β min (Σ s )+σ 2 (10);

[0049] In the formula Σ y Σ represents the covariance matrix of the noisy data block. s Represents the covariance matrix of the noiseless data block, β min (Σ y ) represents the smallest eigenvalue of the covariance matrix of the noisy data block, β min (Σ s β is the smallest eigenvalue of the covariance matrix of the noiseless data block. For real-world data, β... min (Σ y The value of the weak signal block set is unknown, but it can be replaced by the minimum eigenvalue of the covariance matrix of the weak signal block set. The weak signal block set can be selected by the signal strength of the local gradient matrix of the data and its statistical characteristics.

[0050] In training an adaptive convolutional denoising network, combining both model and real-world data can improve the generalization ability of the trained network model. During actual training, model data and real-world data are used alternately. The real-world training data is obtained by denoising near-field microwave data using other conventional methods. Multiple denoising methods are employed, and data with good denoising effects and significant improvements in the signal-to-noise ratio before and after denoising are selected from the denoising results to construct the database.

[0051] To verify the effectiveness of the algorithm of this invention, a denoising experiment was first conducted on the model data. The model data is a sinusoidal signal. Denoising processing was performed on the noisy model data after the noise was added. The processing result is as follows: Figure 2 As shown.

[0052] As can be seen from the waveforms of a single set of data, the training network structure used in this invention is more sensitive to detecting random noise in the data, effectively suppressing random noise while preserving the amplitude of the valid signal. Model experiments show that the adaptive convolutional denoising model used in this invention can better retain the valid information in the data.

[0053] The key points and protection points of this invention are as follows:

[0054] (1) A deep learning algorithm based on an adaptive convolutional denoising network is used to denoise the near-field microwave signal. The denoising network structure is constructed by combining a noise level estimation network and a convolutional denoising network. The noise level estimation network can improve the training speed and denoising performance of the convolutional denoising network, thereby achieving adaptive intelligent denoising.

[0055] (2) Combining the noise estimation method based on weak signal blocks, and considering the characteristics of microwave near-field detection and acquisition data, it can provide more accurate data for network training.

[0056] (3) Batch normalization is adopted to speed up training and improve the accuracy of the training model in estimating the standard deviation of data noise;

[0057] (4) In order to reduce the impact of noise level estimation error on denoising, an asymmetric loss function is adopted to eliminate the underfitting and overfitting problems of the convolutional denoising network.

[0058] (5) The network training uses two types of data, model noise and actual noise, to train the network model alternately, thereby improving the denoising effect and generalization ability of the network model.

[0059] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for suppressing random noise in near-field microwave signals using an adaptive convolutional denoising network, characterized in that: The adaptive convolutional denoising network consists of two sub-networks: a noise level estimation network and a convolutional denoising network. The steps include: Step 1: Use the block matching algorithm to denoise the actual collected data, select the denoised data and noisy data with good denoising effect as training data, and build a data training library using two types of data: model noise and actual noise. Step 2: Use a noise estimation method based on weak signal blocks to estimate the noise level of the training data, obtain the noise standard deviation, and then use the obtained data to train the noise level estimation network. When training a network, noise is added to the noiseless model data to obtain noisy data. The noise standard deviation is known, but the noisy model data and the collected noisy actual data differ in terms of data characteristics. As for the collected actual data, the noise is caused by the collection environment, and it is impossible to obtain accurate actual noiseless data. We can only use the denoised data that is close to noiseless data as a reference to calculate the noise standard deviation of the actual data in the training library. Therefore, a noise estimation method based on weak signal blocks is used to calculate the actual noise standard deviation. This method utilizes probabilistic and statistical thinking, fitting the standard deviation of each segmented data block by statistically analyzing the data segmentation results, and obtaining the noise standard deviation by decomposing the covariance matrix of the noisy data block to minimize the eigenvalue. : （1）； In the formula, Represents the covariance matrix of the noisy data block. Represents the covariance matrix of the noiseless data block. This represents the smallest eigenvalue of the covariance matrix of the noisy data block. It represents the smallest eigenvalue of the covariance matrix of the noiseless data block; Based on actual data, It is unknown, but it can be replaced by the minimum eigenvalue of the covariance matrix of the weak signal block set. The weak signal block set can be selected by the signal strength of the local gradient matrix of the data and its statistical properties. Step 3: Calculate the noise estimation standard deviation of the noisy data y using a noise level estimation network. Then, the data and parameters are input into a convolutional denoising network to train the denoising model, and the standard deviation is estimated using noise. Adjust network training parameters; Step 4: Use a convolutional denoising network to extract the noisy data y and estimate the standard deviation of the noise. and noise data As input, an asymmetric loss function is used to eliminate the underfitting and overfitting problems of the convolutional denoising network, and the final denoising model is obtained. Convolutional denoising networks are asymmetrically sensitive to the error in noise estimation results. An asymmetric loss function is used to eliminate this asymmetric sensitivity and suppress noise. Estimated noise level at point i and actual value The objective loss function for constructing the entire adaptive convolutional denoising network is: （2）； In the formula, An asymmetric loss function for estimating the noise level of the network; Here is the loss function for noise suppression, used to ensure that noise can be sufficiently suppressed: （3）； when In order to fully suppress noise, a denoising weighting factor is used. By setting the weight parameter α: 0 < α < 0.5, the denoising weights of data with different noise standard deviations can be balanced. To address the overfitting loss function and ensure sufficient preservation of the effective signal after denoising, a total variation regularization term is introduced to constrain it. Smoothness: （4）； in, Its weighting factor; The loss function for the convolutional denoising network is defined using the L2 norm to define the reconstruction error: （5）； Step 5: Use the trained denoising model to denoise other data, achieving adaptive intelligent denoising based on the characteristics of microwave near-field data, while protecting the effective signal from being suppressed, and obtaining the final denoising result.

2. The method for suppressing random noise in near-field microwave signals using an adaptive convolutional denoising network according to claim 1, characterized in that: In step 3, the noise level estimation network uses a five-layer convolutional network with a kernel size of 3×3×32 and employs batch normalization to accelerate training and improve the accuracy of the training model in estimating the standard deviation of data noise.

3. The method for suppressing random noise in near-field microwave signals using an adaptive convolutional denoising network according to claim 1, characterized in that: In step 4, the convolutional denoising network uses a 16-layer U-Net network structure and employs residual learning to learn the residual data. In other words, noise data is obtained by removing noise data from the input noisy data.

Images