Sea surface floating target identification method based on recurrence plot

Abstract

The invention relates to a sea surface floating target identification method based on a recurrence plot, and belongs to the technical field of radar target characteristic and identification. Comprising the following steps: 1) acquiring a target radar echo sequence; step 2) generation and processing of a recurrence plot; and step 3), constructing and identifying the CNN-LSTM network. According to the invention, sea surface floating target identification can be completed under the condition of short observation time.

Description

Technical Field

[0001] The invention relates to a method for identifying sea surface floating targets based on a recursive graph, and belongs to the technical field of radar target characteristics and identification. Background Art

[0002] Accurately identifying and classifying sea surface targets helps determine the purpose of the target and take appropriate countermeasures. In civil shipping, accurate identification of various marine floating objects can help passenger and cargo ships ensure route safety and avoid collisions. In the development of marine resources, it can provide accurate target location information for fisheries, energy collection, etc., and improve the efficiency of marine operations. In the military, the identification of passive interference such as diagonal reflectors helps commanders make correct decisions and commands during combat. Most existing research currently extracts target features from one-dimensional radar echo sequences and their transform domains. In most cases, differential feature extraction requires long-term accumulation of target echo sequences, otherwise the proposed features are insufficient to describe the difference between floating targets on the sea surface and ship targets.

[0003] Therefore, how to effectively distinguish the characteristic differences between targets within a short observation time is the key to solving the above problem. Currently, there is no relevant public technology to solve the above technical problem. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the above-mentioned prior art and provide a method for identifying floating targets on the sea surface based on a recursive graph, which can complete the identification of floating targets on the sea surface under short observation time conditions.

[0005] The present invention provides a method for identifying floating targets on the sea surface based on a recursive graph, which is special in that it comprises the following steps:

[0006] Step 1) Acquisition of target radar echo sequences: The test radar emits electromagnetic waves to detect ships and buoy targets floating on the sea surface. The echo signal sequences of the ship targets and the sea surface buoys are divided into batches to obtain the pulse time series used to generate the recurrence graph;

[0007] Step 2) Generate and process recursion graphs: For the divided pulse time series of ship targets and buoy targets, the autocorrelation coefficient method and the geometric invariant method are used to determine the delay time and embedding dimension, respectively. Then, a recursion graph is generated for each pulse series, and labels are set according to the ship and buoy categories.

[0008] Step 3) Construction and identification of the CNN-LSTM network: Pulse sequences are divided from the echoes of ship targets and buoy targets to generate recursive graphs, forming the required training set and test set feature samples. The training set samples are input into the CNN-LSTM network for training. Finally, the test set samples are input into the trained CNN-LSTM network to identify ship targets and buoy targets.

[0009] Preferably, the specific steps of step 1) are: the sea detection radar emits electromagnetic waves to detect ship targets and buoy targets floating on the sea surface. Assuming that a pulse sequence of length L is received on the distance unit of a ship target and a buoy target, its N sequences are divided into a group, then each target can generate L / N recursive graphs.

[0010] Preferably, the specific steps of step 2) are:

[0011] Phase space reconstruction technology has two key parameters: the embedding dimension m and the delay time τ. The autocorrelation function is a method to calculate the delay time. For a time series x(n) of length N, its autocorrelation function is written as follows:

[0012]

[0013] When the autocorrelation function drops to 1-e of the initial value R(0) -1 When times, the obtained time τ is also the delay time of reconstructing the phase space;

[0014] In order to determine the embedding dimension m, the common method in practical applications is to calculate certain geometric invariants such as the correlation dimension. For a time series x(n) of length N, after determining the delay time τ, a smaller embedding dimension m is given first, and the sequence after phase space reconstruction is:

[0015] X i =(x i ,x i+τ ,...,x i+(m-1)τ ) (2)

[0016] The distance between any two points in the reconstructed phase space is:

[0017] d ij =||X i -X j || (3)

[0018] Where i, j = 1, 2, ... N-(m-1)τ, ‖·‖ represents the two-norm of the vector, and a correlation function is given:

[0019]

[0020] where θ(z) is the Heaviside function:

[0021]

[0022] r is the distance threshold, C(r) is a cumulative distribution function. For a certain appropriate range of r, the correlation dimension d and the cumulative distribution function C(r) should satisfy a log-linear relationship, that is, d(m) = lnC(r) / lnr. The corresponding correlation dimension estimate is obtained by fitting. The correlation dimension is calculated by gradually increasing the value of the embedding dimension m until the minimum m at which the correlation dimension stops changing is the required embedding dimension. Therefore, the recurrence graph is formed as follows:

[0023] 1) For a time series x(n) of length N, first determine its delay time. When the autocorrelation function drops to 1-e of the initial value R(0), -1 times, that is, R(τ)=(1-e -1 )R(0), the obtained time τ is also the delay time of reconstructing the phase space;

[0024] 2) After determining the delay time, gradually increase the value of the embedding dimension m to calculate the correlation dimension until the minimum m at which the correlation dimension stops changing is the desired embedding dimension;

[0025] 3) After determining the embedding dimension m and the delay time τ, the sequence after phase space reconstruction is obtained;

[0026] 4) Calculate the distance between any two points in the reconstructed phase space;

[0027] 5) The distance matrix obtained is the threshold-free recursive graph.

[0028] Preferably, the specific steps of step 3) are:

[0029] The neural network structure is used to extract image features. The pulse time series received by the radar is first converted into a recursive graph. The recursive graph is then applied to the CNN-LSTM network to automatically extract features and realize the recognition and detection of floating targets at sea.

[0030] The LSTM model is used in sequence prediction, time series analysis and other fields. For an LSTM memory unit, at time t, the input of the memory unit includes the hidden layer state variable h at the previous moment. t-1 , memory unit state variable c t-1 and the current input information x t ; Then the model passes through the forget gate f i , input gate i t , output gate o t Get the hidden layer state variable h at time t t and the memory cell state variable ct ;Finally h t It will be passed to the output layer to generate the calculation result y of LSTM at time t t , and c t Pass them together for calculation at the next moment.

[0031] During the calculation process at time t, the first thing to be calculated is the forget gate f t , the forget gate determines which information is discarded from the unit, and determines whether to retain or forget information by looking at the hidden state of the previous moment and the current input:

[0032] f t =σ(U f x t +W f h t-1 +b f ) (6)

[0033] Where U f ,W f and b f is the adjustable parameter matrix or vector of the forget gate, and σ is the Sigmoid activation function:

[0034]

[0035] Then, calculate the input gate i t , in the input gate "candidate value" Once created, it may be added to the state, and then decide which values need to be updated by looking at the hidden state of the previous moment and the current input:

[0036] i t =σ(U i x t +W i h t-1 +b i ) (8)

[0037] Where U i ,W i and b i is the adjustable parameter matrix or vector of the input gate, and the new information The calculation formula is as follows:

[0038]

[0039] Next, use the results calculated above to update the neuron state. The calculation formula is as follows:

[0040]

[0041] Where, U c ,Wc and b c is an adjustable parameter matrix or vector, and ⊙ represents the matrix element product.

[0042] Then, calculate the output gate o t The output gate determines which information to generate the hidden layer state variable h based on the latest cell state. t ,The hidden state is then passed to the next time point or used to generate the output, which is calculated as follows:

[0043] o t =σ(U o x t +W o h t-1 +b o ) (11)

[0044] h t =o t ⊙tanh(c t ) (12)

[0045] Where, U o ,W o and b o is the adjustable parameter matrix or vector of the output gate;

[0046] Finally, h t The final output y of LSTM at time t is obtained after calculation. t .

[0047] y t =W d h t +b d (13)

[0048] Where: W d and b d is the weight parameter matrix or vector of the output layer.

[0049] The CNN-LSTM network combines convolutional neural networks with LSTM layers. The LSTM layer is connected after the traditional neural network pooling layer. The convolutional layer and LSTM layer are used to train the data. The CNN-LSTM network structure is as follows:

[0050] 1) First, the recursive graph training set is input into the CNN-LSTM network input layer;

[0051] 2) After the input layer, the first network layer is connected in sequence to the convolution layer to extract input image features, the batch normalization layer to accelerate network training, the activation function layer to perform nonlinear transformation, the maximum pooling layer to reduce the dimension of the feature map, and the dropout layer to improve the generalization ability of the model;

[0052] 3) Then connect the convolution layer, batch normalization layer, activation function layer, maximum pooling layer and dropout layer to the second layer of the network;

[0053] 4) The obtained feature map is flattened by the flatten layer and then input into the LSTM layer;

[0054] 5) After passing through the LSTM layer, connect to the fully connected layer and the dropout layer;

[0055] 6) Finally, connect to the fully connected layer, softmax layer, and classification layer to obtain the recognition result.

[0056] Compared with the existing technology, the method for identifying floating targets on the sea surface based on recursive graph proposed in the present invention has the following beneficial effects:

[0057] (1) The method proposed in the present invention utilizes recursive graphs to directly perform feature analysis on the entire or local image, and more directly reflects the spatial structure and nonlinearity of the image's constituent elements.

[0058] (2) The method proposed in the present invention has a high recognition accuracy of over 90% for targets with large mass differences, such as ship targets and buoy targets.

[0059] (3) The method proposed in this invention can clearly display the mutation points and state changes in the data, and reveal the characteristics of time series similarity, periodicity, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the time domain echogram of the ship target and the buoy target used in the present invention;

[0061] Figure 2 The recursion diagrams of the ship target and the buoy target used in the present invention are shown in FIG. 1 : (a) is the recursion diagram of the ship target obtained from the measured data; (b) is the recursion diagram of the buoy target obtained from the measured data;

[0062] Figure 3 This is the recognition result diagram after CNN-LSTM network training; in the figure: (a) is the recognition accuracy diagram of ships and buoys; (b) is the sample classification result diagram of ships and buoys. DETAILED DESCRIPTION

[0063] For better understanding and implementation, a specific implementation detailed description of a method for identifying floating targets on the sea surface based on a recursive graph of the present invention is given below in conjunction with the accompanying drawings. It should be noted that this embodiment introduces the method proposed by the present invention by taking the identification of ship targets and buoy targets as an example.

[0064] A method for identifying floating targets on the sea surface based on a recursive graph in this embodiment includes the following steps:

[0065] 1) Collection of target radar echo sequence

[0066] The sea detection radar emits electromagnetic waves to detect the ship targets and buoy targets floating on the sea surface, and the obtained time domain echo map is as shown in the attached figure. Figure 1 As shown in Figure 2, assuming that a pulse sequence of length L is received at the range unit of a ship target and a buoy target, and its N sequences are divided into a group, then each target can generate L / N recursive graphs.

[0067] 2) Generation and processing of recursive graphs

[0068] A recurrence plot (RP) is an image obtained from a time series that represents the distance between each time point. Due to the complex weather environment and sea clutter on the sea surface, the echo signals received by marine radars are mostly non-stationary. Recurrence plots are a common method for processing non-stationary signals. They can reconstruct the time series of echo pulses collected by the radar through phase space and display them as two-dimensional graphics. The threshold-free recurrence plot is not restricted by the threshold, and the pixel points of the graph directly reflect the distance between the two points. It can retain more information of the signal and avoid losing a large amount of detailed features. This example uses the N pulse time series of a ship target as an example to introduce how to generate a recurrence plot from a time series.

[0069] Phase space reconstruction technology has two key parameters: the embedding dimension m and the delay time τ. The autocorrelation function is a relatively simple method to calculate the delay time. For a time series x(n) of length N, its autocorrelation function can be written as follows:

[0070]

[0071] When the autocorrelation function drops to 1-e of the initial value R(0) -1 When the time τ is times, the obtained time τ is the delay time of reconstructing the phase space.

[0072] In order to determine the embedding dimension m, the usual method in practical applications is to calculate certain geometric invariants such as the correlation dimension. For a time series x(n) of length N, after determining the delay time τ, a smaller embedding dimension m is first given, and the sequence after phase space reconstruction is:

[0073] X i =(x i ,x i+τ ,...,x i+(m-1)τ ) (15)

[0074] The distance between any two points in the reconstructed phase space is

[0075] d ij =||Xi -X j || (16)

[0076] Where i, j = 1, 2, ... N-(m-1)τ, ‖·‖ represents the bi-norm of the vector. Given a correlation function:

[0077]

[0078] Where θ(z) is the Heaviside function,

[0079]

[0080] r is the distance threshold, and C(r) is a cumulative distribution function. For a certain appropriate range of r, the correlation dimension d and the cumulative distribution function C(r) should satisfy a log-linear relationship, i.e., d(m) = lnC(r) / lnr. The corresponding correlation dimension estimate is then obtained by fitting. The correlation dimension is calculated by gradually increasing the value of the embedding dimension m until the minimum m at which the correlation dimension stops changing is the desired embedding dimension. Therefore, the recurrence graph is formed as follows:

[0081] 1) For a time series x(n) of length N, first determine its delay time. When the autocorrelation function drops to 1-e of the initial value R(0), -1 times, that is, R(τ)=(1-e -1 )R(0), the obtained time τ is also the delay time of reconstructing the phase space.

[0082] 2) After determining the delay time, gradually increase the value of the embedding dimension m to calculate the correlation dimension until the minimum m at which the correlation dimension stops changing is the desired embedding dimension.

[0083] 3) After determining the embedding dimension m and the delay time τ, the sequence after phase space reconstruction is obtained.

[0084] 4) Calculate the distance between any two points in the reconstructed phase space.

[0085] 5) The distance matrix obtained is the threshold-free recursive graph.

[0086] The recursive diagram of ship targets and buoy targets obtained from the measured data is shown in the attached figure. Figure 2As shown in the figure, the recurrence plot for the ship target has many horizontal and vertical lines, while the recurrence plot for the buoy is more irregular. This is because the ship's tonnage is heavier, and the sway amplitude caused by the impact of waves is weak. The echo sequence values collected by the radar do not vary much, so the sequences after phase space reconstruction are relatively close, and the distance between any two points is approximately zero. Therefore, the recurrence plot has a dense appearance of parallel lines. On the other hand, the channel buoy is lighter and is more affected by waves. The radar echo sequence values fluctuate greatly, resulting in a large difference in the distance between any two points after phase space reconstruction. Therefore, the difference between the two types of targets can be more clearly seen from the recurrence plot, providing differentiating features for distinguishing and identifying the two types of targets.

[0087] 3) Construction and identification of CNN-LSTM network

[0088] With the continuous development of artificial intelligence, methods such as neural networks have shown promising results in target recognition and classification. Neural network models can very effectively extract local features from images and automatically learn complex image distributions through multi-layer nonlinear transformations. Compared to traditional feature extractor design methods, neural networks can learn the essential characteristics of images while also having higher processing efficiency. Therefore, the present invention utilizes a neural network structure to extract image features. The pulse time series received by the radar is first converted into a recursive graph, which is then applied to a CNN-LSTM network for automatic feature extraction, enabling the recognition and detection of floating targets at sea.

[0089] Among them, the LSTM model is widely used in the fields of sequence prediction, time series analysis, etc. For an LSTM memory unit, at time t, the input of the memory unit includes the hidden layer state variable h at the previous moment t-1 , memory unit state variable c t-1 and the current input information x t ; Then the model passes through the forget gate f i , input gate i t , output gate o t Get the hidden layer state variable h at time t t and the memory cell state variable c t ;Finally h t It will be passed to the output layer to generate the calculation result y of LSTM at time t t , and c t Pass them together for calculation at the next moment.

[0090] During the calculation process at time t, the first thing to be calculated is the forget gate f t The forget gate determines which information is discarded from the unit, and determines whether to retain or forget information by looking at the hidden state of the previous moment and the current input.

[0091] f t=σ(U f x t +W f h t-1 +b f ) (19)

[0092] Where U f ,W f and b f is the adjustable parameter matrix or vector of the forget gate, and σ is the Sigmoid activation function:

[0093]

[0094] Then, calculate the input gate i t . In the input gate, the "candidate value" Once created, it may be added to the state, and then decide which values need to be updated by looking at the hidden state of the previous moment and the current input.

[0095] i t =σ(U i x t +W i h t-1 +b i ) (twenty one)

[0096] Where U i ,W i and b i is the adjustable parameter matrix or vector of the input gate; new information The calculation formula is as follows:

[0097]

[0098] Next, use the results calculated above to update the neuron state. The calculation formula is as follows:

[0099]

[0100] Where, U c ,W c and b c is an adjustable parameter matrix or vector, and ⊙ represents the matrix element product.

[0101] Then, calculate the output gate o t The output gate determines which information to generate the hidden layer state variable h based on the latest cell state. t ,The hidden state is then passed to the next time point or used to generate output. Its calculation formula is as follows:

[0102] o t =σ(U o x t +Wo h t-1 +b o ) (twenty four)

[0103] h t =o t ⊙tanh(c t ) (25)

[0104] Where, U o ,W o and b o is the adjustable parameter matrix or vector of the output gate;

[0105] Finally, h t The final output y of LSTM at time t is obtained after calculation. t .

[0106] y t =W d h t +b d (26)

[0107] Where: W d and b d is the weight parameter matrix or vector of the output layer.

[0108] The CNN-LSTM network combines a convolutional neural network with an LSTM layer. The LSTM layer is connected after the traditional neural network pooling layer, and the convolutional and LSTM layers are used to train the data. The CNN-LSTM model leverages the strengths of both CNNs and LSTMs, capturing both local features in an image and the temporal relationships between sequences. This allows it to clearly reflect the correlation differences between the time-domain echo sequences of two target types, while simultaneously considering both temporal and spatial information in the image. It also effectively prevents problems such as overfitting and gradient explosion. The CNN-LSTM network structure is as follows:

[0109] 1) First, input the recursive graph training set into the CNN-LSTM network input layer

[0110] 2) After the input layer, the first network layer is connected in sequence to the convolution layer to extract input image features, the batch normalization layer to accelerate network training, the activation function layer to perform nonlinear transformation, the maximum pooling layer to reduce the dimension of the feature map, and the dropout layer to improve the generalization ability of the model.

[0111] 3) Then connect the convolution layer, batch normalization layer, activation function layer, maximum pooling layer and dropout layer to form the second network layer.

[0112] 4) The obtained feature map is flattened by the flatten layer and then input into the LSTM layer.

[0113] 5) After passing through the LSTM layer, connect to the fully connected layer and the dropout layer.

[0114] 6) Finally, connect to the fully connected layer, softmax layer, and classification layer to obtain the recognition result.

[0115] After the CNN-LSTM network training is completed, the test set recursive graph is input into the network, and the recognition results are shown in the attached figure. Figure 3 As shown in the figure, the recognition accuracy of ship target 1 and ship target 2 reaches 99.1% and 100% respectively, which can effectively distinguish between ships and buoys.

Claims

1. A method for identifying floating targets on the sea surface based on recursive graph, characterized in that The following steps are involved: Step 1) Acquisition of target radar echo sequences: The test radar emits electromagnetic waves to detect ships and buoy targets floating on the sea surface. The echo signal sequences of the ship targets and the sea surface buoys are divided into batches to obtain the pulse time series used to generate the recurrence graph; Step 2) Generate and process recursion graphs: For the divided pulse time series of ship targets and buoy targets, the autocorrelation coefficient method and the geometric invariant method are used to determine the delay time and embedding dimension, respectively. Then, a recursion graph is generated for each pulse series, and labels are set according to the ship and buoy categories. Step 3) Construction and identification of the CNN-LSTM network: Pulse sequences are divided from the echoes of ship targets and buoy targets to generate recursive graphs, forming the required training set and test set feature samples. The training set samples are input into the CNN-LSTM network for training. Finally, the test set samples are input into the trained CNN-LSTM network to identify ship targets and buoy targets.

2. A method for identifying floating targets on the sea surface based on a recursive graph according to claim 1, characterized in that The specific steps of step 1) are: the sea detection radar emits electromagnetic waves to detect ship targets and buoy targets floating on the sea surface. Assuming that a pulse sequence of length L is received on the distance unit of a ship target and a buoy target, its N sequences are divided into a group, then L / N recursive graphs can be generated for each target.

3. A method for identifying floating targets on the sea surface based on a recursive graph according to claim 1, characterized in that In step 2), for the divided pulse time series of the ship target and the buoy target, the specific steps of using the autocorrelation coefficient method to determine the delay time and the geometric invariant method to determine the embedding dimension are as follows: Phase space reconstruction technology has two key parameters: the embedding dimension m and the delay time τ. The autocorrelation function is a method to calculate the delay time. For a time series x(n) of length N, its autocorrelation function is written as follows: When the autocorrelation function drops to 1-e of the initial value R(0) -1 When times, the obtained time τ is also the delay time of reconstructing the phase space; For a time series x(n) of length N, after determining the delay time τ, a smaller embedding dimension m is given first, and the sequence after phase space reconstruction is: X i =(x i ,x i+τ ,...,x i+(m-1)τ ) (28) The distance between any two points in the reconstructed phase space is: d ij =||X i -X j || (29) Where i, j = 1, 2, ... N-(m-1)τ, ‖·‖ represents the bi-norm of the vector, and a correlation function is given: where θ(z) is the Heaviside function: r is the distance threshold, C(r) is a cumulative distribution function. For a certain appropriate range of r, the correlation dimension d and the cumulative distribution function C(r) should satisfy a log-linear relationship, that is, d(m) = lnC(r) / lnr. The corresponding correlation dimension estimate is then obtained by fitting. The correlation dimension is calculated by gradually increasing the value of the embedding dimension m until the minimum m at which the correlation dimension stops changing is the desired embedding dimension.

4. A method for identifying floating targets on the sea surface based on a recursive graph according to claim 1, characterized in that The method for generating the recursive graph in step 2) is as follows: 1) For a time series x(n) of length N, first determine its delay time. When the autocorrelation function drops to 1-e of the initial value R(0), -1 times, that is, R(τ)=(1-e-1R0, the obtained time τ is also the delay time of reconstructing the phase space; 2) After determining the delay time, gradually increase the value of the embedding dimension m to calculate the correlation dimension until the minimum m at which the correlation dimension stops changing is the desired embedding dimension; 3) After determining the embedding dimension m and the delay time τ, the sequence after phase space reconstruction is obtained; 4) Calculate the distance between any two points in the reconstructed phase space; 5) The distance matrix obtained is the threshold-free recursive graph.

5. A method for identifying floating targets on the sea surface based on a recursive graph according to claim 1, characterized in that In step 3), the specific steps of dividing the pulse sequence from the ship target and the buoy target echo to generate a recursive graph and forming the required training set and test set feature samples are as follows: Using a neural network structure to extract image features, the radar pulse time series is first converted into a recursive graph, which is then applied to the CNN-LSTM network to automatically extract features and achieve recognition and detection of floating targets at sea. The LSTM model is used in sequence prediction, time series analysis and other fields. For an LSTM memory unit, at time t, the input of the memory unit includes the hidden layer state variable h at the previous moment. t-1 , memory unit state variable c t-1 and the current input information x t ; Then the model passes through the forget gate f i , input gate i t , output gate o t Get the hidden layer state variable h at time t t and the memory cell state variable c t ;Finally h t It will be passed to the output layer to generate the calculation result y of LSTM at time t t , and c t Pass them together to the next moment for calculation; During the calculation process at time t, the first thing to be calculated is the forget gate f t , the forget gate determines which information is discarded from the unit, and determines whether to retain or forget information by looking at the hidden state of the previous moment and the current input: f t =σ(U f x t +W f h t-1 +b f ) (32) Where U f ,W f and b f is the adjustable parameter matrix or vector of the forget gate, and σ is the Sigmoid activation function: Then, calculate the input gate i t , in the input gate "candidate value" Once created, it may be added to the state, and then decide which values need to be updated by looking at the hidden state of the previous moment and the current input: i t =σ(U i x t +W i h t-1 +b i ) (34) Where U i ,W i and b i is the adjustable parameter matrix or vector of the input gate, and the new information The calculation formula is as follows: Next, use the results calculated above to update the neuron state. The calculation formula is as follows: Where, U c ,W c and b c is an adjustable parameter matrix or vector, ⊙ represents the matrix element product; Then, calculate the output gate o t , the output gate determines which information to generate the hidden layer state variable h based on the latest cell state t ,The hidden state is then passed to the next time point or used to generate the output, which is calculated as follows: the t =σ(U o x t +W o h t-1 +b o ) (37) h t =o t ⊙tanh(c t ) (38) Where, U o ,W o and b o is the adjustable parameter matrix or vector of the output gate; Finally, h t The final output y of LSTM at time t is obtained after calculation. t ; y t =W d h t +b d (39) Where: W d and b d is the weight parameter matrix or vector of the output layer.

6. A method for identifying floating targets on the sea surface based on a recursive graph according to claim 1, characterized in that In step 3), the training set samples are input into the CNN-LSTM network for training, and finally the test set samples are input into the trained CNN-LSTM network. The specific steps for identifying ship targets and buoy targets are as follows: The CNN-LSTM network combines convolutional neural networks with LSTM layers. The LSTM layer is connected after the traditional neural network pooling layer. The convolutional layer and LSTM layer are used to train the data. The CNN-LSTM network structure is as follows: 1) First, the recursive graph training set is input into the CNN-LSTM network input layer; 2) After the input layer, the first network layer is connected in sequence to the convolution layer to extract input image features, the batch normalization layer to accelerate network training, the activation function layer to perform nonlinear transformation, the maximum pooling layer to reduce the dimension of the feature map, and the dropout layer to improve the generalization ability of the model; 3) Then connect the convolution layer, batch normalization layer, activation function layer, maximum pooling layer and dropout layer to the second layer of the network; 4) The obtained feature map is flattened by the flatten layer and then input into the LSTM layer; 5) After passing through the LSTM layer, connect to the fully connected layer and the dropout layer; 6) Finally, connect to the fully connected layer, softmax layer, and classification layer to obtain the recognition result.

Images