Radar target detection method in sea clutter background based on CNN and dual-view attention

Abstract

The present invention discloses a radar target detection method in sea clutter backgrounds based on a CNN and dual-view attention. The method comprises: preprocessing the radar echo signal to be detected; segmenting the preprocessed radar echo according to the detection window length to obtain a window signal; inputting the window signal into a pre-established and trained radar target detection model to determine whether the detected signal contains a sea surface target; the radar target detection model is implemented using a CNN and dual-view attention. The method of the present invention enables high-precision radar sea surface target detection in sea clutter backgrounds, improving detection accuracy and being more suitable for detecting a variety of different sea surface targets and in strong sea clutter backgrounds. Testing has verified that the present invention achieves detection performance that exceeds existing methods.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing and deep learning technology, and in particular to a radar target detection method under sea clutter background based on CNN and dual-view attention. Background Art

[0002] Offshore surface target detection plays a vital role in monitoring stowaways, intrusion detection, and maritime search and rescue. As the primary means of detecting surface targets, radar is widely used due to its advantages, such as being unaffected by the environment and weather. In the military, sea detection radars typically use pulse Doppler radars, which have a longer observation range, typically exceeding several hundred kilometers. However, these radars are complex and require high deployment costs. Compared to pulse Doppler radars, pulse compression radars are less susceptible to sidelobe clutter suppression, offer better range resolution, and impose lower overall radar design performance requirements, resulting in lower costs. Therefore, pulse compression radars are currently the most widely used in civilian offshore surface target detection. Radar's sea detection performance is limited by numerous interfering signals, such as clutter and noise. When a radar transmitter transmits a signal to scan the sea surface, the receiver receives backscattered echo signals from the surface, known as sea clutter. Complex sea clutter environments reduce the observability of surface targets, making sea detection more challenging for radar systems. Therefore, effective detection of surface targets in the presence of sea clutter has attracted widespread attention from researchers.

[0003] In recent years, deep learning technology has achieved success in many fields, including computer vision, natural language understanding, and speech recognition. With careful design, deep neural networks (DNNs) have excellent capabilities for extracting high-dimensional features from raw data. Inspired by this, many researchers have incorporated DNNs into radar signal processing models, achieving performance improvements over traditional algorithms. However, considering practical application scenarios, applying deep learning to marine surface target detection still faces the following two key challenges.

[0004] First, the complexity of the ocean environment and the diversity of sea surface targets place significant demands on the generalization capabilities of detection models. Under the influence of varying sea clutter, the signal-to-clutter ratio (SCR) of the same target can vary dramatically. In high sea conditions, sea spikes appear on the surface, easily obscuring surface targets in the background noise. Furthermore, sea surface targets vary, ranging from small targets like rafts and buoys to large targets like cruise ships and cargo ships. These targets exhibit varying echo characteristics in different ocean environments, posing significant challenges for detection. Second, a reliable and robust radar system requires a high false alarm rate. Radar target detection involves testing the hypothesis of "target presence" in noisy radar echoes. Compared to typical binary classification detection scenarios, radar systems prioritize misclassification of non-target echoes, requiring them to maximize the detection probability while limiting the false alarm probability. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a radar target detection method in sea clutter background based on CNN and dual-view attention.

[0006] To achieve the above objectives, the present invention proposes a radar target detection method in sea clutter background based on CNN (Convolutional Neural Networks) and dual-view attention, the method comprising:

[0007] Preprocessing the radar echo signal to be detected;

[0008] The pre-processed radar echo is divided according to the detection window length to obtain the window signal;

[0009] The window signal is input into a pre-established and trained radar target detection model to obtain a decision on whether the signal to be detected contains a sea surface target;

[0010] The radar target detection model is implemented by CNN and dual-view attention.

[0011] As an improvement to the above method, the original radar echo information to be detected is preprocessed; specifically, the preprocessing includes: first performing matched filtering on the radar echo signal to be detected, and then using a normalization operation to regularize the amplitude of the radar echo signal after the matched filtering.

[0012] As an improvement to the above method, it is characterized in that the pre-processed radar echo is divided according to the detection window length to obtain the window signal; specifically comprising:

[0013] The pre-processed radar echo signal is divided by a sliding detection window with a window length of m, and a window signal F consisting of sampling points of length m is obtained each time.m .

[0014] As an improvement to the above method, the radar target detection model includes a dynamic convolution module, a dual-view attention module, and a linear prediction module connected in series; wherein,

[0015] The dynamic convolution module is used to adaptively execute different feature encoding strategies in the form of weighted filter banks according to different input window signals, so as to extract effective high-dimensional feature information for sea surface targets of different sizes;

[0016] The dual-view attention module is used to extract feature correlation information from the high-dimensional features obtained after the dynamic convolution module from both global and local perspectives and to fuse the correlation information to obtain correlation features;

[0017] The linear prediction module is used to comprehensively utilize the extracted correlation features to determine whether the current window signal contains a valid sea surface target.

[0018] As an improvement to the above method, the input of the dynamic convolution module is the window signal F m , the output is the encoding feature A, and the dynamic convolution module includes an adaptive convolution layer, a batch normalization layer, an adaptive convolution layer, a batch normalization layer and an adaptive convolution layer connected in series.

[0019] As an improvement to the above method, the adaptive convolution layer is used to assign different weights to multiple convolution filters in the same layer according to the input signal through a compression-excitation method. The processing process of the adaptive convolution layer specifically includes:

[0020] The features of the input signal are pooled by global average pooling to obtain the pooled features f ap , the feature f ap Input into the linear network composed of two fully connected layers and ReLU to obtain the activation feature f l , a set of weights w1, w2, ..., w is obtained through the Softmax activation function k , these weights are fused with k filters to obtain the final weighted convolution filter, and the results obtained by passing the features of the input signal through the weighted convolution filter are used as the output of the adaptive convolution layer.

[0021] As an improvement to the above method, the input of the dual-view attention module is the encoding feature A, and the output is the correlation feature C. The dual-view attention module includes a global correlation attention module, a local correlation attention module, and a correlation fusion module. The processing process of the dual-view attention module specifically includes:

[0022] The output feature A of the dynamic convolution module is subjected to global correlation attention to obtain the global correlation feature G, the local correlation feature L is obtained through the local correlation attention module, and the correlation feature C is obtained by cross-fusion of G and L through the correlation fusion module.

[0023] As an improvement to the above method, the method further includes a step of training a radar target detection model; specifically, the step includes:

[0024] Step 1) establishing a sample set, randomly dividing the radar data in the sample set into m groups according to each group of a segments; the sample set includes a number of radar echoes and corresponding labels indicating whether they contain valid sea surface targets;

[0025] Step 2) Randomly select a set of echo data, input the preprocessed data into the radar target detection model, obtain the prediction result, and use cross entropy to calculate the error between the prediction result and the true label to obtain the true label The error between the predicted result y Repeat step 2) until the sample training is completed, then go to step 3);

[0026] Step 3) According to the error Select the optimal parameter combination to obtain a trained radar target detection model.

[0027] On the other hand, the present invention proposes a radar target detection system in sea clutter background based on CNN and dual-view attention, the system comprising:

[0028] A preprocessing module, used for preprocessing the radar echo signal to be detected;

[0029] A segmentation module is used to segment the pre-processed radar echo according to the detection window length to obtain a window signal;

[0030] The detection output module is used to input the window signal into a pre-established and trained radar target detection model to obtain a decision on whether the signal to be detected contains a sea surface target;

[0031] The radar target detection model is implemented by CNN and dual-view attention.

[0032] Compared with the prior art, the advantages of the present invention are:

[0033] 1. The method of the present invention can perform high-precision radar target detection under sea clutter background, improves detection accuracy, and is more suitable for sea surface radar target detection under complex conditions;

[0034] 2. The radar target detection technology solution in sea clutter background provided by the present invention uses dynamic convolution to perform adaptive feature encoding on the input signal. It can adaptively adjust the weight parameters of the filter bank according to the echo signal received by the radar system, thereby achieving generalized and effective detection of various sea surface targets in complex ocean environments.

[0035] 3. Based on the obtained high-dimensional coding features, a dual-view correlation mining module is used to extract correlation features from the initially encoded radar echo signal from both global and local perspectives using an attention method. The correlation aggregation module then organically integrates the two different types of correlation information.

[0036] 4. When testing benchmark datasets, the present invention achieves performance that exceeds existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the radar target detection method in sea clutter background based on CNN and dual-view attention of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0039] Example 1

[0040] like Figure 1 As shown, embodiment 1 of the present invention proposes a radar target detection method in a sea clutter background based on CNN and dual-view attention, the method comprising:

[0041] Preprocessing the radar echo signal to be detected;

[0042] The pre-processed radar echo is divided according to the detection window length to obtain the window signal;

[0043] The window signal is input into a pre-established and trained radar target detection model to obtain a decision on whether the signal to be detected contains a sea surface target;

[0044] The radar target detection model is implemented by CNN and dual-view attention.

[0045] The radar target detection method under sea clutter background proposed in this invention completes data processing by a detection window, a dynamic convolution module, a dual-view attention module, and a linear prediction module. The specific process includes:

[0046] Step 1) Data preprocessing: The echo samples required for training are randomly divided into n groups with a segment per group, and each segment of radar echo is processed by matched filtering. Then, the amplitude of the filtering result is regularized by normalization operation. After regularization, the data amplitude is distributed between 0 and 1;

[0047] Step 2) 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;

[0048] Step 3) The window signal F obtained in step 2) is converted to m , high-dimensional feature encoding is performed through the adaptive convolution layer, batch normalization layer, adaptive convolution layer, batch normalization layer and adaptive convolution layer connected in series in the dynamic convolution module to obtain the encoded feature A.

[0049] Step 4) The encoding feature A obtained in step 3) is subjected to global correlation extraction and local correlation extraction through different attention mechanisms, capturing the feature correlation information of the original radar echo in high-dimensional space from two different perspectives to obtain the global correlation feature f g and local correlation features f l Perform cross fusion processing to integrate the related features of the same echo band from different perspectives to obtain the fused dependent features f c .

[0050] Step 5) The feature map f obtained in step 4) is c , transform it into a one-dimensional feature, and obtain the prediction result of whether the current radar echo window contains a valid sea surface target through the linear prediction module.

[0051] Step 6) Update the parameters in the above model using the gradient descent method; iterate repeatedly until the optimal parameter combination is trained;

[0052] Step 7) Implement radar target detection in sea clutter background based on the optimal parameter model.

[0053] The method includes: dividing the length of the pre-processed radar echo signal through a detection window to obtain a window signal; using a dynamic convolution module to adaptively encode the radar echo window signal, and this process is mainly achieved by weighted fusion of multiple convolution filters in the same network layer according to different inputs, and using the fused filters to implement feature encoding; based on the obtained encoding features, a dual-view attention module is used to extract global and local correlation information, and the extraction of these two types of correlation information is mainly completed through the global correlation attention and local correlation attention mechanisms, and a correlation fusion module is used to integrate the respectively extracted global and local correlation features; finally, a linear prediction module composed of a fully connected layer and an activation function is used to make an inference decision on whether the current detection window contains a valid sea surface target.

[0054] like Figure 1As shown in the figure, the dynamic convolution module proposed in the present invention consists of an adaptive convolution layer and a batch normalization layer.

[0055] The step 3) specifically includes:

[0056] Step 3-1) Create k convolution kernels with different receptive fields, according to the input feature F m The dynamic weight is obtained through the compression-excitation mechanism. The calculation process of compression-excitation is as follows: First, use the average pooling operation to m Perform dimensionality reduction:

[0057] P=AvgPool(F m )

[0058] The obtained feature map P is then fed into two fully connected layers:

[0059] f P =W A2 (W A1 (P)+b A1 )+b A2

[0060] Where W A1 and W A2 They are the weight parameters of the two fully connected layers, b A1 and b A2 is the bias parameter. The number of neurons in the second fully connected layer is four times that of the first layer.

[0061] Step 3-2) Use Softmax function to calculate f P Perform normalized activation processing to obtain the weight coefficients w1, w2, ..., w of the filter group k , use these parameters to perform weighted fusion of k convolution kernels:

[0062]

[0063] Where W i and b i Represents the learnable parameters in the convolution kernel. Using the fused filter parameters W A and b A , we can perform effective feature encoding:

[0064]

[0065] Among them, σ represents the activation function, Represents batch normalization processing

[0066] Step 3-3) Take A1 as input and perform the operations of steps 3-1) and 3-2) to obtain A2

[0067] Step 3-4) Take A2 as input and perform the operations of steps 3-1) and 3-2) to obtain A3

[0068] Step 3-5) Perform the maximum pooling operation on A3 to obtain A as the final encoding feature.

[0069] like Figure 1 As shown in the figure, the dual-view attention module proposed in the present invention consists of global correlation attention, local correlation attention and correlation fusion modules.

[0070] The step 4) specifically includes:

[0071] Step 4-1) Input the encoded feature A obtained in step 3) into the global correlation attention to extract global correlation. First, three different fully connected layers are used to perform different linear mappings on C:

[0072]

[0073] in, is the weight matrix of the i-th head. Then the following sequence of operations is performed:

[0074]

[0075] Among them head (i) Represents the attention operation result of the i-th head, is the scale factor. The attention operation results of different heads are then concatenated and added to the original encoding features through a skip connection:

[0076] G=Concat(head (1) , head (2) ,…,head (h) )W O +A

[0077] Where W O is the weight matrix. G is the global correlation feature finally extracted.

[0078] In step 4-2), the encoding feature A obtained in step 3) is input into the local correlation attention for local correlation extraction. First, the original encoding feature is aggregated and dimensionally reduced using average pooling and maximum pooling. The two results are concatenated and then passed through convolution layers with different kernel sizes and sigmoid activation function layers to obtain the attention weight coefficient. This coefficient is multiplied point by point with the original feature to obtain the weighted feature map. The calculation formula is as follows:

[0079] F p =Concat(AvgPool(A), MaxPool(A))

[0080] M A =σ(F s )

[0081] L=M A ⊙F s +A

[0082] Among them, AvgPool and MaxPool represent average pooling and maximum pooling operations, F s Represents the obtained attention weight coefficient, σ is the activation function, ⊙ represents point-by-point multiplication, and L represents the local correlation feature obtained after weighting by the attention weight.

[0083] Step 4-3) The global correlation features and local correlation features obtained in step 4-1) are fused. First, the two features are spliced:

[0084] F C =concat(G,L)

[0085] The feature map F C Through two fully connected layers:

[0086] F f1 =W c2 (σ(W c1 (F C )+b c1 ))+b c2

[0087] Where W c1 and W c2 is the weight parameter, b c1 and b c2 is the bias parameter. σ represents the activation function.

[0088] In addition, the feature map F C Through three convolutional layers

[0089]

[0090] in, and is the convolution operation, and σ is the activation function. The feature maps obtained from the two branches are concatenated to obtain the final correlation features:

[0091] C=concat(F f1 , F f2 )

[0092] The linear prediction module includes a fully connected layer and a Sigmoid function. The fully connected layer includes two neurons to achieve feature mapping of high-level abstract features to the output layer output results. The Sigmoid normalization function is used for normalization processing to obtain a classification result of whether the input radar echo signal contains secret information.

[0093] Step 6) specifically includes:

[0094] Step 6-1) Calculate the error between the model output and the true label; the error is calculated using cross entropy, which can be expressed as:

[0095]

[0096] Among them, Los(y,y) represents the error between the true label y and the predicted result y, p(x i ) represents the forward output of the model, q(x i ) represents the true label, Σ represents the sum;

[0097] In step 6-2), 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;

[0098] Step 6-3) Reshuffle the training echo samples and go to step 1); repeat until the optimal parameter combination is trained.

[0099] Step 7) specifically includes:

[0100] 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;

[0101] Step 7-2) Input the regularized radar echo into the optimal parameter model and obtain the prediction result through forward transmission of the model;

[0102] 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.

[0103] Step 7-4) executes step 7-1), step 7-2) and step 7-3) until all echoes to be measured are measured.

[0104] Example 2

[0105] Embodiment 2 of the present invention proposes a radar target detection system in a sea clutter background based on CNN and dual-view attention, which is implemented based on the method of embodiment 1. The system includes:

[0106] A preprocessing module, used for preprocessing the radar echo signal to be detected;

[0107] A segmentation module is used to segment the pre-processed radar echo according to the detection window length to obtain a window signal;

[0108] The detection output module is used to input the window signal into a pre-established and trained radar target detection model to obtain a decision on whether the signal to be detected contains a sea surface target;

[0109] The radar target detection model is implemented by CNN and dual-view attention.

[0110] 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 radar target detection method in sea clutter background based on CNN and dual-view attention, the method comprising: Preprocessing the radar echo signal to be detected; The pre-processed radar echo is divided according to the detection window length to obtain the window signal; The window signal is input into a pre-established and trained radar target detection model to obtain a decision on whether the signal to be detected contains a sea surface target; The radar target detection model is implemented by CNN and dual-view attention; The radar target detection model includes a dynamic convolution module, a dual-view attention module and a linear prediction module connected in series; wherein, The dynamic convolution module is used to adaptively adjust the input window signal F m Different feature encoding strategies are implemented in the form of weighted filter banks, so that effective high-dimensional feature information can be extracted for sea surface targets of different sizes. The dual-view attention module is used to extract feature correlation information from the high-dimensional features obtained after the dynamic convolution module from both global and local perspectives and to fuse the correlation information to obtain correlation features; The linear prediction module is used to comprehensively utilize the extracted correlation features to determine whether the current window signal contains a valid sea surface target.

2. The radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 1 is characterized in that The preprocessing of the original radar echo information to be detected specifically includes: firstly performing matched filtering processing on the radar echo signal to be detected, and then using a normalization operation to regularize the amplitude of the radar echo signal after the matched filtering processing.

3. The radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 2 is characterized in that The pre-processed radar echo is divided according to the detection window length to obtain the window signal; specifically comprising: The pre-processed radar echo signal is divided by a sliding detection window with a window length of m, and a window signal F consisting of sampling points of length m is obtained each time. m .

4. The radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 1 is characterized in that The input of the dynamic convolution module is the window signal F m , the output is the encoding feature A, and the dynamic convolution module includes an adaptive convolution layer, a batch normalization layer, an adaptive convolution layer, a batch normalization layer and an adaptive convolution layer connected in series.

5. The radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 4 is characterized in that: The adaptive convolution layer is used to assign different weights to multiple convolution filters in the same layer through a compression-excitation method according to the input signal. The processing process of the adaptive convolution layer specifically includes: The features of the input signal are pooled by global average pooling to obtain the pooled features f ap , the feature f ap Input into the linear network composed of two fully connected layers and ReLU to obtain the activation feature f l , a set of weights w1,w2,…,w is obtained through the Softmax activation function k , these weights are fused with k filters to obtain the final weighted convolution filter, and the results obtained by passing the features of the input signal through the weighted convolution filter are used as the output of the adaptive convolution layer.

6. The radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 1 is characterized in that The input of the dual-view attention module is the encoding feature A, and the output is the correlation feature C. The dual-view attention module includes a global correlation attention module, a local correlation attention module, and a correlation fusion module. The processing process of the dual-view attention module specifically includes: The output feature A of the dynamic convolution module is subjected to global correlation attention to obtain the global correlation feature G, the local correlation feature L is obtained through the local correlation attention module, and the correlation feature C is obtained by cross-fusion of G and L through the correlation fusion module.

7. The radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 1 is characterized in that The method also includes a radar target detection model training step; specifically including: Step 1) establishing a sample set, randomly dividing the radar data in the sample set into m groups according to each group of a segments; the sample set includes a number of radar echoes and corresponding labels indicating whether they contain valid sea surface targets; Step 2) Randomly select a set of echo data, input the preprocessed data into the radar target detection model, obtain the prediction result, and use cross entropy to calculate the error between the prediction result and the true label to obtain the true label The error between the predicted result y Repeat step 2) until the sample training is completed, then go to step 3); Step 3) According to the error Select the optimal parameter combination to obtain a trained radar target detection model.

8. A system based on the radar target detection method in sea clutter background based on CNN and dual-view attention according to claim 1, characterized in that: The system comprises: A preprocessing module, used for preprocessing the radar echo signal to be detected; A segmentation module, configured to segment the pre-processed radar echo according to the detection window length to obtain a window signal; and The detection output module is used to input the window signal into a pre-established and trained radar target detection model to obtain a decision on whether the signal to be detected contains a sea surface target; The radar target detection model is implemented by CNN and dual-view attention.

Images