Sea surface target detection method based on cost-sensitive focus loss convolutional neural network

Abstract

The invention discloses a sea surface target detection method based on a cost-sensitive focus loss convolutional neural network, and belongs to the field of radar target classification. The method comprises the following steps: acquiring data of sea clutters and target echoes by adopting a radar, and preprocessing the data of the sea clutters and the target echoes; extracting high-dimensional features of the target time sequence and the sea clutter time sequence according to the preprocessed data, and constructing a feature data set for training a convolutional neural network model; setting a loss function of the convolutional neural network model in combination with a focus loss function and a cost-sensitive learning mechanism; dynamically adjusting misclassification cost values in the cost matrix to optimize a training process; and outputting a detection result by using the convolutional neural network model after optimization training. According to the method, focus loss and cost sensitive learning are ingeniously combined, the attention degree on difficult and easy samples is balanced in the training process, different costs are applied to different types of misclassification, and the strategy that the model processes various types of misclassification is optimized.

Description

Technical Field

[0001] The invention belongs to the field of radar target classification, and in particular relates to a sea surface target detection method based on a cost-sensitive focal loss convolutional neural network. Background Art

[0002] Since sea surface targets are usually non-cooperative targets and are relatively small in number, the sample size of small sea surface targets is relatively small in radar target detection, and sea clutter becomes a large sample, which inevitably leads to an imbalance in the data set samples. For data-driven algorithms such as deep learning, this may affect the training effect of the model and the final detection performance, because the model tends to better identify categories with larger sample sizes and ignore important categories with smaller sample sizes. This bias may lead to deviations in classification results, resulting in incorrect decisions.

[0003] In the field of deep learning, especially in the application of CNN, the loss function plays a vital role. It measures the difference between the model's prediction results and the actual labels, and guides the model to continuously optimize parameters through the learning process to improve the accuracy of predictions. Several common traditional loss functions include Mean Squared Error Loss (MSELoss), Cross-Entropy Loss (CE Loss) and Hinge loss.

[0004] The Mean Squared Error (MSE) loss function measures the performance of a model by calculating the average of the squares of the differences between the model's predictions and the actual values. MSE is continuously differentiable over the entire domain, which makes it well suited for use with optimization algorithms such as gradient descent. By calculating the gradient, the model can gradually adjust parameters to minimize the loss. Since MSE squares the errors, larger errors are amplified, which has a greater impact on the total loss. This means that when there are outliers in the dataset, MSE may cause the model to favor these outliers, affecting the overall performance.

[0005] The core purpose of Cross Entropy Loss (CE Loss) is to quantify the difference between the model's predicted probability distribution and the true label distribution, thereby guiding the model to learn a more accurate classification strategy. Cross entropy loss is highly sensitive to misclassification, especially when the model predicts a low probability for certain categories, which helps the model learn the difference between different categories faster. However, CE Loss also has its limitations. When dealing with highly unbalanced data sets, CELoss may cause the model to tend to optimize the prediction accuracy of the majority class while ignoring the minority class, which may lead to poor classification of the minority class.

[0006] Hinge loss is a commonly used loss function in support vector machines, especially for binary classification problems. Its design goal is to find a decision boundary so that positive and negative samples are not only correctly separated, but also the interval between them is as large as possible. This strategy of maximizing the interval helps to improve the generalization ability of the model. However, Hinge loss also has some limitations. First, Hinge loss is mainly used for binary classification tasks. In multi-classification problems, strategies such as "one-to-one" or "one-to-all" need to be adopted, which will increase the complexity of the model and training time. Secondly, Hinge loss is insensitive to samples that have been correctly classified and are far away from the decision boundary, which may cause the model to ignore the potential information of this part of the data.

[0007] With the increasing attention paid to the problem of sample imbalance, Lin et al. proposed Focal Loss (FL) in 2017 to improve the performance of classification models. Focal Loss is a variant of cross entropy loss. By adjusting the proportion of loss contribution and reducing the weight of easy-to-classify samples, the model pays more attention to samples that are difficult to classify or misclassified. Although Focal Loss performs well in dealing with imbalanced problems, it still fails to consider the actual cost of different types of errors. Therefore, the present invention proposes a cost-sensitive focal loss function, which aims to assign different costs to different classification errors to optimize the performance of the overall model. Summary of the invention

[0008] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a sea surface target detection method based on a cost-sensitive focal loss convolutional neural network, which solves the problems in the prior art.

[0009] The purpose of the present invention can be achieved through the following technical solutions:

[0010] The sea surface target detection method based on cost-sensitive focal loss convolutional neural network includes the following steps:

[0011] The radar is used to obtain the data of sea clutter and target echo, and the data of sea clutter and target echo are preprocessed;

[0012] Extract high-dimensional features of target time series and sea clutter time series based on preprocessed data, and construct feature datasets for training convolutional neural network models;

[0013] Combine the focal loss function and the cost-sensitive learning mechanism to set the loss function of the convolutional neural network model;

[0014] Dynamically adjust the misclassification cost value in the cost matrix to optimize the training process;

[0015] Use the optimized trained convolutional neural network model to output the detection results.

[0016] Furthermore, the steps of constructing the feature data set include:

[0017] S21, extracts two amplitude features, two Doppler features and two time-frequency features from the sea clutter time series and the target time series, a total of six significant features;

[0018] Among them, the two amplitude features are: time domain information entropy and time Hurst exponent; the two Doppler features are: relative Doppler peak height and relative vector entropy; the two time-frequency features are: ridge accumulation and the number of connected regions.

[0019] S22, respectively construct the six significant features of the sea clutter time series and the target time series into six-dimensional feature vectors, and respectively construct the six-dimensional feature vectors into a sea clutter feature matrix G 0 And the target feature matrix G 1 , and then construct the feature dataset G fea for:

[0020] G fea =G 0 ∪G 1

[0021] Where: G fea The dimension of is 6×m, where 6 is the number of features and m is the sum of the number of feature samples of the sea clutter time series and the target time series.

[0022] Furthermore, when calculating the decision threshold Threshold corresponding to the Softmax function of the output layer of the convolutional neural network model, the false alarm probability P to be controlled is fa Introduced here, the calculation formula of the decision threshold Threshold is:

[0023] p a =Softmax(clu(a);H 0 |H 0 ),a=1…A

[0024] output = {p 1 ,p 2 ,…p a ,…p β ,…p A}

[0025] Threshold = p β ,β=P fa *A

[0026] Where: H 0indicates that the sample to be detected does not contain the target, clu(a) represents the feature sequence obtained by convolution and pooling the feature vector of the a-th sea clutter time series, and p a is the output probability of the feature sequence being judged as sea clutter after the Softmax function, output is the set of output probabilities, the samples of this set are sorted from small to large according to the probability value, A is the number of sea clutter sample feature sequences; using the Monte Carlo method, take the output probability p of the βth position β as the detection threshold.

[0027] Furthermore, the cost matrix in the cost-sensitive learning mechanism is:

[0028]

[0029] Among them, ω TP ,ω FN ,ω FP ,ω TN denote the costs of true positive, false negative, false positive, and true negative, respectively;

[0030] The loss function expression of the convolutional neural network model is:

[0031]

[0032] Among them, d (i) represents the actual label of the i-th sample, p (i) Represents the probability that the model predicts that the i-th sample is a positive class, μ (i) Used to adjust the weight of positive samples in loss calculation.

[0033] Furthermore, the process of dynamically adjusting the misclassification cost in the cost matrix is:

[0034] The misclassification cost ω FN ,ω FP It is expressed as:

[0035]

[0036] Among them, ω (i) Represents the misclassification cost corresponding to the i-th sample, that is, if the sample is positive, it represents the false negative cost ω FN , otherwise if it is a negative class, it means the false positive cost ω FP θ (i) and R (i) Represent the fixed factor and the adjustment factor after each training iteration, respectively. (i) To control the calculation of the loss function; N total Represents the total number of samples, N irepresents the number of samples in the category corresponding to the i-th sample, σ is a hyperparameter used to adjust the scale of the category cost weight; N 10 Represents the number of positive samples that are incorrectly predicted as negative samples, N 11 N 11 Represents the number of correctly predicted positive samples, and when the i-th sample is a negative class, its R (i) The value of is always 0; at the same time, when d (i) =1, R (i) The initial value of is set to 1;

[0037] Dynamically adjusted cost matrix ω * The expression is:

[0038]

[0039] Furthermore, the step of outputting the detection result includes:

[0040] S51, extracting features of the echo to be detected to construct a feature data set, and sending it to the optimized and trained convolutional neural network model to obtain the classification probability after the Softmax function;

[0041] S52, for each echo to be detected, if it is judged to be a target, the probability p c If the value is greater than the threshold, the echo to be detected is judged as a target, otherwise it is judged as sea clutter.

[0042] The sea surface target detection system based on cost-sensitive focal loss convolutional neural network includes:

[0043] Data receiving module: uses radar to obtain data of sea clutter and target echo, and pre-processes the data of sea clutter and target echo;

[0044] Feature extraction module: extracts high-dimensional features of target time series and sea clutter time series based on preprocessed data, and constructs feature data sets for training convolutional neural network models;

[0045] Loss function setting module: Combines the focal loss function and the cost-sensitive learning mechanism to set the loss function of the convolutional neural network model;

[0046] Dynamic adjustment module: dynamically adjusts the misclassification cost value in the cost matrix to optimize the training process;

[0047] Detection module: Use the optimized and trained convolutional neural network model to output the detection results.

[0048] A computer storage medium stores a readable program, which can execute the above-mentioned sea surface target detection method based on cost-sensitive focal loss convolutional neural network when the program is running.

[0049] An electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus;

[0050] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the above-mentioned sea surface target detection method based on cost-sensitive focal loss convolutional neural network.

[0051] A computer program product includes computer instructions, wherein the computer instructions instruct a computing device to perform operations corresponding to the above-mentioned sea surface target detection method based on a cost-sensitive focal loss convolutional neural network.

[0052] Beneficial effects of the present invention:

[0053] 1. The target detection method of the present invention cleverly combines focal loss with cost-sensitive learning, which not only balances the attention to difficult and easy samples during the training process, but also optimizes the model's strategy for handling various types of misclassification by imposing different costs on different types of misclassification.

[0054] 2. The dynamic misclassification cost adjustment mechanism within the target detection of the present invention helps to achieve a balance between the attention paid to the minority class and the majority class at various stages of training, thereby enhancing the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 is a flow chart of the sea surface target detection method of the present invention;

[0057] Figure 2 This is an operational flow chart of the present invention for implementing sea surface target classification;

[0058] Figure 3 It is a graph of experimental results based on measured data of the present invention;

[0059] Figure 4 It is a block diagram of the sea surface target detection system of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Example 1

[0062] like Figure 1 and Figure 2 As shown, the sea surface target detection method based on the cost-sensitive focal loss convolutional neural network includes the following steps:

[0063] S1, using radar to obtain data of sea clutter and target echo, and preprocessing the data of sea clutter and target echo.

[0064] The specific steps of the pretreatment include:

[0065] S11, data segmentation, divides the measured sea surface data into short time series with a length of 1000.

[0066] The phase space is reconstructed for the time series obtained by segmentation to obtain the embedded vector after phase space reconstruction. The formula is as follows:

[0067] v(r)={x(r),x(r+τ),x(r+2τ),...,x(r+(m-1)τ)}

[0068] Among them, x(r) represents a time series of length N, m represents the embedding dimension, τ represents the embedding delay, r represents the time sample, and v(r) represents the embedding vector;

[0069] Based on the embedded vector, a recursive matrix can be obtained, and the elements in the recursive matrix can be expressed as:

[0070] R r,s =||v(r)-v(s)||

[0071] Among them, 1≤r,s≤n, n=N-(m-1)τ represents the length of the time sample, and v(s) represents the embedding vector of the time sample s. From this, the recursive matrix R can be created as:

[0072]

[0073] S12, grayscale normalization is performed on the recursive matrix, and the grayscale normalization formula is:

[0074]

[0075] Among them, R(i,j) and G(i,j) are the grayscale value of the original image and the grayscale value after transformation, respectively, 1≤i,j≤n, min and max represent the minimum grayscale value and maximum grayscale value of the original image R(i,j), respectively.

[0076] S2, extracts high-dimensional features of the target time series and sea clutter time series based on the preprocessed data to construct a feature dataset for training the convolutional neural network model.

[0077] The specific steps of extracting high-dimensional features of target time series and sea clutter time series include:

[0078] S21, extracts two amplitude features, two Doppler features and two time-frequency features from the sea clutter time series and the target time series, a total of six significant features;

[0079] Among them, the two amplitude features are: time domain information entropy and time Hurst exponent; the two Doppler features are: relative Doppler peak height and relative vector entropy; the two time-frequency features are: ridge accumulation and the number of connected regions.

[0080] S22, respectively construct the six significant features of the sea clutter time series and the target time series into six-dimensional feature vectors, and respectively construct the six-dimensional feature vectors into a sea clutter feature matrix G 0 And the target feature matrix G 1 , and then construct the feature dataset G fea for:

[0081] G fea =G 0 ∪G 1

[0082] Where: G fea The dimension of is 6×m, where 6 is the number of features and m is the sum of the number of feature samples of the sea clutter time series and the target time series.

[0083] The convolutional neural network model includes a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer which are connected in sequence.

[0084] When training the convolutional neural network model, the obtained feature data set is divided into two parts, the training set and the validation set, where the ratio of the training set to the validation set is 8:2. After the feature data set is divided, the feature data sets are labeled separately, and the labels are two categories: target and clutter; the training set and the validation set are input into the convolutional neural network model for training and verification.

[0085] Since the Softmax function of the output layer of the convolutional neural network model can convert the classification results into classification probabilities, the false alarm probability can be accurately controlled by setting different decision thresholds for the Softmax function. fa Introduce it and calculate the decision threshold Threshold corresponding to the Softmax function of the output layer:

[0086] p a =Softmax(clu(a);H 0 |H 0 ),a=1…A

[0087] output = {p 1 ,p 2 ,…p a ,…p β ,…p A}

[0088] Threshold = p β ,β=P fa *A

[0089] Among them, H 0 indicates that the sample to be detected does not contain the target, clu(a) represents the feature sequence obtained by convolution and pooling the feature vector of the a-th sea clutter time series, and p a is the output probability of the feature sequence being judged as sea clutter after the Softmax function, output is the set of output probabilities, the samples of this set are sorted from small to large according to the probability value, A is the number of sea clutter sample feature sequences; using the Monte Carlo method, take the output probability p of the βth position β as the detection threshold.

[0090] S3, combining the focal loss function and the cost-sensitive learning mechanism to set the loss function of the convolutional neural network model;

[0091] The focal loss function is a variant of the cross entropy loss. It reduces the weight of easy-to-classify samples by adjusting the proportion of loss contribution, so that the model pays more attention to samples that are difficult to classify or misclassified.

[0092] The focal loss function formula is expressed as follows:

[0093] FL(p t )=-μ t (1-p t ) γ log(p t )

[0094] Among them, pt It represents the probability that the model correctly classifies the current sample. Its calculation method depends on the actual label d of the sample. When the sample is actually a positive class (d = 1), p t =p, that is, the probability that the model predicts that the sample is a positive class; when the sample is actually a negative class (d = 0), p t =1-p, that is, the probability that the model predicts that the sample is a negative class; μ t is a pre-set weight coefficient. When the sample is actually a positive class, μ t =μ, 0≤μ≤1, when the sample is actually a negative class, μ t =1-μ, γ is the focusing parameter.

[0095] In the cost-sensitive learning mechanism, each classification error is assigned a cost value that reflects the severity or impact of the misclassification.

[0096] In a binary classification problem, the cost matrix is ​​a 2×2 matrix used to represent the cost of various situations in the classification decision results; assuming there are two categories: positive and negative; the cost matrix can be expressed as follows:

[0097]

[0098] where ω TP ,ω FN ,ω FP ,ω TN denote the costs of true positive, false negative, false positive, and true negative, respectively.

[0099] Cost-sensitive learning emphasizes the difference in costs of different types of errors, so a new loss function can be established by combining the cost-sensitive learning mechanism with the focal loss function, which is expressed as follows:

[0100]

[0101] Among them, d (i) represents the actual label of the i-th sample, p (i) Represents the probability that the model predicts that the i-th sample is a positive class, μ (i) It is used to adjust the weight of positive samples in the loss calculation, and γ is the focusing parameter.

[0102] S4, dynamically adjusts the misclassification cost value in the cost matrix to optimize the training process;

[0103] The process of dynamically adjusting the misclassification cost in the cost matrix is:

[0104] For the two-classification problem, the misclassification cost is ω FN ,ω FP It is expressed as:

[0105]

[0106] Among them, ω (i) Represents the misclassification cost corresponding to the i-th sample, that is, if the sample is positive, it represents the false negative cost ω FN , otherwise if it is a negative class, it means the false positive cost ω FP ;

[0107] ω (i) The calculation formula is as follows:

[0108]

[0109] Among them, θ (i) and R (i) Represent the fixed factor and the adjustment factor after each training iteration, respectively. (i) To control the calculation of the loss function;

[0110] θ (i) It represents the reciprocal of the ratio of the number of samples of the category corresponding to the i-th sample to the total number of samples, which is calculated as follows:

[0111]

[0112] Among them, N total Represents the total number of samples, N i represents the number of samples in the category corresponding to the i-th sample, and σ is a hyperparameter used to adjust the scale of the category cost weight;

[0113] R (i) It represents the ratio of positive samples that are incorrectly predicted as negative samples to the total number of positive samples after each iteration. Its expression is as follows:

[0114]

[0115] Among them, N 10 Represents the number of positive samples that are incorrectly predicted as negative samples, N 11 N 11 Represents the number of correctly predicted positive samples, and when the i-th sample is a negative class, its R (i) The value of is always 0; at the same time, when d (i) =1, R (i) The initial value of is set to 1.

[0116] Therefore, a new cost matrix is ​​proposed. Compared with the traditional cost matrix, two aspects are optimized: first, the cost value of correct classification is changed from 0 to 1; second, the fixed cost value of misclassification is changed to dynamic. The new cost matrix expression is as follows:

[0117]

[0118] In the formula, ω * is the new cost matrix.

[0119] Afterwards, the steps of using the trained convolutional neural network model to perform echo detection according to the decision threshold include:

[0120] 1) Extract features from the samples to be tested and use them as test sets, which are then fed into the trained convolutional neural network model;

[0121] 2) Based on the trained convolutional neural network model, the probability p of each sample to be detected being judged as sea clutter is quickly given c .

[0122] p c =Softmax(tar(b);H 0 |H 1 ),b=1,…B

[0123] Where: tar(b) represents the feature sequence obtained after the feature vector of the bth sample to be detected is subjected to the convolution pooling operation, B is the number of feature sequences of the samples to be detected, H 0 Assume that the sample to be detected does not contain the target, H 1 Assumption indicates that the sample to be detected contains the target echo.

[0124] S5, uses the optimized trained convolutional neural network model to output the detection results.

[0125] The steps of outputting the test results specifically include:

[0126] S51, extracting features of the echo to be detected to construct a feature data set, and sending it to the optimized and trained neural network model to obtain the classification probability after the Softmax function;

[0127] S52, for each echo to be detected, if it is judged to be a target, the probability p c If the value is greater than the threshold, the echo to be detected is judged as a target, otherwise it is judged as sea clutter. The judgment formula is:

[0128]

[0129] Example 2

[0130] The effect of the target detection method of the present invention is verified by Example 2;

[0131] In this embodiment, the data measured by Professor S. Haykin of McMaster University in Canada using IPIX radar in a real ocean environment is used as the radar to obtain the sea surface echo amplitude sequence of the present invention. In the experiment, several commonly used loss functions are compared, including MSE, CE, Hinge Loss, Focal Loss and the cost-sensitive focal loss (CoSen-FL) proposed by the present invention. The operation flow of the method of the present invention is as follows: Figure 2 As shown in Figure 2, the detection results of different loss functions are as follows: Figure 3 As shown, from Figure 3 It can be seen that the CoSen-FL proposed in the present invention effectively alleviates the problem of imbalanced dataset categories and improves the probability of target detection to a certain extent. The CoSen-FL-CNN detector effectively avoids the data enhancement and sample generation steps by directly incorporating cost sensitivity and focal loss into the loss function, significantly reducing the training time and required computing resources. The directness of this training strategy not only simplifies the model architecture and improves the stability of the training process, but also provides greater flexibility for adjusting the model according to specific task requirements.

[0132] Therefore, in application scenarios with special requirements for fast response or limited resources, the CoSen-FL-CNN detector demonstrates more outstanding practical application value with its efficient training process and low resource consumption.

[0133] Example 3

[0134] In this embodiment, a sea surface target detection system based on a cost-sensitive focal loss convolutional neural network is proposed. Figure 4 As shown, the detection system includes:

[0135] Data receiving module: uses radar to obtain data of sea clutter and target echo, and pre-processes the data of sea clutter and target echo;

[0136] Feature extraction module: extracts high-dimensional features of target time series and sea clutter time series based on preprocessed data, and constructs feature data sets for training convolutional neural network models;

[0137] Loss function setting module: Combines the focal loss function and the cost-sensitive learning mechanism to set the loss function of the convolutional neural network model;

[0138] Dynamic adjustment module: dynamically adjusts the misclassification cost value in the cost matrix to optimize the training process;

[0139] Detection module: Use the optimized and trained convolutional neural network model to output the detection results.

[0140] The method of the present invention may be implemented in hardware, firmware, or as software or computer code that may be stored in a recording medium (such as a CDROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded over a network and will be stored in a local recording medium, so that the method described herein may be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that a computer, processor, microprocessor controller, or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, processor, or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown herein.

[0141] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A sea surface target detection method based on a cost-sensitive focal loss convolutional neural network, characterized in that: The following steps are involved: The radar is used to obtain the data of sea clutter and target echo, and the data of sea clutter and target echo are preprocessed; Extract high-dimensional features of target time series and sea clutter time series based on preprocessed data, and construct feature datasets for training convolutional neural network models; Combine the focal loss function and the cost-sensitive learning mechanism to set the loss function of the convolutional neural network model; Dynamically adjust the misclassification cost in the cost matrix to optimize the training process; Use the optimized trained convolutional neural network model to output the detection results.

2. The method for sea surface target detection based on cost-sensitive focal loss convolutional neural network according to claim 1 is characterized in that: The steps to construct a feature dataset include: S21, extracts two amplitude features, two Doppler features and two time-frequency features from the sea clutter time series and the target time series, a total of six significant features; Among them, the two amplitude features are: time domain information entropy and time Hurst exponent; the two Doppler features are: relative Doppler peak height and relative vector entropy; the two time-frequency features are: ridge accumulation and the number of connected regions. S22, respectively construct the six significant features of the sea clutter time series and the target time series into six-dimensional feature vectors, and respectively form the six-dimensional feature vectors into the sea clutter feature matrix G0 and the target feature matrix G1, and then construct the feature data set G fea for: G fea =G0∪G1 Where: G fea The dimension of is 6×m, where 6 is the number of features and m is the sum of the number of feature samples of the sea clutter time series and the target time series.

3. The method for sea surface target detection based on cost-sensitive focal loss convolutional neural network according to claim 1 is characterized in that: When calculating the decision threshold Threshold corresponding to the Softmax function of the output layer of the convolutional neural network model, the false alarm probability P to be controlled is fa Introduced here, the calculation formula of the decision threshold Threshold is: p a =Softmax(clu(a);H0|H0),a=1…A output={p1,p2,…p a ,…p β ,…p A } Threshold=p β ,β=P fa *A Where: H0 means that the sample to be detected does not contain the target, clu(a) represents the feature sequence obtained by convolution pooling the feature vector of the a-th sea clutter time series, and p a is the output probability of the feature sequence being judged as sea clutter after the Softmax function, output is the set of output probabilities, the samples of this set are sorted from small to large according to the probability value, A is the number of sea clutter sample feature sequences; using the Monte Carlo method, take the output probability p of the βth position β as the detection threshold.

4. The method for sea surface target detection based on cost-sensitive focal loss convolutional neural network according to claim 1, characterized in that: The cost matrix in the cost-sensitive learning mechanism is: Among them, ω TP ,ω FN ,ω FP ,ω TN denote the costs of true positive, false negative, false positive, and true negative, respectively; The loss function expression of the convolutional neural network model is: Among them, d (i) represents the actual label of the i-th sample, p (i) Represents the probability that the model predicts that the i-th sample is a positive class, μ (i) Used to adjust the weight of positive samples in loss calculation.

5. The method for sea surface target detection based on cost-sensitive focal loss convolutional neural network according to claim 4 is characterized in that: The process of dynamically adjusting the misclassification cost in the cost matrix is: The misclassification cost ω FN ,ω FP It is expressed as: oh (i) =θ (i) ·e R(i) Among them, ω (i) Represents the misclassification cost corresponding to the i-th sample, that is, if the sample is positive, it represents the false negative cost ω FN , otherwise if it is a negative class, it means the false positive cost ω FP θ (i) and R (i) Represent the fixed factor and the adjustment factor after each training iteration, respectively. (i) To control the calculation of the loss function; N total Represents the total number of samples, N i represents the number of samples in the category corresponding to the i-th sample, σ is a hyperparameter used to adjust the scale of the category cost weight; N 10 Represents the number of positive samples that are incorrectly predicted as negative samples, N 11 N 11 Represents the number of correctly predicted positive samples, and when the i-th sample is a negative class, its R (i) The value of is always 0; at the same time, when d (i) =1, R (i) The initial value of is set to 1; Dynamically adjusted cost matrix ω * The expression is:

6. The method for sea surface target detection based on cost-sensitive focal loss convolutional neural network according to claim 1, characterized in that: The steps to output the test results include: S51, extracting features of the echo to be detected to construct a feature data set, and sending it to the optimized and trained convolutional neural network model to obtain the classification probability after the Softmax function; S52, for each echo to be detected, if it is judged to be a target, the probability p c If the value is greater than the threshold, the echo to be detected is judged as a target, otherwise it is judged as sea clutter.

7. A sea surface target detection system based on a cost-sensitive focal loss convolutional neural network, characterized in that: include: Data receiving module: uses radar to obtain data of sea clutter and target echo, and pre-processes the data of sea clutter and target echo; Feature extraction module: extracts high-dimensional features of target time series and sea clutter time series based on preprocessed data, and constructs feature data sets for training convolutional neural network models; Loss function setting module: Combines the focal loss function and cost-sensitive learning mechanism to set the loss function of the convolutional neural network model; Dynamic adjustment module: dynamically adjusts the misclassification cost value in the cost matrix to optimize the training process; Detection module: Use the optimized and trained convolutional neural network model to output the detection results.

8. A computer storage medium storing a readable program, characterized in that: When the program is running, it can execute the sea surface target detection method based on cost-sensitive focal loss convolutional neural network described in any one of claims 1 to 6.

9. An electronic device, characterized in that: include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the sea surface target detection method based on the cost-sensitive focal loss convolutional neural network as described in any one of claims 1-6.

10. A computer program product comprising computer instructions, characterized in that: The computer instructions instruct the computing device to perform operations corresponding to the sea surface target detection method based on the cost-sensitive focal loss convolutional neural network as described in any one of claims 1-6.

Images