Radar sea surface target classification method based on channel attention and dynamic rectification function

Abstract

The invention discloses a radar sea-surface target classification method based on channel attention and a dynamic rectification function, and the method comprises the steps: 1, carrying out the classification of a radar sea-surface target based on the classification scene and classification demands of the radar sea-surface target, and reference to the completeness and balance of data set construction; based on a self-constructed or existing radar, samples are extracted from the sea exploration target classification data set, and a training data set and a test data set are constructed; step 2, using a radar sea surface target classification network based on channel attention and a dynamic rectification function to train the training data set; 3, testing the classification network and the network weight coefficient obtained by training, and recording a recall rate and an average precision mean value; 4, training iteration is carried out until the maximum number of times is reached; and step 5, according to the recorded recall rate and average precision mean value of the classifier test, obtaining a classifier with the optimal performance index in all iterations for sea radar target classification. The weak target classification capability of the classification method under the complex sea surface background is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of radar target classification technology, and in particular relates to a radar sea surface target classification method based on channel attention and dynamic rectification function. Background Technology

[0002] With the increasing demands for marine resource development and maritime security, the detection and classification of sea surface targets play a crucial role in marine monitoring, security protection, and maritime rescue. Radar systems, by transmitting and receiving electromagnetic echo signals, can acquire information on the electromagnetic scattering characteristics of sea surface targets around the clock and in all weather conditions. However, the complex and ever-changing marine environment, coupled with issues such as the coupling effect between targets and the dynamically changing sea surface in the electromagnetic domain, the imbalance between target and background categories, and the difficulty in modeling maneuvering targets on the sea surface, pose challenges to traditional radar signal processing methods in the field of sea surface target detection and classification.

[0003] However, traditional statistical radar target classification methods rely on manually designed features, and in complex scenarios, their performance in extracting electromagnetic features and classifying probabilistic values ​​for weak targets on the sea surface is often unsatisfactory. Summary of the Invention

[0004] The purpose of this invention is to address the problems of weak electromagnetic scattering intensity of radar targets on the sea surface, which are easily submerged by background clutter / noise and have insufficient features. The invention proposes a radar target classification method based on channel attention and dynamic rectification function.

[0005] To achieve the objectives of this invention, the present invention provides a radar sea surface target classification method based on channel attention and dynamic rectification function, comprising the following steps:

[0006] Step 1: Based on the radar sea surface target classification scenario and classification requirements, and referring to the completeness and balance of the dataset construction, samples are extracted from the self-constructed or existing radar sea detection target classification dataset to construct training and test datasets; the sample form in the dataset is a local radar target slice of size Q1×Q2 extracted from two-dimensional radar video data.

[0007] Step 2: Train the radar sea surface target classification network based on channel attention and dynamic rectification function on the training dataset;

[0008] A radar sea surface target classification network based on channel attention and dynamic rectification function is determined based on deep convolutional neural networks, channel attention, and dynamic rectification function. The network structure consists of 10 layers. The first layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the second layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the third layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the fourth layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the fifth layer contains convolutional layers and dynamic rectification function in sequence; the sixth layer contains a channel attention mechanism module; the seventh layer contains convolutional layers; the eighth layer contains a random deactivation layer; the ninth layer contains convolutional layers; and the tenth layer contains a flattening layer.

[0009] A network classifier is obtained by training a radar sea surface target classification network based on channel attention and dynamic rectification function; the classifier includes the radar sea surface target classification network based on channel attention and dynamic rectification function, and the corresponding network weights;

[0010] Step 3: Use the radar sea detection target classification dataset test set samples to test the training obtained classification network and network weight coefficients, and record the two test metrics: recall and mean average precision (mAP).

[0011] Step 4: Determine whether the number of training iterations for the radar sea surface target classification network based on channel attention and dynamic rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weight coefficients and jump to step 2; if the number of iterations has reached the set maximum value, proceed to step 5.

[0012] Steps: Based on the recorded recall and mean precision of the classifier tests, determine the classifier with the best performance in all iterations for classifying marine radar targets.

[0013] The ratio of training dataset to test dataset size is between 6:4 and 8:2; the completeness of the radar sea surface target classification dataset means that, for application scenarios, the dataset should cover various influencing factors of the object to be processed; the balance of the radar sea surface target classification dataset means that, for application scenarios, the distribution of the number of samples in the dataset should follow a reasonable proportion; the sample form in the dataset is a local slice of radar target with size Q1×Q2 extracted from two-dimensional radar video data, where Q1×Q2 has a value of Q1≥32 and Q2≥32; the radar target is located in the middle of the slice image.

[0014] Furthermore, in step 2, the radar sea surface target classification network based on channel attention and dynamic rectification function uses an input sample image of size Q1×Q2, with the specific size determined according to the dataset and actual conditions. The network consists of 10 layers, and its structure is as follows:

[0015] The first layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N1×N1, 8 channels, a stride of S1, and zero-padding size of P1. The parameters of the dynamic rectified functions include 8 channels and a two-dimensional structure. The parameters of the max pooling layers include a kernel size of M1×M1 and a stride of S2. The kernel size, stride, and zero-padding size of the first layer of the network are determined based on the specific application.

[0016] The second layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N²×N², 16 channels, a stride of S³, and zero-padding size of P². The parameters of the dynamic rectified functions include 16 channels and a two-dimensional structure. The parameters of the max pooling layers include a kernel size of M¹×M¹ and a stride of S⁴. The kernel size, stride, and zero-padding size of the second layer are determined based on the specific application.

[0017] The third layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N³×N³, 32 channels, a stride of S⁵, and zero-padding of P³. The dynamic rectified functions have 32 channels and are two-dimensional. The max pooling layers have a kernel size of M³×M³ and a stride of S⁶. The kernel size, stride, and zero-padding of the third layer are determined based on the specific application.

[0018] The fourth layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N4×N4, 64 channels, a stride of S7, and zero-padding size of P4. The parameters of the dynamic rectified functions include 64 channels and a two-dimensional structure. The parameters of the max pooling layers include a kernel size of M4×M4 and a stride of S8. The kernel size, stride, and zero-padding size of the fourth layer are determined based on the specific application.

[0019] The fifth layer of the network consists of convolutional layers and a dynamic rectifier function. The parameters of the convolutional layers are: kernel size N5×N5, number of channels 128, stride S9, and zero-padding size P5. The parameters of the dynamic rectifier function are: number of channels 128 and two-dimensional. The kernel size, stride, and zero-padding size of the fifth layer of the network are determined according to the actual application.

[0020] The 6th layer of the network contains a channel attention mechanism module, and the dimensionality reduction coefficient of the channel attention mechanism module is D. E Determine the parameter values ​​for the channel attention mechanism module in the 6th layer of the network based on the actual application.

[0021] The 7th layer of the network contains convolutional layers. The parameters of the convolutional layers are: kernel size N6×N6, number of channels 256, and stride S. 10 The zero-padding size is P6; the kernel size, stride, and zero-padding size of the 7th layer group of the network are determined according to the actual application.

[0022] The network's 8th layer group contains a random deactivation layer, where the deactivation rate in the parameters of the random deactivation layer is D. r D r The value range is 0 <D r <1; Random deactivation layer with D r The neurons are randomly inactivated with a probability of ×100%.

[0023] The 9th layer of the network contains a convolutional layer with a kernel size of N7×N7, 256 channels, and a stride of S. 11 The zero-padding size is P7; the kernel size, stride, and zero-padding size of the 9th layer group of the network are determined according to the actual application.

[0024] The 10th layer of the network contains a flattening layer; the output of the 9th layer is flattened into one-dimensional vector data after being processed by the flattening layer group, which is used to determine the target category.

[0025] Furthermore, the radar sea surface target classification network based on channel attention and dynamic rectification function in step 2 includes a dynamic rectification function, which comprises a hyperfunction and a dynamic piecewise linear function:

[0026] The hyperfunction η(x) is implemented through a network module. It performs global context encoding on the input feature map data to generate the parameters required for the activation function. For η(x), the dimension of the input feature map data x is C∩H∩W, where C represents that the feature map data has C channels, H represents that the feature map data has H rows, and W represents that the feature map data has W columns. First, the input feature map data undergoes average pooling (AP). Then, the output result is sequentially input into a fully connected (FC) layer W1, a ReLU function, and a fully connected layer W2. The number of output parameters is 2K, where K is the number of functions. The output result is normalized using the Sigmoid function, so that the output parameters are constrained within the closed interval [-1, 1]. The output parameter is Δp. k (x) and Δq k (x); u 1:KLet v represent a vector of dimension K1 consisting of 1s. 1:K This represents a vector of dimension K1 consisting of 0s; parameter p k and q k We obtain this from the following two formulas:

[0027] (1);

[0028] (2);

[0029] Dynamic piecewise linear function y=f η(x) (x): The static form of the traditional rectifier function is y=max(x,0), where max represents taking the maximum value among all variables; while the dynamic rectifier function is extended to a two-piece linear function, and all spaces and channels share the same set of parameters. The expression of the dynamic piecewise linear function is as follows:

[0030] (3).

[0031] Furthermore, the radar sea surface target classification network based on channel attention and dynamic rectification function in step 2 includes a channel attention mechanism module. This module first performs average pooling on the C×H×W feature map data as input, compressing the H×W matrix of each channel into a single-valued statistic. After average pooling, a C×1×1 vector is obtained, representing the global response intensity of each channel. Subsequently, the average pooling result is sequentially input into a fully connected layer W3, a ReLU function, a fully connected layer W4, and a Sigmoid function for processing. After processing by the fully connected layer W3, the vector dimension is reduced to (C / D). E The vector dimension is C×1×1, where C is the dimensionality reduction coefficient. After processing by the fully connected layer W3, the vector dimension returns to C×1×1. The channel weights generated by the channel attention mechanism module are labeled as C. AM (x); The output of the average pooling channel attention mechanism module is ,in This represents the element-wise dot product.

[0032] The significant advancement of this invention compared to existing technologies lies in:

[0033] This invention addresses the problem of insufficient features of weak targets on the sea surface and the difficulty in classifying weak targets due to weak electromagnetic scattering intensity. It employs a channel attention mechanism module to weaken noisy channels and focus more attention on target channels. It also uses a dynamic rectifier function to optimize the static activation function in traditional convolutional neural networks, dynamically adjusting parameters based on input to improve the nonlinear expressive power of the model, thereby enhancing the classification method's ability to classify weak targets in complex sea surface backgrounds.

[0034] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the network structure of the method provided by the present invention.

[0036] Figure 2 This is a schematic diagram of the dynamic rectification function principle in the method provided by the present invention.

[0037] Figure 3 This is a schematic diagram of the channel attention mechanism module in the method provided by the present invention. Detailed Implementation

[0038] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Deep learning technology based on convolutional neural networks has provided a new technological paradigm for radar sea surface target classification due to its powerful multi-dimensional feature extraction capabilities.

[0040] This invention proposes a radar sea surface target classification method based on channel attention and a dynamic rectified function, using a deep convolutional neural network and incorporating a dynamic rectified function and a channel attention mechanism module. This method replaces the static rectified activation function used in conventional deep convolutional neural networks with a dynamic rectified function, thereby adaptively adjusting the parameters of the piecewise linear function using global context information, thus enhancing the nonlinear expressive power of the neural network classification model. Simultaneously, the method employs a channel attention mechanism module to weaken noisy channels and focus more attention on target channels, thereby enhancing the classification method's ability to classify weak targets against complex sea surface backgrounds.

[0041] Step 1: Based on the radar sea surface target classification scenario and classification requirements, and referring to the completeness and balance of the dataset construction, samples are extracted from self-constructed or existing radar sea surface detection target classification datasets to construct training and test datasets. The ratio of the number of training and test datasets is typically between 6:4 and 8:2. The completeness of the radar sea surface target classification dataset construction means that, for the application scenario, the dataset should cover various influencing factors of the object to be processed. The balance of the radar sea surface target classification dataset construction means that, for the application scenario, the distribution of the number of samples in the dataset should follow a reasonable proportion. The sample form in the dataset is a local slice of radar target with a size of Q1×Q2 extracted from two-dimensional radar video data. Typically, Q1×Q2 is Q1≥32 and Q2≥32. The radar target should usually be located in the center of the slice image.

[0042] Step 2: Train the radar sea surface target classification network based on channel attention and dynamic rectification function on the training dataset; this invention determines the radar sea surface target classification network based on channel attention and dynamic rectification function based on deep convolutional neural network, channel attention, and dynamic rectification function; the main structure and parameters of the radar sea surface target classification network based on channel attention and dynamic rectification function are as follows: Figure 1 As shown. The learning rate is 0.001. This network consists of 10 layers, and the network structure is as follows:

[0043] (1) The first layer of the network consists of a convolutional layer, a dynamic rectified function, and a max pooling layer. The parameters of the convolutional layer are: kernel size N1×N1, number of channels 8, stride S1, and zero padding size P1; the parameters of the dynamic rectified function are: number of channels 8 and two-dimensional; the parameters of the max pooling layer are: kernel size M1×M1 and stride S2; the main parameters of the first layer of the network are N1=3, S1=1, P1=0, M1=2, and S2=2.

[0044] The schematic diagram of the dynamic rectification function is as follows: Figure 2 As shown. The dynamic rectifier function includes a hyperfunction and a dynamic piecewise linear function:

[0045] The superfunction η(x) is implemented using a simple network module. It performs global context encoding on the input feature map data to generate the parameters required for the activation function. For η(x), the dimension of the input feature map data x is C∩H∩W, where C represents the number of channels, H represents the number of rows, and W represents the number of columns. First, the input feature map data is processed by Average Pooling (AP). Then, the output is sequentially fed into a fully connected (FC) layer W1, a ReLU function, and a fully connected layer W2. The number of output parameters is 2K, where K is the number of functions. The output is normalized using the Sigmoid function, constraining the output parameters to the closed interval [-1, 1]. The output parameter is Δp. k (x) and Δq k (x). u 1:K Let v represent a vector of dimension K1 consisting of 1s. 1:K This represents a vector of dimension K1 consisting of zeros. The parameter p k and q k We obtain this from the following two formulas:

[0046] (1);

[0047] (2);

[0048] Wherein, parameter λ p and λ q The value of is λ p =1 and λ q =0.5.

[0049] Dynamic piecewise linear function y=f η(x) (x): The static form of the traditional rectifier function is y=max(x,0), where max represents taking the maximum value among all variables; while the dynamic rectifier function extends it to a two-piece linear function, and all spaces and channels share the same set of parameters. The expression of the dynamic piecewise linear function is as follows:

[0050] (3);

[0051] (2) The second layer of the network consists of a convolutional layer, a dynamic rectified function, and a max pooling layer. The parameters of the convolutional layer are: kernel size N2×N2, number of channels 16, stride S3, and zero-padding size P2; the parameters of the dynamic rectified function are: number of channels 16 and two-dimensional; the parameters of the max pooling layer are: kernel size M1×M1 and stride S4; the main parameters of the second layer of the network are N2=4, S3=1, P2=0, M2=2, and S4=2.

[0052] (3) The third layer of the network consists of a convolutional layer, a dynamic rectified function, and a max pooling layer. The parameters of the convolutional layer are: kernel size N3×N3, number of channels 32, stride S5, and zero-padding size P3; the parameters of the dynamic rectified function are: number of channels 32 and two-dimensional; the parameters of the max pooling layer are: kernel size M3×M3 and stride S6; the main parameters of the third layer of the network are N3=3, S5=1, P3=0, M3=2, and S6=2.

[0053] (4) The fourth layer of the network consists of a convolutional layer, a dynamic rectified function, and a max pooling layer. The parameters of the convolutional layer are: kernel size N4×N4, number of channels 64, stride S7, and zero-padding size P4; the parameters of the dynamic rectified function are: number of channels 64 and two-dimensional; the parameters of the max pooling layer are: kernel size M4×M4 and stride S8; the main parameters of the fourth layer of the network are N4=2, S7=1, P4=0, M4=2, and S8=2.

[0054] (5) The fifth layer of the network consists of convolutional layers and dynamic rectifiers. The parameters of the convolutional layers are: kernel size N5×N5, number of channels 128, stride S9, and zero-padding size P5; the parameters of the dynamic rectifiers are: number of channels 128 and two-dimensional; the main parameters of the fifth layer of the network are N5=3, S9=1, and P5=0.

[0055] (6) The 6th layer of the network contains a channel attention mechanism module, and the principle diagram is shown below. Figure 3 As shown; the dimensionality reduction coefficient of this module is D. E The parameters of the channel attention mechanism module in layer 6 of the network were determined based on the actual application. The channel attention mechanism module first performs average pooling on the C×H×W feature map data as input, compressing the H×W matrix of each channel into a single-valued statistic. After average pooling, a C×1×1 vector is obtained, representing the global response intensity of each channel. The average pooling result is then sequentially input into fully connected layer W3, the ReLU function, fully connected layer W4, and the Sigmoid function for processing. After processing by fully connected layer W3, the dimension of the vector is reduced to (C / D). E The vector dimension is C×1×1, where C is the dimensionality reduction coefficient. After processing by the fully connected layer W3, the vector dimension returns to C×1×1. The channel weights generated by the channel attention mechanism module are labeled C. AM (x). The output of the average pooling channel attention mechanism module is ,in This represents the element-wise dot product. The main parameter value for the 6th layer of the network is D. E =16, the number of channels is 128.

[0056] (7) The 7th layer of the network contains a convolutional layer. The parameters of the convolutional layer are: kernel size N6×N6, number of channels 256, and stride S. 10 The zero-padding size is P6; the main parameters of the 7th layer group of the network are N6=3 and S. 10 =1、P6=0.

[0057] (8) The 8th layer of the network contains a random deactivation layer, and the deactivation rate in the parameters of the random deactivation layer is D. r D r The value range is 0 <D r <1; Random deactivation layer with D r Neurons are randomly inactivated with a probability of ×100%, and the main parameter value of the 8th layer of the network is D. r =0.5.

[0058] (9) The 9th layer of the network contains a convolutional layer. The parameters of the convolutional layer are: kernel size N7×N7, number of channels 256, and stride S. 11 The zero-padding size is P7; the main parameters of the 9th layer group of the network are N7=3 and S. 11 =1、P7=0.

[0059] (10) The 10th layer of the network contains a flattening layer. The output of the 9th layer is flattened into one-dimensional vector data after being processed by this layer, which is used to determine the target category.

[0060] A network classifier is obtained by training a radar sea surface target classification network based on channel attention and dynamic rectification function; the classifier includes the radar sea surface target classification network based on channel attention and dynamic rectification function, and the corresponding network weights.

[0061] Step 3: Use the radar sea detection target classification dataset test set samples to test the training obtained classification network and network weight coefficients, and record the two test metrics: recall and mean average precision (mAP).

[0062] Step 4: Check if the number of training iterations for the radar sea surface target classification method based on channel attention and dynamic rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weight coefficients and jump to step 2; if the number of iterations has reached the set maximum value, proceed to step 5.

[0063] Step 5: Based on the recorded recall and mean precision of the classifier tests, determine the classifier with the best performance across all iterations for classifying marine radar targets.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A radar-based sea surface target classification method based on channel attention and dynamic rectification function, characterized in that, The method includes: Step 1: Based on the radar sea surface target classification scenario and classification requirements, and referring to the completeness and balance of the dataset construction, samples are extracted from the self-constructed or existing radar sea detection target classification dataset to construct training and test datasets; the sample form in the dataset is a local radar target slice of size Q1×Q2 extracted from two-dimensional radar video data. Step 2: Train the radar sea surface target classification network based on channel attention and dynamic rectification function on the training dataset; Step 3: Use the radar sea detection target classification dataset test set samples to test the training obtained classification network and network weight coefficients, and record the two test metrics: recall and mean average precision (mAP). Step 4: Determine whether the number of training iterations for the radar sea surface target classification network based on channel attention and dynamic rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weight coefficients and jump to step 2; if the number of iterations has reached the set maximum value, proceed to step 5. Step 5: Based on the recorded recall and mean precision of the classifier tests, determine the classifier with the best performance in all iterations for classifying marine radar targets.

2. The radar sea surface target classification method based on channel attention and dynamic rectification function according to claim 1, characterized in that, Step 2: Train the radar sea surface target classification network based on channel attention and dynamic rectification function on the training dataset; including: A radar sea surface target classification network based on channel attention and dynamic rectification function is determined based on deep convolutional neural networks, channel attention, and dynamic rectification function. The network structure consists of 10 layers. The first layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the second layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the third layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the fourth layer contains convolutional layers, dynamic rectification function, and max pooling layers in sequence; the fifth layer contains convolutional layers and dynamic rectification function in sequence; the sixth layer contains a channel attention mechanism module; the seventh layer contains convolutional layers; the eighth layer contains a random deactivation layer; the ninth layer contains convolutional layers; and the tenth layer contains a flattening layer. A network classifier is obtained by training a radar sea surface target classification network based on channel attention and dynamic rectification function; the classifier includes the radar sea surface target classification network based on channel attention and dynamic rectification function, and the corresponding network weights.

3. The radar sea surface target classification method based on channel attention and dynamic rectification function according to claim 1, characterized in that, In step 1, the ratio of the number of training datasets to test datasets is between 6:4 and 8:2; the classification scenarios include target classification scenarios under typical marine exploration backgrounds such as noisy background, sea clutter background, and cloud and rain clutter background; the classification requirements include various target classification requirements such as low signal-to-noise ratio target classification requirements, low signal-to-noise ratio target classification requirements, low time overhead target classification requirements, and low memory overhead target classification requirements. The completeness of the radar sea surface target classification dataset construction refers to the fact that, oriented towards the application scenario, the dataset should cover all influencing factors of the object to be processed; the balance of the radar sea surface target classification dataset construction refers to the fact that, oriented towards the application scenario, the number distribution of samples in the dataset should follow a reasonable proportion; the sample form in the dataset is a local slice of radar target with size Q1×Q2 extracted from two-dimensional radar video data, where Q1×Q2 has values ​​of Q1≥32 and Q2≥32; The radar target is located in the center of the slice image.

4. The radar sea surface target classification method based on channel attention and dynamic rectification function according to claim 1, characterized in that: In step 2, the radar sea surface target classification network based on channel attention and dynamic rectification function uses an input sample image of size Q1×Q2, with the specific size determined according to the dataset and actual conditions. The network consists of 10 layers, and its structure is as follows: The first layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N1×N1, 8 channels, a stride of S1, and zero-padding size of P1. The parameters of the dynamic rectified functions include 8 channels and a two-dimensional structure. The parameters of the max pooling layers include a kernel size of M1×M1 and a stride of S2. The kernel size, stride, and zero-padding size of the first layer of the network are determined based on the specific application. The second layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N²×N², 16 channels, a stride of S³, and zero-padding size of P². The parameters of the dynamic rectified functions include 16 channels and a two-dimensional structure. The parameters of the max pooling layers include a kernel size of M¹×M¹ and a stride of S⁴. The kernel size, stride, and zero-padding size of the second layer are determined based on the specific application. The third layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N³×N³, 32 channels, a stride of S⁵, and zero-padding of P³. The dynamic rectified functions have 32 channels and are two-dimensional. The max pooling layers have a kernel size of M³×M³ and a stride of S⁶. The kernel size, stride, and zero-padding of the third layer are determined based on the specific application. The fourth layer of the network consists of convolutional layers, dynamic rectified functions, and max pooling layers. The parameters of the convolutional layers include a kernel size of N4×N4, 64 channels, a stride of S7, and zero-padding size of P4. The parameters of the dynamic rectified functions include 64 channels and a two-dimensional structure. The parameters of the max pooling layers include a kernel size of M4×M4 and a stride of S8. The kernel size, stride, and zero-padding size of the fourth layer are determined based on the specific application. The fifth layer of the network consists of convolutional layers and a dynamic rectifier function. The parameters of the convolutional layers are: kernel size N5×N5, number of channels 128, stride S9, and zero-padding size P5. The parameters of the dynamic rectifier function are: number of channels 128 and two-dimensional. The kernel size, stride, and zero-padding size of the fifth layer of the network are determined according to the actual application. The 6th layer of the network contains a channel attention mechanism module, and the dimensionality reduction coefficient of the channel attention mechanism module is D. E Determine the parameter values ​​for the channel attention mechanism module in the 6th layer of the network based on the actual application. The 7th layer of the network contains convolutional layers. The parameters of the convolutional layers are: kernel size N6×N6, number of channels 256, and stride S. 10 The zero-padding size is P6; the kernel size, stride, and zero-padding size of the 7th layer group of the network are determined according to the actual application. The network's 8th layer group contains a random deactivation layer, where the deactivation rate in the parameters of the random deactivation layer is D. r D r The value range is 0 <D r <1; Random deactivation layer with D r The neurons are randomly inactivated with a probability of ×100%. The 9th layer of the network contains a convolutional layer with a kernel size of N7×N7, 256 channels, and a stride of S. 11 The zero-padding size is P7; the kernel size, stride, and zero-padding size of the 9th layer group of the network are determined according to the actual application. The 10th layer of the network contains a flattening layer; the output of the 9th layer is flattened into one-dimensional vector data after being processed by the flattening layer group, which is used to determine the target category.

5. The radar sea surface target classification method based on channel attention and dynamic rectification function according to claim 1, characterized in that: The radar sea surface target classification network based on channel attention and dynamic rectification function in step 2 includes a dynamic rectification function, which comprises a hyperfunction and a dynamic piecewise linear function. The hyperfunction η(x) is implemented through a network module. It performs global context encoding on the input feature map data to generate the parameters required for the activation function. For η(x), the dimension of the input feature map data x is C∩H∩W, where C represents that the feature map data has C channels, H represents that the feature map data has H rows, and W represents that the feature map data has W columns. First, the input feature map data undergoes average pooling (AP). Then, the output result is sequentially input into a fully connected (FC) layer W1, a ReLU function, and a fully connected layer W2. The number of output parameters is 2K, where K is the number of functions. The output result is normalized using the Sigmoid function, so that the output parameters are constrained within the closed interval [-1, 1]. The output parameter is Δp. k (x) and Δq k (x); u 1:K Let v represent a vector of dimension K1 consisting of 1s. 1:K This represents a vector of dimension K1 consisting of 0s; parameter p k and q k We obtain this from the following two formulas: (1)； (2)； Dynamic piecewise linear function y=f η(x) (x): The dynamic rectification function is extended into a two-piece linear function, and all spaces and channels share the same set of parameters. The expression of the dynamic piecewise linear function is as follows: (3)。 6. The radar sea surface target classification method based on channel attention and dynamic rectification function according to claim 1, characterized in that: In step 2, the radar sea surface target classification network based on channel attention and dynamic rectification function includes a channel attention mechanism module. This module first performs average pooling on the C×H×W feature map data as input, compressing the H×W matrix of each channel into a single-valued statistic. After average pooling, a C×1×1 vector is obtained, representing the global response intensity of each channel. Subsequently, the average pooling result is sequentially input into a fully connected layer W3, a ReLU function, a fully connected layer W4, and a Sigmoid function for processing. After processing by the fully connected layer W3, the vector dimension is reduced to (C / D). E The vector dimension is C×1×1, where C is the dimensionality reduction coefficient. After processing by the fully connected layer W3, the vector dimension returns to C×1×1. The channel weights generated by the channel attention mechanism module are labeled as C. AM (x); The output of the average pooling channel attention mechanism module is ,in This represents the element-wise dot product.

Images