CN112966667B

CN112966667B - Method for identifying one-dimensional distance image noise reduction convolution neural network of sea surface target

Info

Publication number: CN112966667B
Application number: CN202110366635.3A
Authority: CN
Inventors: 简涛; 王哲昊; 王海鹏; 刘瑜; 刘传辉; 李刚; 李辉; 杨予昊; 张健
Original assignee: Naval Aeronautical University
Current assignee: Naval Aeronautical University
Priority date: 2021-04-06
Filing date: 2021-04-06
Publication date: 2022-12-16
Anticipated expiration: 2041-04-06
Also published as: CN112966667A

Abstract

The invention discloses a method for identifying a one-dimensional distance image noise reduction convolution neural network of a sea surface target, belonging to the field of radar signal processing. The method reasonably preprocesses original HRRP data under the condition of low signal-to-noise ratio, constructs a plurality of sea surface target data sets under the condition of different signal-to-noise ratios, constructs a one-dimensional noise reduction convolutional neural network by utilizing a deep learning technology, improves the signal-to-noise ratio of the low signal-to-noise ratio data on the basis of keeping the high signal-to-noise ratio data to be free of fluctuation, reduces the learning burden of a deep neural network by utilizing a residual error structure of the convolutional neural network, further constructs an intelligent sea surface target classification and identification model integrating noise reduction and classification, improves the identification accuracy of sea surface targets, improves the identification performance of the sea surface targets under the condition of low signal-to-noise ratio, enhances the classification and identification capacity of sea radars under the complex sea surface environment, and has popularization and application values.

Description

Method for identifying one-dimensional distance image noise reduction convolution neural network of sea surface target

Technical Field

The invention belongs to the field of radar signal processing, and particularly relates to a method for identifying a one-dimensional distance image noise reduction convolution neural network of a sea surface target.

Background

The reliable sea surface target identification method is very important for each link of the sea target reconnaissance. At present, a high-resolution broadband radar technology is widely applied to the field of target identification, and a high-resolution one-dimensional range profile (HRRP) represents the distribution situation of scattering centers along the radar sight direction, contains the structural characteristics of targets, and has an important role in the field of radar sea surface target identification due to the advantages of easiness in acquisition and processing and the like.

Various HRRP target identification methods exist, such as a subspace-based HRRP target identification algorithm, a scattering center model-based HRRP radar target identification method and the like. In the subspace-based identification algorithm, aiming at the nonlinear problem, a kernel function is required to be used for extracting feature information by ascending and then descending dimensions, a plurality of artificially designed parameters exist in the kernel function, and whether the selection of the parameters is reasonable or not directly influences the target identification effect; in the recognition algorithm based on the scattering center model, classification and recognition are carried out according to feature information of targets extracted from each scattering center. However, in the actual radar detection process, the sea surface environment is complex and changeable, the signal-to-noise ratio for obtaining the HRRP is often low, the problem is particularly prominent in a long-distance detection scene, and the low signal-to-noise ratio can greatly affect the two characteristic information extraction methods, so that the recognition accuracy and robustness of the two characteristic information extraction methods face great challenges. The deep learning technology in the artificial intelligence can automatically extract the essential features of the targets, and the end-to-end learning mode has potential advantages in the aspects of improving the accuracy, robustness, intelligence and the like of sea surface target identification.

Aiming at the complex situation that sea surface target identification may face both high signal-to-noise ratio and low signal-to-noise ratio, how to improve the sea surface target identification accuracy under the condition of low signal-to-noise ratio on the basis of keeping the target identification accuracy under the condition of high signal-to-noise ratio is a key for improving the target identification capability under the complex sea surface environment and is also one of the research key points of the sea surface target HRRP identification technology.

Disclosure of Invention

Aiming at the condition of low signal to noise ratio, how to reasonably preprocess original HRRP data, construct a plurality of sea surface target data sets under the condition of different signal to noise ratios, and construct a one-dimensional noise reduction convolution neural network (DnCNN) by utilizing a deep learning technology, on the premise of keeping the high signal to noise ratio data basically free of fluctuation, the adverse effect of noise in the low signal to noise ratio data is reduced, the learning burden of a deep neural network is reduced, and then a sea surface target HRRP intelligent classification identification model integrating noise reduction and classification is constructed, the sea surface target identification accuracy under the condition of low signal to noise ratio is further improved, the influence of human factors is avoided, and the classification identification capability and robustness of a sea radar under the complex sea surface environment are improved.

The invention discloses a method for identifying a one-dimensional distance image noise reduction convolution neural network of a sea surface target, which comprises the following steps:

step 1, acquiring noise-containing target HRRP data under various typical sea surface environments according to sea surface target identification requirements, carrying out noise addition and distance translation expansion on the target HRRP data, and constructing a training set and a test set of a model on an expanded data set according to a fixed proportion; secondly, according to the characteristics of the sea surface target HRRP, the HRRP is converted into a two-dimensional image, and the physical meaning of the coordinate axis of the two-dimensional image is given according to the HRRP information, so that different characteristic information can be conveniently extracted; the method comprises the following specific steps:

in order to be closer to a real sea surface environment, firstly, data is subjected to noise adding processing, random noises with different signal-to-noise ratios are added to simulate the sea surface environment under different conditions, meanwhile, in order to meet the requirement of a sea surface target classification recognition model with deep level feature information autonomous learning capacity on a data set, the data subjected to noise adding processing is subjected to data set expansion, and finally, a training set and a testing set of the model are constructed according to a certain proportion on the expanded data set.

Firstly, obtaining multi-class noise interference-free sea surface target HRRP sample data x _nj (n＝1,2,...,N；j＝1,2,...,C)，x _nj The data is an M-dimensional vector and represents sample data corresponding to the jth class of targets at the nth azimuth, C is the number of classes of targets, N is the number of azimuths, namely each class of targets contains sample data of N azimuths, and the length of the sample data at each azimuth is M. Generating a random NOISE sequence NOISE, and estimating the effective power PN of the random NOISE sequence according to a given SNR (signal to NOISE ratio) in order to ensure that the noiseless data and the random NOISE are superposed at the given SNR:

where PS represents the effective power of the noiseless data, which is calculated as follows

PS＝|x _nj | ² /length(x _nj ) (2)

Wherein, | · | is a module value of the vector, and length (·) is a dimension of the vector.

And combining the effective power PN to standardize the random NOISE sequence NOISE to obtain a NOISE sequence NOISE to be added, thereby completing the NOISE adding processing of the noiseless data. The normalization process is as follows:

noise＝sqrt(PN)/std(NOISE)*NOISE (3)

where sqrt (. Cndot.) is a square root function and std (. Cndot.) is a standard deviation calculation function.

Secondly, expanding the HRRP data set after noise addition, wherein the expanding method comprises two steps:

method 1.1: the original sample data without noise interference is expanded, and noise is added for multiple times to the original sample data under the condition of a given SNR by utilizing the noise adding processing mode;

method 1.2: and expanding the data set through local data cyclic translation, performing cyclic translation in the distance direction on the data subjected to noise processing, and expanding the data set by using the data subjected to translation each time.

And finally, after the data sets are expanded according to the method 1.1 and the method 1.2, in order to ensure the classification and identification performance of the model, the whole group of data sets are randomly disturbed, recombined and then divided into a training set and a testing set according to the proportion of each type of targets A to B.

According to the characteristics of the sea surface target HRRP, the HRRP is converted into a two-dimensional image, and the physical meaning of the coordinate axis of the two-dimensional image is given according to the HRRP information, so that different characteristic information can be conveniently extracted; the method comprises the following specific steps:

method 2.1: when the HRRP is converted into a two-dimensional image, the XYZ three-coordinate of the two-dimensional gray image is needed to be corresponded. To extract intensity information of HRRP scattering sites, the horizontal and vertical axes of HRRP can be defined as distance and scattering intensity, respectively, as compared to a two-dimensional image. The physical meanings of a two-dimensional image on the XY axis are all space dimensions, and the physical meaning on the Z axis is the gray value of the pixel point and can also be regarded as an intensity unit. Therefore, the HRRP can be regarded as a two-dimensional image with one dimension being 1 in the X axis or the Y axis, and the HRRP is directly used as the two-dimensional image for feature extraction.

Method 2.2: in order to extract the structural information of each scattering point on the HRRP curve, the horizontal axis and the vertical axis of the HRRP are respectively corresponding to the XY axes of the two-dimensional image, the whole HRRP image is divided into single pixel points according to distance units, the pixel point value of the curve is set to be 1, the other blank pixel point values are set to be 0, namely the vertical axis of the original HRRP is changed into the Y axis of the two-dimensional image, the horizontal axis is changed into the X axis of the two-dimensional image, and the HRRP is converted into a binary image with the Z-axis intensity value of 0 or 1.

Step 2, constructing a network structure combining a deep convolutional layer and a batch normalization layer, extracting deep noise characteristics, constructing a one-dimensional noise reduction convolutional neural network model of HRRP, and reducing noise of HRRP data under different signal-to-noise ratios; a residual error structure is adopted, the load of deep network learning is reduced, and meanwhile an HRRP converted image is reconstructed; the method comprises the following specific steps:

the one-dimensional noise-reducing convolutional neural network is mainly divided into three parts, the first part is formed by a layer of convolutional layers, and the number of convolutional cores is A ₁ Size B ₁ ×C ₁ The activation function is a RELU function, and the part mainly utilizes the convolution layer to generate a certain number of feature maps to be transmitted to the next part. The second part is the main part of the whole network, a Batch Normalization (BN) layer is added between the convolution layer and the RELU function to form a normalized convolution layer, the network depth of the second part is set as an F-layer structure according to the network noise reduction effect, and the number of convolution kernels is A ₂ Size B ₂ ×C ₂ . And the standardized convolutional layer corrects the input value of each layer of the activation function through the BN layer, and the input value is pulled back to the standard normal distribution with the mean value of 0 and the variance of 1 so as to improve the output sensitivity of the activation function and solve the problem of internal covariate shift. The correction process can be expressed as

Wherein x represents the output result of the convolutional layer, s ₁ And s ₂ As intermediate variables, μ and σ ² Respectively representing the mean value and the variance of x, wherein epsilon is a tiny positive number added to avoid the variance being 0, and gamma and beta are network learnable parameters which are respectively used for scaling and translating the normalized value to improve the expression capability of the network; when the parameters γ and β are initialized to σ and μ, respectively, the normalized variables are reduced to the original values. With the training of the network, the network learns the most appropriate gamma and beta.

The third part of the network is a convolution layer, and the number of convolution kernels is A ₃ Size B ₃ ×C ₃ The part is used for characteristic diagram reconstruction, and a residual error structure is adopted, so that the learning burden of a deep network is reduced, and the problem of gradient disappearance is avoided. The residual structure can be expressed as

x _L+1 ＝x _L +F(x _L ,W _L ) (5)

Wherein x is _L Is a feature of unit L, F (x) _L ,W _L ) Denotes the residual F (x) _L ) Network mapping of W _L Representing the weight parameter of the cell L. Through recursion, the feature x of any unit L can be obtained _L Is expressed as

Wherein x is _l Is a feature of unit l, F (x) _i ,W _i ) Denotes the residual F (x) _i ) Network mapping of W _i Representing the weight parameter of the cell i.

For back propagation, assuming the loss function is E, it can be obtained according to the chain rule of back propagation:

the above formula is divided into two parts, wherein

Not passing through the weight layer,

passing through a weight layer; the former ensures that the signal can be directly transmitted back to any shallow layer

Cannot be-1, and the above formula can avoid the gradient disappearance phenomenon.

Step 3, a feature dimension reduction extraction module is constructed, automatic feature extraction and feature selection are carried out on the sea surface target HRRP, the one-dimensional noise reduction convolutional neural network is connected with a classification identification network, a classification cross entropy and softmax classifier is adopted, a sea surface target HRRP intelligent classification identification model integrating noise reduction and classification is constructed, and the extracted features are classified; the method comprises the following specific steps:

the feature dimension reduction extraction module utilizes the multilayer convolution layers to extract features, the step length of each convolution layer is 1, and the filling modes are the same, namely the input dimension and the output dimension are the same. Meanwhile, the maximum pooling layer is utilized to perform feature dimension reduction, redundant information is eliminated, and effective information is reserved. The module adopts the mode of alternately connecting convolution layers and pooling layers, firstly two convolution layers are adopted, the number of convolution cores is A ₄ Size B ₄ ×C ₄ Then, a maximum pooling layer is followed, the step length is D, and the output dimension is reduced to 1/D of the input dimension; then, two convolution layers are connected, and the number of convolution kernels is A ₅ Size B ₅ ×C ₅ The subsequent maximum pooling layer with the step length of D is also connected to perform characteristic dimension reduction; finally, a convolution layer is formed, and the convolution layer reduces the number of convolution kernels to A ₆ Size of B ₆ ×C ₆ If the number of the convolution kernels of the layer is too large, a large number of features can be formed, and subsequent feature splicing is not convenient.

And splicing the features extracted by the modules by utilizing two full-connection layers, wherein the neuron number of the second full-connection layer is usually the same as the target category number. And then transmitting the output of the second layer full connection layer and the class label together to a loss function and a softmax classifier for classification and identification.

The loss function is the end point of the forward propagation and also the start point of the backward propagation. And selecting the classified cross entropy as a loss function, so that the network parameters are continuously changed towards the direction of minimizing the loss function in each round of training until the loss function is converged. The class cross entropy can be expressed as

Wherein Q represents the total number of samples after expansion, p _zj Denotes the probability, t, of discriminating the z-th sample as the j-th class _zj The jth value representing the z sample class label is presented in One-hot encoded form, e.g., the class label of class 2 of the 5 class samples is [0,1,0,0,0%]。

And performing class probability division on the output of the second fully-connected layer by using a softmax classifier, wherein the result output by the classifier is the probability of the sample being divided into each class, so that the target classification is realized. The class probability is calculated as follows

In the formula, H is the number of the neurons of the second full connecting layer; a is _zh The output value of the z th sample of the h th neuron of the second layer full connection layer is the input value of softmax; y is _zh The h output value of the z sample of the softmax classifier. Because the second layer fully-connected layer neurons count the same as the target class number (i.e., H = C), and y is _z1 +y _z1 +…+y _zH =1, the output value of the softmax classifier is the class probability distribution of the sample, i.e. p _zj ＝y _zh 。

Step 4, evaluating the classification performance of the model, comparing the results of the softmax classifier with the initial class labels, calculating indexes such as accuracy, precision or recall rate and the like, and evaluating the classification identification performance of the model according to the index calculation results; the method comprises the following specific steps:

the calculation process of three indexes of accuracy acc, accuracy P and recall R is expressed as

Wherein, TP represents the number of true positive examples, that is, the prediction result of the positive example is also the number of the positive examples; TN represents the number of true counterexamples, namely the prediction result of the counterexample sample is also the number of the counterexamples; FP represents the number of false positive examples, namely the number of positive examples of the prediction result of the negative example sample; FN represents the number of false counterexamples, namely the number of the positive example sample prediction results which are the counterexamples;

generally, the higher the accuracy and precision, the better the classification and identification performance of the model, and the recall ratio is an index contradictory to the precision, the high precision, the low recall ratio, and vice versa.

Compared with the background art, the invention has the beneficial effects that: 1) Aiming at a complex sea surface environment, an HRRP one-dimensional noise reduction convolution neural network model is constructed, noise reduction is carried out through an end-to-end neural network model, adverse effects of human factors are reduced, universality is high, and the method is suitable for noise reduction processing of different noise scenes; 2) A dimension reduction extraction module is constructed by utilizing multi-level convolution kernels and a maximum pooling layer, target HRRP characteristic information of different levels is extracted, and the capacity and the flexibility of the characteristic information are improved; 3) The HRRP intelligent classification and identification model of the sea surface target integrated with noise reduction and classification is constructed, on the basis of keeping the target identification accuracy under the condition of high signal-to-noise ratio, the identification accuracy of the sea surface target under the condition of low signal-to-noise ratio is improved, the robustness and the intelligent level of the sea surface target identification are enhanced, and the identification capability of a radar on the sea surface target under the complex environment is improved.

Drawings

Fig. 1 is a schematic diagram of an HRRP converted two-dimensional image.

FIG. 2 is a block diagram of the general structure of a one-dimensional distance image noise reduction convolutional neural network recognition model of a sea surface target according to the present invention.

FIG. 3 is a schematic view of a simulation model of a ship target of some type.

FIG. 4 is a comparison graph of the HRRP signal before and after noise reduction under the condition of 1dB signal-to-noise ratio.

FIG. 5 is a recognition confusion matrix for classifying and recognizing five types of sea surface targets under the condition of low signal-to-noise ratio.

Fig. 6 is an identification confusion matrix for classification identification of five types of sea surface targets by a Convolutional Neural Network (CNN) model under a low signal-to-noise ratio condition.

Fig. 7 is a line graph comparing the average recognition accuracy of CNN and the method of the present invention under different SNR conditions.

Fig. 8 is a schematic diagram of a binary image of

HRRP conversion

0 or 1.

Detailed Description

The invention is further described below with reference to the accompanying drawings. The present embodiments are to be considered as illustrative and not restrictive, and all changes and modifications that come within the spirit of the invention and the scope of the appended claims are desired to be protected.

In order to verify the effectiveness of the method of the present invention, the present embodiment provides two embodiments, the first embodiment uses method 2.1 to convert the HRRP into the two-dimensional image with X-axis dimension 1 for classification and identification, and the second embodiment uses method 2.2 to perform classification and identification on the binary image with 0 or 1 converted by the HRRP.

Example 1:

the specific implementation of example 1 is divided into the following steps:

step A-1: firstly, let M =200, randomly generate a NOISE sequence NOISE, calculate the effective power PS of the noiseless HRRP by using formula (2), estimate the effective power PN of the NOISE under the condition of a given SNR by using formula (1), normalize the random NOISE sequence NOISE to the effective power, and superimpose the normalized NOISE on the noiseless HRRP to complete the NOISE adding processing of the HRRP data. And then, adding noises with different signal-to-noise ratios according to the steps, wherein each signal-to-noise ratio is added for 5 times, and the data set is expanded by 5 times. And translating the HRRP data in the distance direction so as to expand the data set, wherein 10 distance units are translated each time, the translation is carried out 21 times totally, and the front data set and the rear data set are enlarged by 105 times totally. And finally, randomly disordering and recombining the whole group of data, and according to each type of target 8: the scale of 2 is divided into a training set and a test set.

Step A-2: as shown in fig. 1, HRRP is converted into a two-dimensional image, the physical meaning of HRRP on the horizontal axis is distance and the physical meaning on the vertical axis is the intensity of the scattering point. The physical meaning of a two-dimensional image on the XY axis is a space dimension, and the physical meaning on the Z axis is the gray value of the pixel point, and can also be regarded as an intensity unit. Therefore, the HRRP can be approximately regarded as a two-dimensional image with the X-axis dimension of 1, and from the principle analysis, the HRRP can be regarded as a two-dimensional image for classification and identification.

Step A-3: connecting the constructed one-dimensional noise reduction convolutional neural network with the feature dimension reduction extraction module, splicing the extracted features by using a full connection layer, transferring the spliced features to a loss function and a softmax classification function for back propagation learning and identification classification, and further constructing a sea surface target HRRP intelligent classification and identification model integrating noise reduction and identification, wherein the overall structure of the model is shown in FIG. 2. The values of the parameters of each layer of the network model are shown in table 1.

Table 1 values of network model parameters in example 1

Parameter(s)	Value taking	Parameter(s)	Value taking	Parameter(s)	Value taking	Parameter(s)	Value taking
								A ₁	64	A ₃	1	A ₅	32	D	2
B ₁	1	B ₃	1	B ₅	1	F	5
								C ₁	3	C ₃	3	C ₅	3	G	200
A ₂	32	A ₄	64	A ₆	16	H	5
								B ₂	1	B ₄	1	B ₆	1
C ₂	3	C ₄	5	C ₆	3

Step A-4: after the network model structure is constructed, selecting a proper network learning round epoch and the number of learnable data batch of each round, and expanding the learning process of the network model. The learning process of the network model is divided into two stages: and in the training stage, internal parameters of the network are continuously adjusted through each round of training, so that the loss function is continuously changed towards the minimum direction, and in the testing stage, the trained model parameters are used for carrying out classification and identification on HRRP data concentrated in the testing process, so that the learning effect of the network model is verified.

The effect of the method for identifying the one-dimensional distance image noise reduction convolution neural network of the sea surface target provided by the invention can be further illustrated by the following simulation result.

Description of simulation data: the data used in this embodiment are HRRP data of five types of targets obtained by using computer simulation software, and the simulation software radar parameters are set as follows: the radar has the center frequency of 10GHz, the bandwidth of 80MHz, the azimuth angle range of 0-360 degrees, the interval of 1 degree, 200 distance units and the resolution of 1.875 meters. Finally, sample data of 360 azimuth angles of each type of targets is obtained, the data length of each azimuth angle is 200, and fig. 3 is a schematic diagram of a simulation model of one type of ship targets.

Fig. 4 (a) is HRRP under SNR =1dB condition, and it can be seen from fig. 4 (a) that the target signal is substantially drowned in noise due to low signal-to-noise ratio. Fig. 4 (b) shows the HRRP after denoising using the one-dimensional denoising convolutional network, in which the signal-to-noise ratio of the HRRP after denoising is significantly improved and the target portion is more clearly visible than that in fig. 4 (a). In a whole view, the one-dimensional noise reduction convolution neural network constructed by the invention can effectively inhibit the interference of random noise and lays a good foundation for accurately identifying the subsequent sea surface target.

Fig. 5 (a) to (c) are identification confusion matrices for classifying and identifying five types of sea surface targets under the conditions of SNR =0.1dB, 0.5dB and 1dB, respectively, and it can be seen that the method of the present invention has a high identification rate for the sea surface targets under the condition of low signal-to-noise ratio. Fig. 6 (a) to (c) are recognition confusion matrices for classifying and recognizing five types of sea surface targets by the classical CNN model under the conditions of SNR =0.1dB, 0.5dB and 1dB, respectively, and it can be seen from a comparison of the results in fig. 5 that the recognition effect of the CNN model is inferior to that of the method of the present invention under the condition of low signal-to-noise ratio. Fig. 7 is a comparison line graph of average recognition accuracy of the two methods in fig. 5 and fig. 6 under different SNR conditions, from which it can be seen more intuitively that the average recognition accuracy of the two methods is continuously increased along with the improvement of the SNR, but the method of the present invention can maintain better recognition rate under different SNR conditions, and has better recognition robustness and recognition performance.

Example 2:

the specific implementation of example 2 is divided into the following steps:

first, referring to step a-1 in example 1, HRRP is subjected to preprocessing such as noise addition, expansion, and division, and then performed according to step B-2 described below.

Step B-2: the HRRP is converted into a binary image of 0 or 1 by adopting a method 2.2, and the structural characteristic information of the HRRP curve is explored. Firstly, dividing the whole HRRP image into small pixel points, setting the pixel point where the curve is located as 1 and setting the rest blank pixel points as 0, and finally converting the longitudinal axis of the HRRP into the Y axis of the two-dimensional image and converting the transverse axis of the HRRP into the X axis of the two-dimensional image to complete the image conversion of the HRRP, wherein the conversion effect is shown in fig. 8.

And then, referring to the step A-3 and the step A-4 in the embodiment 1, building a deep learning network model integrating noise reduction and identification, wherein the values of parameters of each layer of the network model are shown in a table 2, and then training the network model, so as to perform classification identification and performance evaluation on the sea surface target under the condition of low signal-to-noise ratio.

Table 2 values of network model parameters in example 2

Parameter(s)	Value taking	Parameter(s)	Value taking	Parameter(s)	Value taking	Parameter(s)	Value taking
								A ₁	64	A ₃	1	A ₅	32	D	2
B ₁	3	B ₃	3	B ₅	5	F	5
								C ₁	3	C ₃	3	C ₅	3	G	200
A ₂	32	A ₄	64	A ₆	16	H	5
								B ₂	3	B ₄	5	B ₆	3
C ₂	3	C ₄	5	C ₆	3

Claims

1. The method for identifying the one-dimensional distance image noise reduction convolution neural network of the sea surface target is characterized by comprising the following steps of:

step 1, acquiring noise-containing target HRRP data in various typical sea surface environments according to the sea surface target identification requirements, carrying out noise addition and distance translation expansion on the target HRRP data, and constructing a training set and a test set of a model for an expanded data set according to a fixed proportion; then, according to the characteristics of the sea surface target HRRP, the HRRP is converted into a two-dimensional image, and the physical meaning of the coordinate axis of the two-dimensional image is given according to the HRRP information, so that different characteristic information can be conveniently extracted;

step 2, constructing a network structure combining a deep convolutional layer and a batch normalization layer, extracting deep noise characteristics, constructing a one-dimensional noise reduction convolutional neural network model of HRRP, and reducing noise of HRRP data under different signal-to-noise ratios; a residual error structure is adopted, the load of deep network learning is reduced, and meanwhile an HRRP converted image is reconstructed;

the one-dimensional noise reduction convolutional neural network model is divided into three parts, wherein the first part is a convolutional layer, and an RELU function is selected as an activation function; the second part adds a batch normalization layer between the convolution layer and the RELU function to form a standardized convolution layer, and sets the network depth of the part as an F-layer structure according to the noise reduction effect; the third part is a convolution layer;

step 3, a feature dimension reduction extraction module is constructed, automatic feature extraction and feature selection are carried out on the sea surface target HRRP, the one-dimensional noise reduction convolutional neural network is connected with a classification identification network, a classification cross entropy and softmax classifier is adopted, a sea surface target HRRP intelligent classification identification model integrating noise reduction and classification is constructed, and the extracted features are classified;

constructing a characteristic dimension reduction extraction module by a connection mode of alternately combining the convolution layer and the pooling layer; connecting two full-connection layers at the output end of the module for characteristic splicing; sequentially connecting the one-dimensional noise reduction convolutional neural network module, the characteristic dimension reduction extraction module and the full-connection layer end to obtain a noise reduction and classification integrated sea surface target high-resolution one-dimensional distance image intelligent classification identification model;

and 4, acquiring high-resolution one-dimensional range profile data, importing the high-resolution one-dimensional range profile data into the intelligent classification identification model, performing back propagation through a loss function, continuously updating model parameters until the loss function is converged, performing class classification by using a classifier, comparing a classification result with an initial class label, calculating the accuracy, precision or recall rate, and evaluating the classification identification performance of the model according to an index calculation result.

2. The method for identifying the one-dimensional distance image noise reduction convolutional neural network of the sea surface target according to claim 1, wherein before the data is imported into the intelligent classification identification model, the data is preprocessed, and the method specifically comprises the following steps:

firstly, carrying out noise processing on acquired data; generating a random NOISE sequence NOISE, estimating the effective power PN of the random NOISE sequence according to the set SNR:

in the above equation, PS represents the effective power of the noiseless data;

combining the effective power PN, standardizing the random NOISE sequence NOISE to obtain a NOISE sequence NOISE to be added, wherein the standardization process is as follows:

noise＝sqrt(PN)/std(NOISE)*NOISE

wherein sqrt (·) is a square root function, std (·) is a standard deviation calculation function;

secondly, expanding the data set after noise addition; firstly, adding noise for a plurality of times to original sample data under the condition of a given SNR by using the noise adding processing mode; then, circularly translating the data subjected to the noise processing in the distance direction, and expanding a data set by using the data translated each time; then randomly disordering and recombining the whole group of data sets, then dividing the data sets into a training set and a testing set according to the proportion of A to B of each type of target;

finally, converting the HRRP into a two-dimensional image; corresponding to XYZ three coordinates of the two-dimensional gray image, the horizontal axis and the vertical axis of the HRRP are respectively defined as the distance and the scattering intensity, the HRRP is regarded as a two-dimensional image with one dimension being 1 in the X axis or the Y axis, and the HRRP is further directly regarded as the two-dimensional image for feature extraction.

3. The method for identifying the one-dimensional distance image noise reduction convolution neural network of the sea surface target according to claim 2, wherein the method for converting the HRRP into the two-dimensional image is replaced by the following method:

the method comprises the steps of enabling the horizontal axis and the vertical axis of the HRRP to correspond to the XY-axis distance dimension of a two-dimensional image, dividing the whole HRRP image into single pixel points according to distance units, setting the pixel point value of a curve to be 1, setting the values of the rest blank pixel points to be 0, namely, changing the vertical axis of the original HRRP into the Y axis of the two-dimensional image, changing the horizontal axis into the X axis of the two-dimensional image, and further converting the HRRP into a binary image with a Z-axis intensity value of 0 or 1.

4. The method for identifying the one-dimensional distance image noise-reduction convolutional neural network of the sea surface target according to claim 1, wherein when a one-dimensional noise-reduction convolutional neural network module is constructed, a residual structure is adopted, and the input of the module is directly transmitted to the output, specifically:

wherein x is _l Is a feature of unit l, x _L Is a feature of the unit L, x _i Characteristic of cell i, F (x) _i ,W _i ) Denotes the residual F (x) _i ) Network mapping of W _i A weight parameter representing a cell i;

for back propagation, the loss function is E, which is obtained according to the chain rule of back propagation:

in the above formula

Not passed through the weight layer,

passing through the weight layer.

5. The method for identifying the one-dimensional distance image noise reduction convolutional neural network of the sea surface target according to claim 1, wherein the selection of the loss function and the classifier specifically comprises:

selecting a classification cross entropy as a loss function, wherein the classification cross entropy is expressed as follows:

wherein Q represents the total number of samples after expansion, p _zj Denotes the probability, t, of discriminating the z-th sample as the j-th class _zj A jth value representing a z sample class label, wherein the class label is displayed in a form of One-hot coding;

the classifier selects a softmax classifier to perform class probability division on the output of the second full-connection layer, and the calculation process of the class probability is as follows:

in the formula, H is the number of the neurons of the second full connecting layer; a is _zh The output value of the z th sample of the h th neuron of the second layer full connection layer is the input value of softmax; y is _zh And obtaining the h output value of the z sample of the softmax classifier, namely the class distribution probability.