A self-evolutionary detection method for sea surface targets using pulse compression radar

Abstract

The present invention discloses a self-evolutionary detection method for sea surface targets suitable for pulse compression radar. The method comprises: collecting unsupervised echo samples in the current working environment of the radar and performing preprocessing; for each preprocessed unsupervised echo sample, calculating the multi-expert joint information entropy of a pre-trained set of teacher models to achieve reliable sample screening; then performing data expansion based on the target image; optimizing the student model based on the expanded reliable samples through heterogeneous domain mixing loss to obtain a trained detection model; inputting the radar echo to be detected into the detection model after regularization processing to achieve sea surface target detection; the teacher model and the student model have the same structure and are both implemented by a convolutional neural network and an attention mechanism. The method of the present invention can perform high-precision radar sea surface target detection in a sea clutter background, improves detection accuracy, and is more suitable for detection of a variety of different sea surface targets and in strong sea clutter backgrounds.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing and deep learning technology, and in particular to a self-evolving detection method for sea surface targets suitable for pulse compression radar. Background Art

[0002] As an important means of detecting sea targets, radar is widely used due to its advantages, such as being unaffected by weather and capable of operating in all weather conditions and around the clock. However, under real-world conditions, radar's sea detection performance is limited by numerous interfering noise signals, the most typical of which is sea clutter. When a radar transmitter transmits a signal to scan the sea surface, the receiver receives backscattered echo signals from the sea surface, which is sea clutter. Complex sea clutter environments reduce the observability of sea targets and increase the difficulty of sea detection by radar systems, a situation that is often encountered in real-world applications. Therefore, the problem of sea target detection in the presence of sea clutter has become a difficult and hot topic of research.

[0003] In recent years, deep learning methods based on neural networks have made rapid progress in many pattern recognition tasks, such as computer vision and natural language processing. Complex deep neural network models, due to their powerful ability to automatically capture features, can be applied to a variety of downstream tasks. Inspired by this, some researchers have conducted research on radar target detection methods based on deep learning. By leveraging deep neural networks to automatically capture the differential features between target samples and clutter samples, these methods have achieved significant performance improvements in various metrics such as detection probability and false alarm rate. However, the performance of these methods is extremely dependent on the distribution of training data, and it is also difficult to achieve controllable low false alarm detection in practical application scenarios with variable sea conditions. For these reasons, the target warning function of radar monitoring systems in actual deployment is relatively difficult to implement, which is an urgent problem that needs to be solved.

[0004] Radar echo characteristics can vary significantly due to factors such as weather conditions and hardware conditions. In this field, the echoes collected by a radar system at a specific time (including ambient weather information) and under a specific operating state are referred to as a detection domain. The detection performance of existing deep learning-based methods is heavily dependent on the distribution of training samples, meaning they can only achieve accurate detection when the test sample and the training sample belong to the same or similar detection domain. In practical applications, the radar's operating conditions and the sea surface environment cannot be predicted in advance and may differ significantly from the training samples. Theoretically, it is impossible to exhaust all possible detection domains encountered in real-world applications to construct a dataset to train a neural network. Therefore, in order to achieve practical application, the core issue of deep neural network-based radar target detection algorithms is to improve their adaptability to different real-world scenarios. Only by developing radar sea surface target detection methods that can self-evolve based on the current detection domain and improve generalized detection performance can deep learning-based radar early warning systems be widely adopted. Summary of the Invention

[0005] In view of the technical defects of the existing radar target detection methods under sea clutter background, the purpose of the present invention is to overcome the above-mentioned technical defects and propose a self-evolving detection method for sea surface targets suitable for pulse compression radar.

[0006] To achieve the above object, the present invention proposes a self-evolving detection method for sea surface targets suitable for pulse compression radar, the method comprising:

[0007] Step 1) Collect unsupervised echo samples under the current working environment of the radar and perform preprocessing;

[0008] Step 2) For each unsupervised echo sample after preprocessing, the multi-expert joint information entropy is calculated for a set of pre-trained teacher models to achieve reliable sample screening;

[0009] Step 3) Perform data expansion based on the target image for the reliable samples obtained by screening;

[0010] Step 4) Based on the expanded reliable samples, the student model is optimized by heterogeneous domain mixing loss to obtain a trained detection model;

[0011] Step 5) the radar echo to be detected is input into the detection model after regularization processing to realize sea surface target detection;

[0012] The teacher model and the student model have the same structure and are both implemented through convolutional neural networks and attention mechanisms.

[0013] As an improvement to the above method, the detection model includes:

[0014] A stacked 4-layer convolutional network structure with a kernel size of 3 and a multi-head self-attention mechanism are used to encode the features of the input window signal; and

[0015] The fully connected layers and activation functions connected in sequence are used to make inference decisions on whether the current detection window contains valid sea surface targets.

[0016] As an improvement to the above method, the method further includes a step of training a set of teacher models, specifically including:

[0017] The teacher model is trained using historical supervised radar echo samples through binary classification loss. K model parameters are saved at different stages of the training process to obtain k teacher models, i.e. a group of teacher models, and the parameters of the student model are randomly initialized.

[0018] As an improvement to the above method, step 1) includes:

[0019] Echo samples are collected under the radar's current working environment, and each group of segment a is randomly divided into group b. Each segment of radar echo is processed by matched filtering. The amplitude of the filtering result is then normalized using a normalization operation. After normalization, the data amplitude is distributed between 0 and 1, completing the preprocessing.

[0020] Randomly read a set of pre-processed echo data, and divide the matched filtering results through a sliding detection window with a window length of m, and each time obtain a window sequence F consisting of sampling points with a length of m m As the unsupervised echo sample of step 2).

[0021] As an improvement to the above method, step 2) includes:

[0022] Step 2-1) For each unsupervised sample F m , first calculate the joint weighted confidence c(F m ):

[0023]

[0024] in, For the i-th teacher model window sequence F m The prediction results, λ i is the weight coefficient of the i-th teacher model decision;

[0025] Step 2-2) According to c(F m ) Calculate the joint information entropy H(F m ):

[0026] H(F m )=-[c(F m)log(c(F m ))+(1-c(F m ))log(1-c(F m ))]

[0027] Step 2-3) According to H(F m ), the target domain sample F is given by the following formula m Assign pseudo labels l(F m ):

[0028]

[0029] Where σ is the pre-set entropy threshold, sgn(·) represents the sign function;

[0030] If H(F m ) is greater than the entropy threshold, the pseudo label is recorded as -1, which is an unreliable echo sample and does not participate in the subsequent student model training steps. Otherwise, it is a reliable echo sample and participates in the subsequent student model training steps.

[0031] As an improvement to the above method, step 3) includes: executing three different sample expansion strategies for each reliable echo sample, including Gaussian noise superposition, echo sequence reorganization, and target unit reproduction, thereby achieving data expansion based on target mirroring; wherein,

[0032] The Gaussian noise superposition realizes data expansion by superimposing Gaussian noise with a specific signal-to-noise ratio on reliable echo samples;

[0033] The echo sequence reorganization is to realize data expansion by dividing the reliable echo samples into units and then shuffling and rearranging them;

[0034] The target unit reproduction is based on the original sea surface target sample and simulates samples of the target appearing at different positions to achieve data expansion.

[0035] As an improvement to the above method, the echo sequence reassembly process specifically includes:

[0036] The reliable echo samples are divided into units and then rearranged, and the length of the minimum division unit is taken as l c , the number of segmentation units is l m =m / l c , where m is the length of the echo sequence, and the variable l c Determines the granularity of segmentation and rearrangement. c =m, the rearranged sequence is equal to the original sequence; when l c =1, the rearranged sequence is most different from the original sequence.

[0037] As an improvement to the above method, the target unit reproduction processing specifically includes:

[0038] Take n near the target t sampling points, calculate their average power P0, randomly select an area in the echo and calculate its average power P1, set a random scaling factor and implement target embedding accordingly;

[0039] The sampling points at the original target unit are randomly assigned values ​​using the set random numbers to complete the erasure of the original target unit, thereby achieving data expansion.

[0040] As an improvement to the above method, step 4) specifically includes:

[0041] Through the expanded reliable samples, the semantic domain binary classification loss, feature domain distance loss and feature domain contrast loss are calculated respectively. The final optimization target of the student model is obtained by combining the three loss functions. The gradient descent method is used for repeated iterations until the optimal parameter combination is trained to obtain a trained detection model.

[0042] As an improvement to the above method, the step 5) of regularizing the radar echo to be detected specifically includes:

[0043] The radar echo to be detected is processed by matched filtering, and then the amplitude of the filtering result is regularized by normalization operation.

[0044] Compared with the prior art, the advantages of the present invention are:

[0045] 1. The method of the present invention can self-evolve detection model parameters based on unsupervised data collected by current pulse compression radar, thereby improving the generalization detection performance of existing radar target detection methods and being more suitable for sea surface radar target detection in complex situations.

[0046] 2. The radar target self-evolution detection technology provided by the present invention uses multi-expert joint information entropy to perform unsupervised sample screening. By calculating the multi-expert joint information entropy, we can automatically screen the unsupervised samples collected in the actual measurement environment and assign pseudo-labels, thus obtaining supervised data without manual labeling.

[0047] 3. The radar target self-evolution detection technology provided by the present invention uses target mirroring to expand the data of radar reliable samples. Based on the echo characteristics of pulse compression radar, this method can simulate and replicate sea surface targets in different signal-to-noise ratio environments and at different appearance locations, thereby effectively expanding the reliable samples.

[0048] 4. The radar target self-evolution detection technology provided by the present invention includes a student model optimization method based on heterogeneous domain mixing loss. This method can effectively utilize the expanded samples, fully exploit the rich information contained in the radar echo samples in the semantic domain and feature domain, and thus train a more powerful student model;

[0049] 5. When testing benchmark datasets, verification shows that the present invention achieves performance that exceeds existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the sea surface target self-evolution detection method of the present invention. DETAILED DESCRIPTION

[0051] The present invention proposes a self-evolving detection method for sea surface targets using pulse compression radar. The method uses a detection window to length-segment the preprocessed radar echo signal to obtain a window signal. A stacked convolutional network structure and an attention mechanism are used to encode the features of the window signal. Finally, a linear prediction module composed of a fully connected layer and an activation function is used to make an inference decision on whether a valid sea surface target is contained in the current detection window.

[0052] Specifically include:

[0053] Use historical supervised radar echo samples to train a set of teacher models and initialize the student models;

[0054] Collect unsupervised echo samples under the current working environment of the radar and preprocess them;

[0055] For each unsupervised echo sample, reliable samples are screened by calculating the joint information entropy of multiple experts;

[0056] For the reliable samples obtained through screening, data expansion based on the target image is performed;

[0057] Based on the expanded reliable samples, the student model is optimized through heterogeneous domain mixing loss;

[0058] Use the optimized student model to detect sea surface targets;

[0059] The teacher model and the student model are both implemented through structures such as convolutional neural networks and attention mechanisms.

[0060] As an improvement to the above method, the method of using historical supervised radar echo samples to train a group of teacher models specifically includes: saving models with different optimization levels as joint teacher models at different stages of training.

[0061] As an improvement to the above method, the method collects unsupervised echo samples in the current working environment of the radar and preprocesses them; specifically, the method includes:

[0062] First, the collected radar echo signal is processed by matched filtering, and then the normalization operation is used to regularize the amplitude of the radar echo signal after matched filtering. Then, the pre-processed echo information is read, and the matched filtering result is divided by a sliding detection window with a window length of m. Each time, a window sequence consisting of sampling points of length m is obtained as the subsequent single detection unit.

[0063] As an improvement to the above method, reliable sample screening is performed by calculating the joint information entropy of multiple experts; specifically, the method includes:

[0064] For each unsupervised sample, the joint weighted confidence is first calculated by the multi-teacher model:

[0065]

[0066] in, For the i-th teacher model to target domain samples The prediction results, λ i Represents the weight coefficient of the model decision. Then use this confidence to calculate the information entropy:

[0067]

[0068] Finally, reliable echo samples are screened out through a preset threshold.

[0069] As an improvement of the above method, the target mirror-based data expansion specifically includes: executing three different sample expansion strategies for each reliable sample, Gaussian noise superposition, echo sequence reorganization and target unit reproduction, so as to obtain different expanded samples.

[0070] The echo sequence reorganization specifically includes dividing the radar echo into units and then rearranging them. Assuming that the length of a minimum division unit is l c , the number of segmentation units is l m =m / l c , where m is the length of the echo sequence, and the variable l c Determines the granularity of segmentation and rearrangement. c =m, the rearranged sequence is equal to the original sequence; when l c =1, the rearranged sequence is most different from the original sequence.

[0071] The target unit reappears; specifically includes: first, the generation of a new target unit, taking n near the target tsampling points, calculate their average power P0, randomly select an area in the echo and calculate its average power P1, set a random scaling factor and implement target embedding accordingly; secondly, erase the original target unit and use the random number to randomly assign values ​​to the sampling points at the original target unit.

[0072] As an improvement to the above method, the method optimizes the student model by mixing losses in heterogeneous domains; specifically, the method includes: using the expanded feature samples to calculate the losses in the semantic domain and feature domain, and then optimizing the student model after mapping the two to the same plane.

[0073] The semantic domain loss is used to compare the binary classification prediction results output by the network with the pseudo labels, and to optimize the model through the binary classification loss;

[0074] The feature domain loss is used to compare the high-dimensional features output by networks containing different target samples, and to optimize the model through distance loss and contrast loss.

[0075] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0076] Example

[0077] like Figure 1 As shown, an embodiment of the present invention provides a pulse compression radar sea surface target self-evolution detection method, the method comprising:

[0078] Use historical supervised radar echo samples to train a set of teacher models and initialize the student models;

[0079] Collect unsupervised echo samples under the current working environment of the radar and preprocess them;

[0080] For each unsupervised echo sample, reliable samples are screened by calculating the joint information entropy of multiple experts;

[0081] For the reliable samples obtained through screening, data expansion based on the target image is performed;

[0082] Based on the expanded reliable samples, the student model is optimized through heterogeneous domain mixing loss;

[0083] Use the optimized student model to detect sea surface targets;

[0084] Both the teacher model and the student model are implemented using convolutional neural networks and attention mechanisms. Specifically, the architecture consists of a stacked four-layer convolutional network structure with a kernel size of 3 and a multi-head self-attention mechanism, which encodes the features of the input window signal. The fully connected layers and activation functions are then connected to make inferences about whether the current detection window contains a valid sea surface target.

[0085] The specific process of the pulse compression radar sea surface target self-evolution detection method includes:

[0086] Step 1) Use historical data to train the teacher model through binary classification loss, save k model parameters as the joint teacher model at different stages of the training process, and randomly initialize the parameters of the student model;

[0087] Step 2) Collect echo samples in the actual environment where the radar is currently deployed and perform preprocessing. The echo samples required for subsequent training are randomly divided into groups b according to each group a segment. Each radar echo segment is then subjected to matched filtering. The amplitude of the filtering result is then normalized. After normalization, the data amplitude is distributed between 0 and 1.

[0088] Step 3) Randomly read a set of pre-processed echo information, and divide the matched filtering results through a sliding detection window with a window length of m, and each time obtain a window sequence F consisting of sampling points with a length of m m As a subsequent single detection unit;

[0089] Step 4) The window signal obtained in step 3) is used to calculate the multi-expert joint information entropy through the joint teacher model. Based on the calculation result and the preset threshold, it is determined whether the current window signal is a reliable sample. If it is, the subsequent operation is performed; otherwise, the current sample is discarded and the calculation and judgment of the next sample are continued;

[0090] Step 5) Each reliable sample x obtained in step 4) is expanded based on the target image by three methods: signal-to-noise ratio superposition, callback sequence recombination, and target unit reproduction, to obtain x1, x2, and x3 respectively;

[0091] Step 6) Using the expanded samples obtained in step 5), the student model after parameter initialization is trained by the student model optimization method based on heterogeneous domain mixing loss.

[0092] Step 7) Update the parameters in the above model using the gradient descent method; iterate repeatedly until the optimal parameter combination is trained;

[0093] Step 8) Implement radar target detection in sea clutter background based on the optimal parameter model.

[0094] The step 4) specifically includes:

[0095] Step 4-1) Assumption For the i-th teacher model to target domain samples The prediction results, λ i Represents the weight coefficient of the model decision, then for a single sample, the joint confidence is calculated based on multiple teacher models:

[0096]

[0097] Step 4-2) Calculate the multi-expert joint information entropy based on the joint confidence:

[0098] H(F m )=-[c(F m )log(c(F m ))+(1-c(F m ))log(1-c(F m ))]

[0099] Step 4-3) Based on the calculated entropy value, we can use the following formula to give the target domain sample Assign pseudo labels:

[0100]

[0101] Where σ is the pre-set entropy threshold, and sgn(·) represents the sign function. If H(F m ) is greater than the entropy threshold, the pseudo label is an unreliable echo sample (recorded as -1 and does not participate in subsequent training steps), otherwise it is a reliable echo sample (participates in subsequent training steps).

[0102] The step 5) specifically includes:

[0103] Step 5-1) Determine the average power of the currently added Gaussian noise according to the preset signal-to-noise ratio η.

[0104]

[0105] By superimposing the Gaussian noise obtained at a specific signal-to-noise ratio on the original radar echo, we can obtain the data-enhanced echo sample:

[0106]

[0107] Step 5-2) Assume that the length of a minimum segmentation unit is l c , the number of segmentation units is l m =m / l c , where m is the length of the echo sequence, and the variable l c Determines the granularity of segmentation and rearrangement. c=m, the rearranged sequence is equal to the original sequence; when l c =1, the rearranged sequence is the most different from the original sequence. A random sequence of length m can be obtained through a random seed key:

[0108]

[0109] in is the pseudo-random generation algorithm used. With the help of this generation sequence, we rearrange the original radar echo to achieve data enhancement:

[0110]

[0111] Step 5-3) The main idea of ​​this strategy is to use the original sea surface target sample as the basis, and simulate the samples of the target appearing at different positions to achieve data enhancement. Specifically, it can be divided into two steps. The first step is to generate new target units, taking n near the target. t sampling points, calculate their average power P0, and then randomly select an area in the echo and calculate its average power P1:

[0112]

[0113] Set a random scaling factor:

[0114]

[0115] According to the scaling factor Embed in After executing the above process, the second step is to erase the original target cell. First calculate:

[0116]

[0117] Randomly generate a number RS ​​between 0 and Diff

[0118]

[0119] The random number is used to randomly assign values ​​to the sampling points at the original target unit, thereby completing the erasure of the original target unit.

[0120] The step 6) specifically includes:

[0121] Step 6-1) Calculate the semantic domain binary classification loss:

[0122]

[0123] Step 6-2) Calculate the feature domain distance loss:

[0124]

[0125] Step 6-3) Calculate the feature domain contrast loss based on infoNCE:

[0126]

[0127] Step 6-4) The final optimization goal is obtained by combining the above three loss functions:

[0128]

[0129] Among them, λ, β and γ are balance factors. We can use the gradient backpropagation algorithm to update and optimize the weight parameters of the student network.

[0130] In step 6-5), the parameters obtained in step 6-1) are used as the weight values ​​for this iteration; a group of echo samples are randomly selected from the remaining echo samples, and a new parameter combination is obtained through steps 2), 3), 4), 5), and 6-1); the iteration is repeated until one iteration cycle is completed;

[0131] Step 6-6) Reshuffle the training echo samples and go to step 1); repeat until the optimal parameter combination is trained.

[0132] Step 7) specifically includes:

[0133] Step 7-1) For the radar echo to be detected, first perform matched filtering processing, and then use normalization operation to regularize the amplitude of the echo information;

[0134] Step 7-2) Input the regularized radar echo into the optimal parameter model and obtain the prediction result through forward transmission of the model;

[0135] Step 7-3) Compare the category corresponding to the maximum probability value of the predicted result of each radar echo segment with the true label. If they are consistent, the prediction is correct; otherwise, the prediction is wrong.

[0136] Step 7-4) executes step 7-1), step 7-2) and step 7-3) until all echoes to be measured are measured.

[0137] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A self-evolving detection method for sea surface targets applicable to pulse compression radar, the method comprising: Step 1) Collect unsupervised echo samples under the current working environment of the radar and perform preprocessing; Step 2) For each unsupervised echo sample after preprocessing, the multi-expert joint information entropy is calculated for a set of pre-trained teacher models to achieve reliable sample screening; Step 3) Perform data expansion based on the target image for the reliable samples obtained by screening; Step 4) Based on the expanded reliable samples, the student model is optimized by heterogeneous domain mixing loss to obtain a trained detection model; Step 5) the radar echo to be detected is input into the detection model after regularization processing to realize sea surface target detection; The teacher model and the student model have the same structure and are both implemented through convolutional neural networks and attention mechanisms; The step 1) comprises: Echo samples are collected under the radar's current working environment, and each group of segment a is randomly divided into group b. Each segment of radar echo is processed by matched filtering. The amplitude of the filtering result is then normalized using a normalization operation. After normalization, the data amplitude is distributed between 0 and 1, completing the preprocessing. Randomly read a set of pre-processed echo data, and divide the matched filtering results through a sliding detection window with a window length of m, and each time obtain a window sequence F consisting of sampling points with a length of m m As the unsupervised echo sample of step 2); The step 2) comprises: Step 2-1) For each unsupervised sample F m , first calculate the joint weighted confidence c(F m ): in, For the i-th teacher model window sequence F m The prediction results, λ i is the weight coefficient of the i-th teacher model decision; Step 2-2) According to c(F m ) Calculate the joint information entropy H(F m ): H(F m )=-[c(F m )log(c(F m ))+(1-c(F m ))log(1-c(F m ))] Step 2-3) According to H(F m ), the target domain sample F is given by the following formula m Assign pseudo labels l(F m ): Where σ is the pre-set entropy threshold, sgn(·) represents the sign function; If H(F m ) is greater than the entropy threshold, the pseudo label is recorded as -1, which is an unreliable echo sample and does not participate in the subsequent student model training steps. Otherwise, it is a reliable echo sample and participates in the subsequent student model training steps; The step 3) includes: executing three different sample expansion strategies for each reliable echo sample, including Gaussian noise superposition, echo sequence reorganization and target unit reproduction, thereby achieving data expansion based on the target image; wherein, The Gaussian noise superposition realizes data expansion by superimposing Gaussian noise with a specific signal-to-noise ratio on reliable echo samples; The echo sequence reorganization is to realize data expansion by dividing the reliable echo samples into units and then shuffling and rearranging them; The target unit reproduction is based on the original sea surface target sample and simulates samples of the target appearing at different positions to achieve data expansion.

2. The sea surface target self-evolution detection method suitable for pulse compression radar according to claim 1, characterized in that: The detection model includes: A stacked 4-layer convolutional network structure with a kernel size of 3 and a multi-head self-attention mechanism are used to encode the features of the input window signal; and The fully connected layers and activation functions connected in sequence are used to make inference decisions on whether the current detection window contains valid sea surface targets.

3. The sea surface target self-evolution detection method suitable for pulse compression radar according to claim 1, characterized in that: The method further includes a step of training a set of teacher models, specifically comprising: The teacher model is trained using historical supervised radar echo samples through binary classification loss. K model parameters are saved at different stages of the training process to obtain k teacher models, i.e. a group of teacher models, and the parameters of the student model are randomly initialized.

4. The sea surface target self-evolution detection method applicable to pulse compression radar according to claim 1, characterized in that: The processing of the echo sequence reorganization specifically includes: The reliable echo samples are divided into units and then rearranged, and the length of the minimum division unit is taken as l c , the number of segmentation units is l m =m1 / l c , where m1 is the echo sequence length, variable l c Determines the granularity of segmentation and rearrangement. c =m1, the rearranged sequence is equal to the original sequence; when l c =1, the rearranged sequence is most different from the original sequence.

5. The sea surface target self-evolution detection method applicable to pulse compression radar according to claim 1, characterized in that: The target unit recurrence processing specifically includes: Take n near the target t sampling points, calculate their average power P0, randomly select an area in the echo and calculate its average power P1, set a random scaling factor and implement target embedding accordingly; The sampling points at the original target unit are randomly assigned values ​​using the set random numbers to complete the erasure of the original target unit, thereby achieving data expansion.

6. The sea surface target self-evolution detection method applicable to pulse compression radar according to claim 1, characterized in that: The step 4) specifically includes: Through the expanded reliable samples, the semantic domain binary classification loss, feature domain distance loss and feature domain contrast loss are calculated respectively. The final optimization target of the student model is obtained by combining the three loss functions. The gradient descent method is used for repeated iterations until the optimal parameter combination is trained to obtain a trained detection model.

7. The sea surface target self-evolution detection method applicable to pulse compression radar according to claim 1, characterized in that: The step 5) of regularizing the radar echo to be detected specifically includes: The radar echo to be detected is processed by matched filtering, and then the amplitude of the filtering result is regularized by normalization operation.

Images