Deep learning method and device based on joint detection and reconstruction of weak line spectrum under non-gaussian impulse noise

Abstract

The application relates to the field of underwater acoustic signal processing and discloses a deep learning method and device for weak line spectrum joint detection and reconstruction under non-Gaussian pulse noise. In order to solve the problem that in the prior art, a weak line spectrum detection method mainly extracts target features by using experience knowledge, and the performance of the method depends on the cognitive degree of an operator for samples to a great extent, and the problem that underwater acoustic signal processing technology under a Gaussian noise assumption is directly applied to signal detection under non-Gaussian noise, which causes the performance to decrease or even fail, the technical scheme provided by the application is as follows: a weak line spectrum joint detection and reconstruction method comprises the following steps: collecting pulse noise, and constructing a ship line spectrum data set according to the noise; pre-processing the data set to obtain a lofar graph; extracting weak features of the lofar graph, and obtaining a line spectrum detection and reconstruction result; training a preset model according to the lofar graph and the result, and outputting a trained model. The application is suitable for weak line spectrum detection and reconstruction under Gaussian pulse noise.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of underwater acoustic signal processing, in particular to underwater acoustic signal detection and reconstruction. BACKGROUND

[0002] Single frequency detection of underwater radiated noise is a common method for detecting low noise and quiet targets. Using short-time Fourier transform, time-frequency transform is projected onto the time plane and the frequency plane to form a three-dimensional stereogram (lofargram), which embodies the rich characteristics of underwater radiated noise. Therefore, the lofargram is often used to analyze the characteristics of passive sonar signals. However, under low signal-to-noise ratio conditions, irregular fluctuations in frequency caused by moving targets and strong background noise can weaken the ability to detect line spectra.

[0003] For weak line spectrum detection in the lofargram, image processing methods, neural networks and statistical models can be used to detect weak line spectra in the lofargram. Image processing and neural network methods remove noise from complex image semantic features to obtain line spectrum trajectories. Under low signal-to-noise ratio, the line spectrum trajectories are not satisfactory. Statistical models track the optimal line spectrum trajectory in multiple frames of power spectrum data. The above researches are mostly applied to Gaussian distributed ocean ambient noise. In particular, due to the superposition of sea water thermal noise, hydrodynamic noise, ice under noise, biological noise and other noise, the ocean ambient noise has strong pulse characteristics. Under such non-Gaussian noise, underwater acoustic signal processing methods may fail. Some researches have made statistical analysis and modeling of non-Gaussian ocean ambient noise. The generation and propagation of underwater impulse noise conforms to the "heavy tail" statistical characteristics of symmetric alpha stable (S) distribution. Non-Gaussian impulse noise suppression can be achieved using various methods, including median filters and memoryless analog nonlinear preprocessor (MANP). However, weak line spectra are difficult to detect under lower generalized signal-to-noise ratio.

[0004] The above weak line spectrum detection methods mainly use empirical knowledge to extract target features, and their performance largely depends on the operator's cognitive level of the sample. At the same time, directly applying underwater acoustic signal processing techniques under Gaussian noise assumption to signal detection under non-Gaussian noise will degrade its performance or even fail. Deep learning, with its strong non-linear fitting ability and less required prior knowledge, provides a new direction for weak line spectrum detection under non-Gaussian impulse noise. SUMMARY

[0005] To solve the problem in the prior art that weak line spectrum detection methods mainly use empirical knowledge to extract target features, and their performance largely depends on the operator's cognitive level of the sample. At the same time, directly applying underwater acoustic signal processing techniques under Gaussian noise assumption to signal detection under non-Gaussian noise will degrade its performance or even fail, the technical scheme provided by the present application is as follows:

[0006] The deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise is characterized in that the method comprises:

[0007] Step 1: collecting impulse noise, and constructing a ship line spectrum data set according to the noise;

[0008] Step 2: preprocessing the data set to obtain a lofar graph;

[0009] Step 3: extracting weak features of the lofar graph, and obtaining a result of line spectrum detection and reconstruction;

[0010] Step 4: training a preset model according to the lofar graph and the result, and outputting a trained model.

[0011] Further, a preferred embodiment is provided, wherein in step 1, the data set is constructed by using an alpha stable distribution modeling method.

[0012] Further, a preferred embodiment is provided, wherein in step 2, the preprocessing comprises the operations of impulse noise suppression and video analysis.

[0013] Further, a preferred embodiment is provided, wherein in step 3, the weak features of the lofar graph are deeply extracted based on deep learning, a convolutional neural network is used as a basic unit, and a result of line spectrum detection and reconstruction is obtained.

[0014] Further, a preferred embodiment is provided, wherein in step 3, the process of extracting the weak features of the lofar graph further comprises a step of using a line spectrum detection algorithm.

[0015] Further, a preferred embodiment is provided, wherein in step 4, the training method is specifically based on a double classification adaptive weighted loss function.

[0016] Further, a preferred embodiment is provided, wherein the impulse noise comprises Gaussian impulse noise and non-Gaussian impulse noise.

[0017] Based on the same inventive concept, the present application also provides a deep learning device for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise, which comprises:

[0018] Module 1: used for collecting impulse noise, and constructing a ship line spectrum data set according to the noise;

[0019] Module 2: used for preprocessing the data set to obtain a lofar graph;

[0020] Module 3: for extracting weak features of the lofar graph, and obtaining the results of line spectrum detection and reconstruction;

[0021] Module 4: for training a preset model according to the lofar graph and the results, and outputting the trained model.

[0022] Based on the same inventive concept, the application also provides a computer storage medium for storing a computer program for being read by a computer to execute the deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise.

[0023] Based on the same inventive concept, the application also provides a computer comprising a processor and a storage medium, wherein the storage medium stores a computer program for being read by the processor to enable the computer to execute the deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise.

[0024] Compared with the prior art, the application has the advantages of:

[0025] The deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise provided by the application proposes a deep learning algorithm for weak line spectrum joint detection and reconstruction, extracts weak line spectrum features by combining impulse noise suppression preprocessing and deep learning method, and simultaneously gives a lofar graph with detection rate and line spectrum prominence.

[0026] The deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise provided by the application can effectively detect and reconstruct weak spectrum line features.

[0027] The deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise provided by the application provides a new direction for weak line spectrum detection under non-Gaussian impulse noise.

[0028] It is suitable for weak line spectrum detection and reconstruction under Gaussian impulse noise. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The flowchart of the deep learning method for weak line spectrum joint detection and reconstruction under non-Gaussian impulse noise provided for the first embodiment;

[0030] Figure 2 The lofar graph trained by the DRNet mentioned in the eleventh embodiment;

[0031] wherein, First Conv represents the first layer of convolution, The shared Encoder represents the shared encoding layer for feature extraction, The detection decoder represents the detection decoder for obtaining the detection result, The reconstruction decoder represents the reconstruction decoder for reconstructing the extracted features, Reconstruction, Detection task uncertainty respectively represent reconstruction, detection task uncertainty for balancing the dual-task loss function, Dual-task loss represents the dual-task loss;

[0032] Figure 3 the DRNet network structure mentioned in embodiment eleven is shown in the schematic diagram;

[0033] wherein, the abscissa represents frequency, and the ordinate represents time;

[0034] Figure 4 the ROC curve for evaluating the detection rate mentioned in embodiment eleven is shown in the diagram;

[0035] wherein, the abscissa represents the false alarm rate, and the ordinate represents the detection rate;

[0036] Figure 5 the LLA, mIOU diagram for evaluating the reconstruction performance mentioned in embodiment eleven is shown in the diagram;

[0037] wherein, the abscissa represents the generalized signal-to-noise ratio, and the ordinate represents the line spectrum position accuracy and the line spectrum classification accuracy;

[0038] Figure 6 the lofargram example diagram obtained by various methods mentioned in embodiment eleven is shown in the diagram;

[0039] wherein, the abscissa represents frequency, and the ordinate represents time. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely 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 the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0041] The present application will be further described in conjunction with the specific embodiments and drawings:

[0042] Embodiment one, in conjunction with Figure 1To illustrate the embodiments, the embodiments provide a deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise, characterized in that the method comprises:

[0043] Step 1: Collecting impulse noise and constructing a ship line spectrum dataset according to the noise;

[0044] Step 2: Preprocessing the dataset to obtain a lofar graph;

[0045] Step 3: Extracting weak features of the lofar graph and obtaining results of line spectrum detection and reconstruction;

[0046] Step 4: Training a preset model according to the lofar graph and the results, and outputting the trained model.

[0047] Specifically, the method comprises:

[0048] Step 1: Modeling Gaussian / non-Gaussian impulse noise by using alpha stable distribution, and constructing a ship line spectrum dataset under Gaussian / non-Gaussian impulse noise;

[0049] Step 2: Obtaining a lofar graph by using preprocessing methods such as AINP and time-frequency analysis;

[0050] Step 3: Constructing a line spectrum detection and reconstruction framework based on deep learning (AINP+LR-DRNet), using a convolutional neural network as a basic calculation unit, combining a multi-task design double-output structure, deeply extracting weak line spectrum features, and giving results of line spectrum detection and reconstruction;

[0051] Step 4: An adaptive weighted loss function based on double classification is proposed for training the model. The adaptive weighted loss function based on double classification is as follows:

[0052]

[0053]

[0054]

[0055] where σ d , σ r are the observation noise parameters of the models for detection and reconstruction tasks, h∈{0,1} represents H0 and H1 hypotheses, p is the probability of output sample class represented by the Softmax function, λ=|G - | / |G| and 1-λ=|G + | / |G| represent the proportion of noise pixels and line spectrum pixels in the actual dataset, respectively, and p f,tL d represents the detection loss function, L r represents the reconstruction loss function, N1 represents the batch size, i represents the sample index, p represents the sample probability, N2 represents the number of samples of H1 in the batch size, p f,t represents the probability at position (f, t), G represents the total number of pixels in the sample, G - represents the number of background pixels in the sample, G + represents the number of line spectrum pixels in the sample.

[0056] Step five: propose a line spectrum detection algorithm (LR-DRNet) based on the framework.

[0057] Step one: offline training.

[0058] In offline training, the lofargram dataset after AINP preprocessing is constructed and labeled:

[0059] (O l , x) = {(l (m) , x (m) ), (l i , x f ),..., (l LR-DRNet , x i )};

[0060] where O i represents the lofargram sample set, l represents the sample, x represents the sample label, and M represents the index of the label.

[0061] For test statistics, LR-DRNet can extract the potential spectral line features in lofargram. After a series of convolution layers, pooling layers and activation functions, the probability of lofargram belonging to H1 or H0 is obtained as follows:

[0062]

[0063]

[0064] H1 represents the hypothesis under the existence of line spectrum, H0 represents the hypothesis under the non-existence of line spectrum, P represents the probability, x k represents the predicted sample category, l k represents the input sample, represents the model parameters, represents the probability of the model predicting that the sample belongs to H1, represents the probability of the model predicting that the sample belongs to H0.

[0065] The detection output of the fully trained LR-DRNet is as follows:

[0066]

[0067] According to the Bayesian and NP criteria, the test statistics is constructed as follows:

[0068]

[0069] wherein, represents the probability of the pre-trained model inferring that the sample belongs to H i , represents the probability of the pre-trained model inferring that the sample belongs to H1, represents the probability of the pre-trained model inferring that the sample belongs to H0.

[0070] The detection threshold is calculated as follows:

[0071]

[0072] wherein P f = P[Λ LR-DRNet |H0> η], and M is the number of H0 samples.

[0073] Step 2: Online detection.

[0074] In the online detection stage, the unlabeled sample is directly input into the fully trained LR-DRNet, and compared with the detection threshold to detect whether the line spectrum exists or not.

[0075]

[0076] wherein, represents the hypothesis under the unlabeled sample, represents the probability of the pre-trained model under H1, represents the probability of the pre-trained model under H0.

[0077] Step 6: Take the results of the framework proposed in steps 4, 5 and 3 as the final output.

[0078] The embodiment aims at the problem of weak line spectrum detection and reconstruction under non-Gaussian impulse noise, and proposes a deep learning algorithm for joint detection and reconstruction of weak line spectrum. The algorithm combines impulse noise suppression preprocessing and deep learning method to extract weak line spectrum features, and simultaneously gives the detection rate and lofargram of the significant line spectrum. The results show that the weak line spectrum joint detection and reconstruction algorithm based on impulse noise preprocessing and deep learning proposed in the embodiment can effectively detect and reconstruct the weak spectrum line features.

[0079] Embodiment two, the embodiment is further limited to the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment one, in step 1, the method for constructing the data set is: using the method of alpha stable distribution modeling.

[0080] Specifically, in step one,

[0081] First step: simulate Gaussian / non-Gaussian impulse noise. Non-Gaussian impulse noise can be described by alpha stable distribution, and its characteristic function can be expressed as:

[0082]

[0083]

[0084] Where 0 < a ≤ 2, -1 ≤ β ≤ 1, γ > 0, -∞ < δ < ∞, a is the pulse intensity, defined as the intensity of the pulse of the distribution, the position parameter δ determines the center of the distribution, and the dispersion coefficient γ defines the degree of deviation of the distribution from its mean value, j represents the imaginary number, t represents the time sequence, sgn represents the sign function, and the skewness parameter β determines the symmetry of the distribution, when β = 0, the distribution is SαS distribution.

[0085] Second step: simulate ship line spectrum under Gaussian / non-Gaussian impulse noise. The ship spectrum observed in the kth time interval can be described as:

[0086]

[0087] Where a i represents the signal amplitude, I is the number of spectrum lines. f i (t k ) is the frequency that changes with t k . is the initial phase, n(t k ) is the sampling point of SαS distribution noise at t k .

[0088] Embodiment three, the embodiment is further limited to the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment one, in step 2, the preprocessing includes: pulse noise suppression and video analysis operations.

[0089] Specifically, in step two,

[0090] First step: establish a pulse noise suppression preprocessor. A nonlinear suppression method for abnormal amplitude in non-Gaussian impulse noise is used for preprocessing, and y(t) is obtained after the input signal x(t) is subjected to nonlinear suppression:

[0091]

[0092] where θ(t) can be obtained by the following formula:

[0093] θ(t) = (1 + 2θ0)Q2(t)

[0094] where Q2(t) is the second quartile of the absolute value of the input signal, and θ0 is a constant coefficient set to 1.5.

[0095] Second step: Lofar drawing. After the above nonlinear suppression, time-frequency analysis is performed to obtain the lofargram.

[0096] Embodiment four, this embodiment is a further limitation of the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment one, and step 3 is specifically: based on deep learning, using a convolutional neural network as a basic unit, deeply extracting the weak features of the lofargram, and obtaining the results of line spectrum detection and reconstruction.

[0097] Embodiment five, this embodiment is a further limitation of the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment four, and step 3 further includes: using a line spectrum detection algorithm in the process of extracting the weak features of the lofargram.

[0098] Embodiment six, this embodiment is a further limitation of the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment one, and in step 4, the training method is specifically: training based on a double classification adaptive weighted loss function.

[0099] Embodiment seven, this embodiment is a further limitation of the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment one or two, and the impulse noise includes Gaussian impulse noise and non-Gaussian impulse noise.

[0100] Embodiment eight, this embodiment provides a deep learning device for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise, and the device comprises:

[0101] Module 1: for collecting impulse noise and constructing a ship line spectrum data set according to the noise;

[0102] Module 2: for preprocessing the data set to obtain a lofargram;

[0103] Module 3: for extracting weak features of the lofargram and obtaining results of line spectrum detection and reconstruction;

[0104] Module 4: training a preset model according to the lofar graph and the result, and outputting the trained model.

[0105] Embodiment nine, the embodiment provides a computer storage medium for storing a computer program, the computer program is read by a computer to execute the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in any one of embodiments one to seven.

[0106] Embodiment ten, the embodiment provides a computer including a processor and a storage medium, the storage medium stores a computer program, the computer program is read by the processor, so that the computer executes the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in any one of embodiments one to seven.

[0107] Embodiment eleven, the embodiment is a specific embodiment of the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise provided in embodiment one, and is used to verify the advantages and benefits of the method, specifically:

[0108] The overall flow of the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise under non-Gaussian impulse noise is as shown in Figure 1 The specific steps are as follows:

[0109] (1) Establish a Gaussian / non-Gaussian impulse noise under ship line spectrum data set and divide it into training and test set.

[0110] When simulating SαS distribution noise, randomly select α in the range of [1.3, 2], β is set to 0, and λ and δ are set to 1 and 0 respectively. Considering the ship line spectrum within 100Hz and the MSNR within [-26, -19]dB. The sampling rate is 1000Hz. The number of line spectrum is only 5. Add SαS distribution noise to the sinusoidal signal by MSNR. Figure 2 The lofargram of multiple sinusoidal signals with different MSNR is given. In the lofargram below-22dB, the existence of line spectrum cannot be perceived. For MSNR in the range of-22dB to-26dB, we repeated 1200 times of Monte Carlo simulation to simulate the situation under various parameters, and divided the data set into 85% training and 15% test. Therefore, our training data set contains 9600 and 6800 lofargrams, and the test set contains 1440 and 1200 lofargrams.

[0111] (2) Preprocess the received signal.

[0112] A nonlinear suppression method for abnormal amplitude in non-Gaussian impulse noise is used for preprocessing, and y(t) is obtained after the input signal x(t) is subjected to nonlinear suppression:

[0113]

[0114] Where θ(t) can be obtained by the following formula:

[0115] θ(t) = (1 + 2θ0)Q2(t);

[0116] Where Q2(t) is the second quartile of the absolute value of the input signal, and θ0 is a constant coefficient set to 1.5.

[0117] And then the short-time Fourier transform is performed to obtain the loggram.

[0118] (3) The LR-DRNet as shown in Figure 3 is trained offline to obtain optimal network parameters.

[0119] Step 1: Use the adaptive weighted loss function based on binary classification to train the model. The adaptive weighted loss function based on binary classification is as follows:

[0120]

[0121]

[0122]

[0123] Where σ d , σ r are the observation noise parameters of the model for detection and reconstruction tasks respectively. h ∈ {0, 1} represents H0 and H1 hypotheses, and p is the probability of output sample class represented by the Softmax function. λ = |G - | / |G| and 1-λ = |G + | / |G| represent the proportion of noise pixels and line spectrum pixels in the actual data set respectively. p f,t is the predicted value of the sample at the (f, t) position after Sigmoid under the H1 hypothesis.

[0124] Step 2: Use the line spectrum detection algorithm (LR-DRNet) to obtain the threshold.

[0125] In offline training, the lofargram data set after AINP preprocessing is constructed and labeled:

[0126]

[0127] For the test statistics, LR-DRNet can extract the potential spectral line features in lofargram. After a series of convolution layers, pooling layers and activation functions, the probability of lofargram belonging to H1 or H0 is as follows:

[0128]

[0129]

[0130] The detection output of the fully trained LR-DRNet is:

[0131]

[0132] According to the Bayesian and NP criteria, the test statistics is constructed as follows:

[0133]

[0134] The detection threshold is calculated as follows:

[0135]

[0136] Where P f =P[Λ LR-DRNet |H0>η], M is the number of H0 samples.

[0137] (4) Online detection and reconstruction are performed on the LR-DRNet as shown in Figure 3 .

[0138] Parameter fine-tuning is performed using the network parameters obtained in (3) to reduce the inadaptability of the model. Through the comparison of the test statistics and the threshold in (3) and the reconstruction of lofargram, the results of Figure 4 , Figure 5 and Figure 6 are finally obtained. The present application has a high detection rate, especially when the MSNR is-24dB to-26dB. Among the five methods, the present application has the best reconstruction performance in terms of the difference in reconstruction performance, and can reconstruct more spectral line information. Therefore, the feasibility of the present application in detection and reconstruction under Gaussian / non-Gaussian impulse noise is proved.

[0139] The present embodiment constructs a data set for Gaussian / non-Gaussian impulse noise through the SaS distribution model, performs impulse noise suppression using the impulse noise suppression method, and captures potential spectral line features through a specially designed dual-output network structure and corresponding adaptive dual-classification loss for the dual-task requirements of detection and reconstruction. The results prove that the method has good detection and reconstruction effect.

[0140] The technical solutions of the present application are described in further detail through several specific embodiments above, in order to highlight the advantages and benefits of the technical solutions provided by the present application. However, the above several specific embodiments are not used as a limitation to the present application, and any reasonable modifications and improvements, reasonable combinations and equivalent replacements of the embodiments, etc. based on the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A deep learning method based on joint detection and reconstruction of weak line spectrum under non-Gaussian impulsive noise, characterized in that, The method comprises: Step 1: collecting impulse noise, and constructing a ship line spectrum dataset according to the noise; Step 2: preprocessing the dataset to obtain a lofar graph; Step 3: extracting weak features of the lofar graph, and obtaining the results of line spectrum detection and reconstruction; Step 4: training a preset model according to the lofar graph and the results, and outputting the trained model; Specifically, the method comprises: Step one: modeling Gaussian / non-Gaussian impulse noise by alpha stable distribution, and constructing a ship line spectrum dataset under Gaussian / non-Gaussian impulse noise; Step two: obtaining a lofar graph through pulse noise suppression (AINP) and time-frequency analysis and other preprocessing methods; Step three: constructing a line spectrum detection and reconstruction framework (AINP+LR-DRNet) based on deep learning, using convolutional neural network as the basic calculation unit, combining multi-task design double output structure, deeply extracting weak line spectrum features, and giving the results of line spectrum detection and reconstruction; Step four: an adaptive weighted loss function based on double classification is proposed for training the model; the adaptive weighted loss function based on double classification is as follows: ； ； ； in , These are the observation noise parameters of the models used for detection and reconstruction tasks, respectively. express and Assuming, The probability of the output sample class is represented by the Softmax function. and These represent the proportions of noise pixels and line spectrum pixels in the actual dataset, respectively. for Assume the following sample is in The predicted value at the location after Sigmoid. This represents the dual-task loss function. The weights represent the weights of the detection loss function. This represents the detection loss function. Represents the reconstruction loss function. Indicates batch size. Indicates the sample index. Represents the sample probability. Indicates batchsize The number of samples, Let f represent the probability at position (f,t). This represents the total number of pixels in the sample. This represents the number of background pixels in the sample. Indicates the number of pixels in the sample line spectrum; Step five: a line spectrum detection algorithm (LR-DRNet) based on the framework is proposed; First step: offline training; In offline training, the lofar graph dataset after AINP preprocessing is constructed and labeled: ； wherein, represents a set of lofar gram samples, represents a sample, represents a sample label, represents an index of a label; For the test statistics, LR-DRNet can extract the potential spectral line features in lofargram; after a series of convolution layers, pooling layers and activation functions, the probability that lofargram belongs to or is as follows: ； represents a hypothesis under the presence of line spectra, represents a hypothesis under the absence of line spectra, represents a probability, represents a predicted sample class, represents an input sample, represents a model parameter, represents a probability that a model hypothesizes that a sample belongs to represents a probability that a model hypothesizes that a sample belongs to represents a probability that a model hypothesizes that a sample belongs to represents a probability that a model hypothesizes that a sample belongs to 2. The deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulsive noise according to claim 1, characterized in that, In step 1, the method for constructing the dataset is to model alpha stable distribution. 3.The deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulsive noise according to claim 1, wherein, In step 2, preprocessing includes pulse noise suppression and video analysis. 4.The deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulsive noise according to claim 1, wherein, In step 3, based on deep learning, convolutional neural network is used as the basic unit to deeply extract weak features of the lofar graph, and the results of line spectrum detection and reconstruction are obtained.

5. The deep learning method for joint detection and reconstruction of weak linespectrum under non-Gaussian impulsive noise according to claim 4, characterized in that, In step 3, it also includes the steps of using line spectrum detection algorithm in the process of extracting weak features of the lofar graph. 6.The deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulsive noise according to claim 1, wherein, In step 4, the training method is specifically based on adaptive weighted loss function based on double classification.

7. The deep learning method for joint detection and reconstruction of weak linespectrum under non-Gaussian impulsive noise according to claim 1 or 2, characterized in that, The pulse noise includes Gaussian pulse noise and non-Gaussian pulse noise.

8. A deep learning device based on joint detection and reconstruction of weak line spectrum under non-Gaussian impulsive noise, characterized in that, The device comprises: Module 1: for collecting impulse noise, and constructing a ship line spectrum dataset according to the noise; Module 2: for preprocessing the dataset to obtain a lofar graph; Module 3: for extracting weak features of the lofar graph, and obtaining the results of line spectrum detection and reconstruction; Module 4: for training a preset model according to the lofar graph and the results, and outputting the trained model; Specifically, the method comprises: Step one: modeling Gaussian / non-Gaussian impulse noise by alpha stable distribution, and constructing a ship line spectrum dataset under Gaussian / non-Gaussian impulse noise; Step two: obtaining a lofar graph through pulse noise suppression (AINP) and time-frequency analysis and other preprocessing methods; Step three: constructing a line spectrum detection and reconstruction framework (AINP+LR-DRNet) based on deep learning, using convolutional neural network as the basic calculation unit, combining multi-task design double output structure, deeply extracting weak line spectrum features, and giving the results of line spectrum detection and reconstruction; Step four: an adaptive weighted loss function based on double classification is proposed for training the model; the adaptive weighted loss function based on double classification is as follows: Step five: a line spectrum detection algorithm (LR-DRNet) based on the framework is proposed; Step three: Constructing a deep learning-based line spectrum detection and reconstruction framework (AINP+LR-DRNet), using a convolutional neural network as the basic computing unit, combining a double-output structure with multi-task design, deeply extracting weak line spectrum features while giving the results of line spectrum detection and reconstruction; Step four: An adaptive weighted loss function based on double classification is proposed for training the model; the adaptive weighted loss function based on double classification is as follows: ； ； ； in , These are the observation noise parameters of the models used for detection and reconstruction tasks, respectively. express and Assuming, The probability of the output sample class is represented by the Softmax function. and These represent the proportions of noise pixels and line spectrum pixels in the actual dataset, respectively. for Assume the following sample is in The predicted value at the location after Sigmoid. This represents the dual-task loss function. The weights represent the weights of the detection loss function. This represents the detection loss function. Represents the reconstruction loss function. Indicates batch size. Indicates the sample index. Represents the sample probability. Indicates batchsize The number of samples, Let f represent the probability at position (f,t). This represents the total number of pixels in the sample. This represents the number of background pixels in the sample. Indicates the number of pixels in the sample line spectrum; Step five: A line spectrum detection algorithm (LR-DRNet) based on the framework is proposed; First step: Offline training; In offline training, the lofargram dataset after AINP preprocessing is constructed and labeled: ； wherein, represents a set of lofar gram samples, represents a sample, represents a sample label, represents an index of a label; For the test statistics, LR-DRNet can extract the potential spectral line features in lofargram; after a series of convolution layers, pooling layers and activation functions, the probability that lofargram belongs to or is as follows: ； This represents the assumption that the line spectrum exists. This represents the assumption that the line spectrum does not exist. Represents probability. Indicates the predicted sample category, Indicates the input sample. Indicates model parameters, The model predicts that the sample belongs to The probability, The model predicts that the sample belongs to The probability of.

9. Computer storage medium for storing a computer program, characterized in that The computer program is used to be read by a computer to execute the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise according to any one of claims 1-7.

10. A computer comprising a processor and a storage medium, said storage medium having stored therein a computer program, characterized in that, The computer program is used to be read by the processor, so that the computer executes the deep learning method for joint detection and reconstruction of weak line spectrum under non-Gaussian impulse noise according to any one of claims 1-7.

Images