SAR active interference segmentation method based on mixed attention and multiple loss constraint network

Abstract

The invention relates to an SAR active interference segmentation method based on a mixed attention and multiple loss constraint network. The method specifically comprises the following steps: firstly, obtaining radar echo time-frequency maps under different interferences; generating a channel attention matrix by using a channel attention module (extracting global and local information through maximum / average pooling); a spatial attention weight matrix is generated through a spatial attention module (spatial convolution after pooling); taking the weighted feature map as input, and extracting low-resolution features through an interference feature extraction network; a high-resolution interference segmentation image is reconstructed through up-sampling of the interference feature recovery network; and finally, outputting a classification result by the pixel-level classification network. According to the scheme, accurate interference segmentation is realized by fusing a double attention mechanism and a coding and decoding structure.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a SAR active jamming segmentation method based on a hybrid attention and multiple loss constraint network. BACKGROUND

[0002] SAR has been widely applied in civil fields due to its high resolution and all-weather and all-day advantages, but also faces serious challenges from various types of active jamming in the environment. Active jamming seriously affects the quality of SAR imaging and restricts the information acquisition capability of SAR. In order to improve the SAR jamming suppression capability, it is necessary to perceive the jamming, which can not only guide the SAR system to select the optimal anti-jamming method, but also assist in the design of anti-jamming method. Compared with jamming detection and jamming recognition, jamming segmentation can not only provide the category information of jamming, but also provide the location information of jamming. Therefore, in order to ensure the information acquisition capability of SAR system in complex electromagnetic environment, it is of great significance and application value to study the jamming segmentation algorithm.

[0003] Most of the existing SAR jamming segmentation algorithms are used as pre-task of jamming suppression algorithm to locate the position, frequency and bandwidth of jamming. And in the segmentation process, the type of jamming is single, which can only be used in the process of jamming suppression to assist the use of specific type of jamming, and it is difficult to be popularized to the segmentation task of multiple types of jamming in complex electromagnetic environment. At the same time, the existing algorithm also has the problems of fuzzy boundary division and positioning deviation in the segmentation process. SUMMARY

[0004] In order to solve the above problems in the prior art, the application provides a SAR active jamming segmentation method based on a hybrid attention and multiple loss constraint network. The technical problems to be solved by the application are realized by the following technical scheme: The application provides a SAR active jamming segmentation method based on a hybrid attention and multiple loss constraint network, which comprises: The radar echoes generated under different interference conditions are acquired to generate multiple two-dimensional time-frequency diagrams; a channel attention module is used to perform maximum pooling and average pooling on each two-dimensional time-frequency diagram to obtain global information and local information, and a channel attention matrix is obtained based on the global information and the local information, each element in the channel attention matrix falling within an open interval (0, 1); a spatial attention module is used to perform maximum pooling and average pooling on the channel attention matrix to obtain a spatial information matrix, and a spatial attention weight matrix is obtained by performing spatial dimension convolution on the spatial information matrix, each element in the spatial attention weight matrix falling within an open interval (0, 1); an interference feature extraction network is used to extract features of an input graph containing interference features to obtain an interference time-frequency spectrum with a resolution less than a first threshold; the input graph containing interference features is obtained based on the spatial attention weight matrix; an interference feature recovery network is used to recover features of the interference time-frequency spectrum to obtain an interference segmentation graph with a resolution greater than a second threshold, the first threshold being less than the second threshold; a pixel-level interference classification network is used to identify the interference segmentation graph to obtain multiple probability values, and the maximum value in the multiple probability values is selected as a classification result of the interference segmentation graph.

[0005] Compared with the prior art, the present application has the following advantages: (1) The SAR active jamming segmentation method based on mixed attention and multiple loss constraint network has the core advantage of significantly improving the jamming segmentation accuracy and robustness. By cooperatively using the channel attention mechanism and the spatial attention mechanism, the method realizes fine extraction of interference features: the channel attention dynamically enhances the key interference feature channel and suppresses noise, and the spatial attention accurately locates the significant spatial area of the interference, so that the model can accurately distinguish the interference types under the suppression interference (such as comb spectrum), deception interference (such as dense false target) and composite interference scenes, effectively solving the similar interference confusion and boundary ambiguity problem. Ablation experiments verify that the mixed attention significantly improves the average segmentation accuracy (AA) and significantly optimizes the boundary clarity.

[0006] (2) The method effectively solves the key technical difficulties in jamming segmentation. The encoder-decoder structure cooperates with the learnable deconvolution upsampling to effectively recover the spatial details of low-resolution features; the triple loss constraint (cross-entropy loss to ensure global classification correctness, set similarity loss to optimize shape matching, and pixel-level reconstruction error to strengthen detail recovery) cooperatively optimizes the model, significantly improving the boundary alignment accuracy. Even under extreme conditions of low jam-to-noise ratio (JNR=0dB), the model can still maintain a segmentation accuracy of more than 98%, showing strong extraction ability and anti-confusion ability for weak interference features.

[0007] (3) The differences of different types of interference and target echoes in high-dimensional features are strengthened by using channel attention and spatial attention mechanisms, the feature extraction capability of the interference segmentation model for interference signals is improved, the boundary between interference and target echoes is refined, and a multiple loss function is introduced to enhance the network parameter updating and interference feature self-calibration capability, further optimize the performance of the interference segmentation model, and improve the segmentation accuracy and robustness of the network. This method can solve the problems of fuzzy segmentation boundary and poor updating and calibration capability of existing interference segmentation methods when facing increasingly complex and variable active interference.

[0008] (4) Channel and spatial attention mechanism modules are added in the down-sampling module and the up-sampling module to improve the extraction and recovery capability of the interference segmentation network model for interference features, and to make the interference segmentation boundary clear.

[0009] (5) A full convolution architecture is used to realize end-to-end pixel-level prediction. The core is to recover the semantic features of the deep spatial dimension through transpose convolution operation: first, the high-level abstract features are obtained through multi-level down-sampling of the encoder (convolution part), then the low-resolution feature map output by the final convolution is progressively up-sampled through the decoder (deconvolution part), and finally the prediction map corresponding to the input image size is reconstructed. This full convolution architecture has double advantages: first, by eliminating the restriction of full connection layer on input size, the model can process image input of any resolution; second, the feature map space correspondence mechanism in the encoding and decoding structure effectively preserves the original spatial information, so that each pixel point can obtain semantic class prediction based on its local context information. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of the SAR active interference segmentation method based on the mixed attention and multiple loss constraint network provided by the embodiment of the present application; Figure 2 is a part of the data set sample example graph of eight radar interference time-frequency graphs provided by the embodiment of the present application; Figure 3 is a connection diagram of related modules in steps 120-160 provided by the embodiment of the present application; Figure 4 is a data processing flowchart of the channel attention module provided by the embodiment of the present application; Figure 5 is a data processing flowchart of the spatial attention module provided by the embodiment of the present application; Figure 6 is a comparison graph of the comb spectrum interference segmentation visualization results of the existing basic framework network UNet and the present application provided by the embodiment of the present application; Figure 7is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the dense false target jamming segmentation of the application; Figure 8 is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the intermittent sampling and forwarding jamming segmentation of the application; Figure 9 is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the multi false target jamming segmentation of the application; Figure 10 is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the sinusoidal frequency sweeping jamming segmentation of the application; Figure 11 is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the sawtooth frequency sweeping jamming segmentation of the application; Figure 12 is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the comb spectrum jamming and intermittent sampling and forwarding jamming compound segmentation of the application; Figure 13 is the existing basic framework network UNet provided by the embodiment of the application and the comparison chart of the visualization results of the sinusoidal frequency sweeping jamming and multi false target jamming compound segmentation of the application; Figure 14 is the example chart of the confusion matrix obtained after the GCnet, Deeplabv3+ and Segformer and the method provided by the application perform the jamming segmentation task on the test set samples provided by the embodiment of the application; Figure 15 is the result of each jamming segmentation evaluation index obtained after the method provided by the application and the GCnet, Deeplabv3+ and Segformer are tested under different JNR provided by the embodiment of the application; Figure 16 is the result of each jamming segmentation evaluation index obtained after the method provided by the application and the GCnet, Deeplabv3+ and Segformer are tested under different sample training sets provided by the embodiment of the application. DETAILED DESCRIPTION

[0011] The application will be further described in detail below in combination with specific embodiments, but the embodiments of the application are not limited thereto.

[0012] In the description of the present application, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0013] An SAR active jamming segmentation method based on a hybrid attention and multiple loss constraint network is described in detail in combination with the drawings.

[0014] Figure 1 is a flowchart of the SAR active jamming segmentation method based on the hybrid attention and multiple loss constraint network provided by the embodiments of the present application. As shown in Figure 1 , the SAR active jamming segmentation method comprises: Step 110: Obtain multiple two-dimensional time-frequency graphs generated by radar echoes under different jamming conditions.

[0015] Specifically, step 110 comprises establishing multiple radar jamming mathematical models under different jamming conditions based on different radar jamming echo generation principles; wherein the jamming conditions at least include: suppressive jamming, deceptive jamming and composite jamming; and using the multiple radar jamming mathematical models to generate corresponding multiple two-dimensional time-frequency graphs.

[0016] Exemplarily, the suppressive jamming includes sinusoidal sweep jamming, sawtooth sweep jamming and comb spectrum jamming. The deceptive jamming includes intermittent sampling and forwarding jamming, multi-false target jamming and dense false target jamming. The composite jamming type is the composite jamming of comb spectrum jamming and intermittent sampling and forwarding jamming, and the composite jamming of sinusoidal sweep jamming and multi-false target jamming. Wherein, part of the parameters of the SAR signal and the simulated jamming are shown in Table 1.

[0017] Table 1

[0018] Here, the jamming sample JNR is set in the range of 0dB~30dB. The jamming time-frequency image size is set to 128x128, the saving format is gray image, and the amplitude is between 0~255. When constructing the data set, the training set, the validation set and the test set are generated in the ratio of 7:1:2, and there are 1000 samples of each jamming. Part of the data set sample of the eight radar jamming time-frequency graphs is shown in Figure 2 . Among them, Figure 2 (a) in (a) is a JNR=0dB different type radar jamming time-frequency spectrum sample, Figure 2 (b) in (b) is a JNR=10dB different type radar jamming time-frequency spectrum sample, Figure 2(c) is a time-frequency spectrum sample of different types of radar jamming when JNR=20dB in the above table, Figure 2 (d) is a time-frequency spectrum sample of different types of radar jamming when JNR=30dB in the above table.

[0019] After obtaining the data set, the related modules in steps 120-160 are trained and learned by using the training set, the validation set is used for monitoring in the training process to avoid overfitting, and finally the test set is used for evaluation. Figure 3 is a connection diagram of the related modules in steps 120-160 provided by the embodiment of the present application. Now, the above steps will be described in combination with Figure 3 The following steps are described.

[0020] Step 120: using a channel attention module to perform maximum pooling and average pooling on each two-dimensional time-frequency graph respectively to obtain global information and local information, and based on the global information and the local information, a channel attention matrix is obtained, and each element in the channel attention matrix falls within the open interval (0, 1).

[0021] Here, the channel attention (Channel Attention Module, CAM) module focuses on the importance of the interference time-frequency spectrum or feature map in the channel dimension. Its goal is to assign a weight to each interference feature channel, increase the weight of important interference feature channels, and reduce the weight of secondary interference feature channels, thereby improving the interference feature representation and extraction capability. Therefore, using the channel attention mechanism can help the Mixed Attention and Multiple Loss Constrained Network (MAMLC-Net) to extract channel dimension key recognizable interference features.

[0022] Specifically, Figure 4 is a data processing flowchart of the channel attention module provided by the embodiment of the present application. As Figure 4 shown, step 120 includes performing maximum pooling on the two-dimensional time-frequency graph to obtain a corresponding maximum value vector; performing average pooling on the two-dimensional time-frequency graph to obtain a corresponding average value vector; using a multi-layer perception, respectively, to perform re-weighting processing on the maximum value vector and the average value vector to obtain a processed maximum value vector and a processed average value vector; performing element-wise addition on the processed maximum value vector and the processed average value vector to obtain a channel fusion feature map; using a sigmod activation function, mapping the value of each element in the channel fusion feature map to the open interval (0, 1) to obtain a channel attention matrix.

[0023] Here, the expression of the maximum value vector is ; the expression of the average value vector is . Among them, is an input matrix of two-dimensional time-frequency diagram, is the number of input channels.

[0024] The multi-layer perceptron (MLP) re-weights the maximum value vector and the average value vector using a convolution kernel with a size of 1, dynamically enhances the response of key feature channels, and simultaneously weakens the response of unimportant or noise channels. Subsequently, the fused feature map is mapped into an open interval (0, 1) by a sigmod activation function, and a channel attention weight matrix is obtained, and the mathematical expression of the sigmod function is Finally, the channel attention matrix is represented as . Wherein, is a sigmod function, , is the height dimension of the channel attention feature map, is the width dimension of the channel attention feature map.

[0025] Step 130: Max-pooling and average-pooling the channel attention matrix using the spatial attention module to obtain a spatial information matrix, and performing spatial dimension convolution on the spatial information matrix to obtain a spatial attention weight matrix, each element in the spatial attention weight matrix falls in an open interval (0, 1).

[0026] Here, the spatial attention (SAM) module focuses on important information of the interference time-frequency diagram or the interference feature map in the spatial dimension. Its goal is to assign a weight to each position in the input feature, highlighting important areas and suppressing unimportant areas. The close structure of the spatial attention module after the channel attention module can realize the mutual complementation of feature representation, help the model to capture local and global important information faster, further improve the extraction ability of the network to the interference feature, and perfect the demand of the interference segmentation task for fine feature boundary.

[0027] Figure 5 is a data processing flowchart of the spatial attention module provided by the embodiment of the application. As shown in Figure 5 , step 130 includes: performing max-pooling on the channel attention matrix to obtain a spatial maximum value response vector; performing average-pooling on the channel attention matrix to obtain a spatial average value response vector; splicing the spatial maximum value response vector and the spatial average value response vector to obtain a spatial information matrix; performing spatial dimension convolution on the spatial information matrix to obtain a spatial fused feature map; using a sigmod activation function, mapping the value of each element in the spatial fused feature map to an open interval (0, 1) to obtain a spatial attention weight matrix.

[0028] Here, the expression of the spatial maximum response vector is the same as that of the maximum vector in step 120, and the expression of the spatial average response vector is the same as that of the average vector in step 120. The expression of the spatial attention weight matrix is wherein is the parameter of the convolution layer.

[0029] Step 140: performing feature extraction on the input graph containing interference features by using the interference feature extraction network to obtain an interference time-frequency spectrum with a resolution less than the first threshold; the input graph containing interference features is obtained based on the spatial attention weight matrix.

[0030] Here, please refer to Figure 3 , the input graph containing interference features is obtained by multiplying each two-dimensional time-frequency graph and the corresponding channel attention matrix to obtain a first matrix, and multiplying the spatial attention weight matrix and the first matrix to obtain the input graph containing interference features.

[0031] Here, first, the specific structure of the interference feature extraction network is described, which includes a plurality of independent down-sampling modules; each down-sampling module includes a first channel attention unit, a first spatial attention unit and two layers of convolution units connected in sequence; each convolution unit includes a first 3*3 convolution layer, a first BatchNorm normalization layer and a first ReLU activation layer connected in sequence.

[0032] Here, the number of down-sampling modules is 4. The core component of the convolution module is the convolution kernel, which is essentially a weight matrix, and feature extraction is performed by multiplication and accumulation operation with the corresponding size of the local sub-region in the input interference feature graph. Assuming that the convolution kernel is , the size is ; the input interference feature graph (i.e. the input graph containing interference features) , the size is , and the output feature graph (i.e. the interference time-frequency spectrum) is , the size is . The convolution operation extracts features by sliding the convolution kernel on the input interference feature graph, which maps the local features of the input data to the next layer, and its mathematical expression is: wherein is a two-dimensional discrete convolution operator, is a trainable bias parameter that controls the proportion of dropped information. In addition to the size of the convolution kernel, the stride and padding in the convolution operation are important parameters that control the size of the output feature map. The stride determines the step distance of the convolution kernel sliding on the input feature map, while the padding adjusts the spatial resolution of the interference feature map and preserves the edge information by adding specific values to the edge of the input interference feature map. The adjustment of these two parameters affects the receptive field and computational efficiency of the model. Assuming that the stride and padding in the high direction are and , and the stride and padding in the width direction are and , then the size of the output interference feature map is: , where is the size of the convolution kernel in the height direction, is the size of the convolution kernel in the width direction.

[0033] Convolution calculation is essentially a local weighted sum performed by sliding the convolution kernel over the input data. Each convolution kernel only focuses on a local region, and the same convolution kernel repeatedly uses weights at different positions, significantly reducing the number of parameters, improving computational efficiency, and enhancing the generalization ability of the model. At the same time, by designing different convolution kernels, spatial local patterns in the interference feature image can be effectively recognized, helping the network build semantic representations layer by layer.

[0034] During the training of deep neural networks, as the hierarchical structure progresses, parameter iterative updates will cause input distribution bias to accumulate layer by layer, leading to dynamic instability in the feature space of deep networks. This phenomenon is manifested as follows: small perturbations in bottom-layer parameters are amplified through forward propagation by high-order composite functions, causing severe oscillations in the input distribution of high-layer neurons, and further leading to gradient anomalies and model convergence difficulties. Batch normalization systematically optimizes this through statistical reconstruction strategies, implementing standardization remapping of feature tensors before each activation function, and forcing the output distribution to be a stable state with zero mean and unit variance. This process is achieved through the formula where is the normalized data, is the input data, and are the mean and standard deviation of the mini-batch data, respectively, is a small constant to prevent division by zero. The activation function forms a specific mapping relationship between the input and output of the neural network. Using a linear activation function results in a linear function after transformation, while using a nonlinear activation function can form a nonlinear mapping relationship between the input and output, which can improve the network's ability to learn complex data.

[0035] The algorithm proposed in this invention uses the ReLU activation function to address the problems of vanishing gradients and slow convergence. The output expression of the ReLU function is: .

[0036] Step 150: Use the interference feature recovery network to perform feature recovery on the interference time spectrum map to obtain an interference segmentation map with a resolution greater than the second threshold, where the first threshold is less than the second threshold.

[0037] Here, the interference feature recovery network includes: multiple independent downsampling modules; each downsampling module includes a second channel attention unit, a second spatial attention unit, a 2*2 deconvolution layer and two convolutional units connected in sequence; each convolutional unit includes a second 3*3 convolutional layer, a second BatchNorm normalization layer and a second ReLU activation layer connected in sequence.

[0038] In interference segmentation tasks, the feature extraction process of convolutional neural networks (CNNs) significantly reduces the spatial resolution of the output interference feature map. To reconstruct the spatial dimension that matches the original input interference spectrogram, upsampling techniques are needed to map the low-dimensional interference features to a high-resolution interference feature space. This process essentially involves mathematically expanding the geometric scale of the interference feature map, with the core objective of recovering the spatial details compressed by downsampling. MAMLC-Net chooses deconvolution as its core upsampling mechanism. The advantage of this method lies in its ability to achieve end-to-end feature reconstruction through learnable dynamic parameters. Deconvolution first expands the spatial dimension of the input feature map with zero-padding according to a set dilation rate, then applies a forward convolution kernel for feature recombination, and finally automatically establishes a nonlinear mapping relationship between the input low-dimensional features and the output high-dimensional features through parameter optimization. This upsampling method, with its differentiable parameters, is more adaptable to the interference segmentation needs of complex scenarios compared to traditional fixed interpolation methods. (Deconvolution output size) From input dimensions ,filling kernel size and step length Decision, when When (% represents the remainder), the output size is: ;when At that time, the output size is .

[0039] Step 160: Use a pixel-level interference classification network to identify the interference segmentation map, obtain multiple probability values, and select the maximum value among the multiple probability values ​​as the classification result of the interference segmentation map.

[0040] Here, before performing step 160, the training process of the pixel-level interference classification network is described in detail. Specifically, this SAR active interference segmentation method also includes: S1: Obtain the training dataset; Note: The training dataset here is generated in step 110.

[0041] S2: Train the i-th interference classification network using the training dataset, and calculate the corresponding i-th group loss value using the loss function.

[0042] Here, the expression for the loss function is: ; ; ; ; in, It is the cross-entropy loss function. It is a set similarity loss function. It is the reconstruction error loss function. The value is 0.7. The value is 1e-3. It is a vector composed of the real interference segmentation labels of all pixels. It is a vector composed of the predicted interference segmentation results of all pixels. This represents the number of segmentation categories in the interference segmentation task. It is an interference segmentation of the true label of the image. This is the result of interference segmentation prediction. This represents the total number of pixels in the spectrogram for each interference event. Representing the Real interference segmentation labels for each pixel, Representing the Predicted interference segmentation results for each pixel. To prevent extremely small constants with a denominator of 0.

[0043] Here, the cross-entropy loss function calculates the model's accuracy by measuring the error between the predicted interference segmentation result and the actual interference segmentation label, while the ensemble similarity loss function measures the similarity between the predicted interference segmentation result and the actual interference segmentation label. The reconstruction error loss function obtains the interference segmentation prediction result through a pixel-level interference classification network, and then uses the reconstruction error function to calculate the distance between the predicted interference segmentation result and the input interference segmentation label.

[0044] S3: Determine whether the maximum value in the i-th group of loss values ​​is less than or equal to a preset value. It should be noted that this training iteration process can also stop when the maximum number of iterations is reached.

[0045] S4: If yes, stop training and use the i-th interference classification network as the pixel-level interference classification network.

[0046] S5: If no, update the model parameters of the i-th interference classification network to obtain an i+1-th interference classification network, and re-substitute S2-S4 for training.

[0047] The above is the whole content of the SAR active jamming segmentation method based on the mixed attention and multiple loss constraint network provided by the application. In order to verify the effectiveness of the method, the proposed method is compared with several existing image intelligent segmentation algorithms, and the comparison methods include GCnet, Deeplabv3+ and Segformer.

[0048] Now the indicators for evaluating performance are introduced, including Average Accuracy (AA), Kappa coefficient and IoU coefficient.

[0049] Average Accuracy (AA) is calculated by calculating the average of the segmentation accuracy of each type of jamming image, which represents the overall accuracy in the segmentation task, and describes the consistency of the jamming classification result of each pixel point with the ideal jamming classification type. The mathematical expression is: Wherein, is the number of jamming categories, is the total number of jamming time-frequency image pixels, is the number of correctly segmented pixel points.

[0050] Kappa coefficient calculates the proportion of correct classification and error classification, which can be used to measure the consistency test of classification results. The mathematical expression is: Wherein, represents the number of pixel points with actual classification result as class, represents the number of pixel points with ideal classification result as class.

[0051] IoU coefficient is a geometric similarity measure function, which can be used to calculate the similarity between ideal classification result and actual classification result. The mathematical expression is: .

[0052] Firstly, the effectiveness of the increased channel attention module and the spatial attention module is verified through ablation experiments. The ablation experiment explores the influence of the model performance by selectively excluding the modules in the model, which can reveal the contribution of the modules to the final result and ensure the rationality and reliability of the design. Table 2 below shows the results of each interference segmentation index obtained by the ablation experiment. Compared with the basic framework network UNet, the average accuracy increases by 0.97% after adding the channel attention module; the average accuracy increases by 1.12% after adding the spatial attention, and the rest of the indicators are also optimized to varying degrees. The average accuracy of the network with both modules increased by 2.44%, and all indicators are the optimal results. Therefore, the experiment proves the effectiveness and necessity of adding the two modules to the network.

[0053] Table 2

[0054] To enhance the intuitive comparison results of the ablation experiment, the interference segmentation visualization results of the proposed method and the framework network will be shown below, and the specific comparison results are shown in Figures 6 to 16 The contents in the figure are the time-frequency graphs of 8 types of interference, interference segmentation labels, UNet segmentation visualization results of the framework network, and the proposed algorithm segmentation visualization results. The interference name abbreviations in the figure are consistent with Table 1. Combined with the interference segmentation visualization results of the two algorithms, it can be verified that the interference classification accuracy of the proposed algorithm is higher, the boundary segmentation is clearer, and the identification results of each type of interference in the composite interference are more accurate. The segmentation results of the two methods for each type of interference will be analyzed in detail below.

[0055] Figure 6 is the comparison chart of the existing basic framework network UNet provided by the embodiment of the present application and the comb spectrum interference segmentation visualization result of the present application; Figure 7 is the comparison chart of the existing basic framework network UNet provided by the embodiment of the present application and the dense false target interference segmentation visualization result of the present application; Figure 8 is the comparison chart of the existing basic framework network UNet provided by the embodiment of the present application and the intermittent sampling and forwarding interference segmentation visualization result of the present application; Figure 9 is the comparison chart of the existing basic framework network UNet provided by the embodiment of the present application and the multi-false target interference segmentation visualization result of the present application; Figure 10 is the comparison chart of the existing basic framework network UNet provided by the embodiment of the present application and the sine sweep interference segmentation visualization result of the present application; Figure 11 is the comparison chart of the existing basic framework network UNet provided by the embodiment of the present application and the sawtooth sweep interference segmentation visualization result of the present application; Figure 12is a comparison chart of the existing basic framework network UNet and the composite segmentation visualization result of the comb spectrum interference and intermittent sampling and forwarding interference provided by the embodiment of the present application; Figure 13 is a comparison chart of the existing basic framework network UNet and the composite segmentation visualization result of the sine sweep interference and multi-false target interference provided by the embodiment of the present application. Among them, Figures 6 to 13 The four small graphs in each of,,, and, the first two small graphs (a) represent the interference video graph, (b) represents the segmentation label, and the last two small graphs (c) represent the segmentation result of the existing basic framework network UNet, and (d) represents the segmentation result of the present application (MAMLC-Net). The side column text in (b), (c) and (d) includes “saw sweep”, “positive sweep”, “multi-false”, “intermittent”, “dense false”, “comb” and “target” in turn.

[0056] From Figure 6 It can be seen that the UNet has a more fine spectrum line segmentation result compared with the segmentation label, and the boundary is more blurred. The time-frequency spectrum line of the segmentation result of the present application is closer to the label, the boundary is clearer, and the segmentation effect is better.

[0057] From Figure 7 It can be seen that UNet classifies part of the low-energy dense false target interference harmonics as target echoes, and has insufficient feature extraction capability for low SNR signals, resulting in reduced segmentation accuracy. The present application has good segmentation effect on both high-power interference bodies and low-power interference harmonics.

[0058] From Figure 8 and Figure 9 It can be seen that UNet has multiple cases of confusing the two types of interference when segmenting intermittent sampling and forwarding interference and multi-false target interference, which have similar time-frequency domain characteristics. The segmentation method proposed by the present application is accurate for similar interference, which further proves that the present application can significantly improve the extraction capability of interference features.

[0059] From Figure 10 and Figure 11 It can be seen that UNet is prone to misclassification at the falling edge of the sine sweep interference and the falling edge of the sawtooth sweep interference, and has low classification accuracy for local features of the interference. The method proposed by the present application performs well in global classification and local classification of interference, which proves the effectiveness of the method proposed by the present application.

[0060] From Figure 12 and Figure 13It can be seen that the problem of UNet in the single-class interference segmentation task still exists in the composite interference, and in addition, there is a problem of confusion at the intersection of the two interferences in the composite interference segmentation task. The method provided in the application has good classification effect on each interference in the composite interference segmentation task, and also has obvious improvement on the confusion problem of the composite interference intersection, and realizes accurate segmentation of different interferences.

[0061] Figure 14 is an example diagram of the confusion matrix obtained after the GCnet, Deeplabv3+ and Segformer provided in the embodiment of the application and the method provided in the application perform the interference segmentation task on the test set samples. As shown in Figure 14 , (a) represents the confusion matrix corresponding to GCnet, (b) represents the confusion matrix corresponding to Deeplabv3+, (c) represents the confusion matrix corresponding to Segformer, and (d) represents the confusion matrix corresponding to MAMLC-Net. The confusion matrices of GCnet and Segformer show that the two algorithms are prone to confuse the sinusoidal sweep interference and the multi-false target interference, because the two algorithms are prone to confuse different interference components in the composite interference or classify the intersection part of the two as one category in the segmentation task of the sinusoidal sweep and multi-false target composite interference. GCnet and Deeplabv3+ can better distinguish different types of interferences, but the two algorithms are insufficient in distinguishing between interference and target echo, for example, classifying target echoes as comb spectrum interference, dense false target interference and sinusoidal sweep interference, and classifying part of the false target interference and sweep interference as target echoes. The image processing resolution of the two algorithms is low, and the interference occupies a small proportion in the interference segmentation task, so a high-resolution image segmentation algorithm is needed to accurately segment the interference, and therefore the boundary blur problem is obvious, which is prone to confuse the interference with the target echo. The method provided in the application improves the feature extraction capability of the interference signal through the mixed attention mechanism and enhances the self-calibration capability of the network through the multiple loss constraint, which makes the confusion matrix result of the method provided in the application concentrate on the diagonal line, indicating that the method provided in the application realizes accurate classification of each type of interference. At the same time, the number of interference segmented as target echoes and the number of target echoes classified as interference by the algorithm provided in the application are also obviously reduced compared with other methods, which proves that the method provided in the application also has good segmentation effect on the interference boundary, and improves the boundary blur problem in the interference segmentation.

[0062] Table 3 provided in the embodiment of the application is the evaluation index results corresponding to GCnet, Deeplabv3+ and Segformer and the method provided in the application (the optimal result of each index is in black and bold).

[0063] Table 3

[0064] The results show that the method has advantages in all indicators. In addition to the improvement of feature extraction and self-calibration ability, the encoder-decoder structure of the framework network of the method allows the model to utilize local and global information of the image simultaneously, thereby realizing higher resolution segmentation of the interference time-frequency spectrogram. In the calculation of the segmentation indicators, the AA changes little because the background accounts for a large proportion in the interference segmentation time-frequency spectrogram, while the IoU changes greatly because it represents the set similarity. The results of the evaluation indicators verify that the method has stronger feature extraction and self-calibration ability, and can better solve the problems of boundary ambiguity and interference type confusion.

[0065] To further verify the effectiveness of the method, considering that the interference power will have a great impact on the interference segmentation result, this paper will discuss the performance of each algorithm under different JNR conditions. Figure 15 The figure shows the interference segmentation evaluation indicator results of the method, GCnet, Deeplabv3+, and Segformer under different JNR. As shown in the figure, Figure 15 (a) represents the average accuracy distribution, (b) represents the Kappa coefficient distribution, and (c) represents the IoU coefficient distribution. The smaller the JNR, the more blurred the interference features compared to the target echo, and the worse the evaluation indicators of the segmentation result. When JNR is higher than 10 dB, the interference features are clear, and the indicators change little. Among them, the MAMLC-Net is optimal in all indicators under different JNR, and the segmentation accuracy of the MAMLC-Net can still remain above 98% when JNR=0 dB, which shows that the MAMLC-Net can extract distinguishable interference features from sample data under different conditions, verifying the effectiveness of the method.

[0066] Considering the difficulty of obtaining interference samples and the scarcity of sample quantity in complex real battlefield environments, this application intends to discuss the influence of the sample quantity of the training set on the algorithm performance by reducing the training sample quantity, and analyze the performance of the algorithm through the evaluation indicator results. Figure 16 The figure shows the interference segmentation evaluation indicator results of the method, GCnet, Deeplabv3+, and Segformer under different sample training sets. As shown in the figure, Figure 16As shown, (a) represents the average accuracy distribution, (b) represents the Kappa coefficient distribution, and (c) represents the IoU coefficient distribution. The segmentation accuracy of all algorithms decreases to varying degrees as the number of training set samples decreases. Among them, the indicators of the method are optimal under different sample quantity conditions, and when the number of training samples is very small, the average accuracy of the segmentation result of the method can still remain above 97%, which not only verifies the effectiveness of the data augmentation function of the input module of the method, but also shows that the method has stronger feature extraction capability and better robustness.

[0067] The beneficial effects of the present application include: (1) The SAR active jamming segmentation method based on mixed attention and multiple loss constraint network has the core advantage of significantly improving the jamming segmentation accuracy and robustness. By cooperatively using the channel attention mechanism and the spatial attention mechanism, the method realizes fine extraction of jamming features: the channel attention dynamically enhances the key jamming feature channel and suppresses noise, and the spatial attention accurately locates the significant spatial area of the jamming, so that the model can accurately distinguish the jamming type under the suppression jamming (such as comb spectrum), deception jamming (such as dense false target) and composite jamming scene, effectively solving the similar jamming confusion and boundary ambiguity problem. Ablation experiment verifies that the mixed attention significantly improves the average segmentation accuracy (AA) and significantly optimizes the boundary definition.

[0068] (2) The method effectively overcomes the key technical difficulties in jamming segmentation. The encoder-decoder structure cooperates with the learnable deconvolution upsampling to effectively recover the spatial details of the low-resolution features; the triple loss constraint (cross-entropy loss to ensure global classification correctness, set similarity loss to optimize shape matching, and pixel-level reconstruction error to strengthen detail recovery) cooperatively optimizes the model, significantly improving the boundary alignment accuracy. Even under extreme conditions of low jam-to-noise ratio (JNR=0dB), the model can still maintain a segmentation accuracy of more than 98%, showing strong extraction ability and anti-confusion ability for weak jamming features.

[0069] (3) The channel attention and spatial attention mechanisms are used to strengthen the differences of different types of jamming and target echoes in high-dimensional features, improve the feature extraction ability of the jamming segmentation model for jamming signals, refine the boundary distinction between jamming and target echoes, and introduce a multiple loss function to enhance the network parameter update and jamming feature self-calibration ability, further optimize the performance of the jamming segmentation model, and improve the segmentation accuracy and robustness of the network. The method can solve the problems of fuzzy segmentation boundary and poor updating and calibration ability of existing jamming segmentation methods when facing increasingly complex and variable active jamming.

[0070] (4) In the down-sampling module and the up-sampling module, a channel and spatial attention mechanism module is added to improve the interference feature extraction and recovery ability of the interference segmentation network model, and to make the interference segmentation boundary clear.

[0071] (5) An end-to-end pixel-level prediction is realized by using a full convolution architecture. The core is to recover the semantic features of the deep spatial dimension by using the transpose convolution operation: first, the high-level abstract features are obtained by using the encoder (convolution part) to perform multi-level down-sampling, then the low-resolution feature map output by the final convolution is gradually up-sampled by using the decoder (deconvolution part), and finally the prediction map corresponding to the input image size is reconstructed. This full convolution architecture has double advantages: first, by eliminating the limitation of the full connection layer on the input size, the model can process image input of any resolution; second, the feature map space correspondence mechanism in the encoding and decoding structure effectively preserves the original spatial information, so that each pixel point can obtain semantic class prediction based on the local context information.

[0072] The above is a further detailed description of the present application in combination with specific preferred embodiments, and cannot be regarded as limiting the specific implementation of the present application to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, some simple deductions or substitutions can be made without departing from the concept of the present application, and all of them should be regarded as falling within the protection scope of the present application.

Claims

1. A SAR active jamming segmentation method based on a hybrid attention and multi-loss constraint network, characterized in that, The SAR active jamming segmentation method comprises: a plurality of two-dimensional time-frequency diagrams generated by radar echoes under different jamming conditions are obtained; a channel attention module is used to perform maximum pooling and average pooling on each two-dimensional time-frequency diagram respectively to obtain global information and local information, and a channel attention matrix is obtained based on the global information and the local information, each element in the channel attention matrix falling within an open interval (0, 1); a spatial attention module is used to perform maximum pooling and average pooling on the channel attention matrix to obtain a spatial information matrix, and a spatial attention weight matrix is obtained by performing spatial dimension convolution on the spatial information matrix, each element in the spatial attention weight matrix falling within an open interval (0, 1); a jamming feature extraction network is used to extract features from an input graph containing jamming features to obtain a jamming time-frequency spectrum with a resolution less than a first threshold; the input graph containing jamming features is obtained based on the spatial attention weight matrix; a jamming feature recovery network is used to recover features from the jamming time-frequency spectrum to obtain a jamming segmentation graph with a resolution greater than a second threshold, the first threshold being less than the second threshold; a pixel-level jamming classification network is used to identify the jamming segmentation graph to obtain a plurality of probability values, and the maximum value in the plurality of probability values is selected as the classification result of the jamming segmentation graph.

2. The SAR active jamming segmentation method based on the hybrid attention and multi-loss constraint network according to claim 1, wherein, The input graph containing jamming features is obtained by the following method: the multiplication of each two-dimensional time-frequency diagram and the corresponding channel attention matrix is performed to obtain a first matrix; the multiplication of the spatial attention weight matrix and the first matrix is performed to obtain the input graph containing jamming features.

3. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 1, characterized in that, Before the pixel-level jamming classification network is used to identify the jamming segmentation graph to obtain a plurality of probability values, and the maximum value in the plurality of probability values is selected as the classification result of the jamming segmentation graph, the SAR active jamming segmentation method further comprises: S1: obtaining a training data set; S2: training an i-th jamming classification network using the training data set, and calculating a corresponding i-th set of loss values using a loss function; S3: determining whether the maximum value in the i-th set of loss values is less than or equal to a preset value; S4: if yes, stopping training and taking the i-th jamming classification network as the pixel-level jamming classification network; S5: if no, updating the model parameters of the i-th jamming classification network to obtain an i+1-th jamming classification network, and re-entering S2-S4 for training.

4. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 1, characterized in that, The channel attention module is used to perform maximum pooling and average pooling on each two-dimensional time-frequency diagram to obtain global information and local information, and a channel attention matrix is obtained based on the global information and the local information, comprising: maximum pooling is performed on the two-dimensional time-frequency diagram to obtain a corresponding maximum value vector; average pooling is performed on the two-dimensional time-frequency diagram to obtain a corresponding average value vector; a multi-layer perception is used to perform re-weighting processing on the maximum value vector and the average value vector respectively to obtain a processed maximum value vector and a processed average value vector; Element-wise addition is performed on the processed maximum value vector and the processed average value vector to obtain a channel fusion feature map; A sigmod activation function is used to map the value of each element in the channel fusion feature map to the open interval (0, 1) to obtain the channel attention matrix.

5. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 1, characterized in that, The channel attention matrix is subjected to maximum pooling and average pooling by using the spatial attention module to obtain a spatial information matrix, and the spatial information matrix is subjected to spatial dimension convolution to obtain a spatial attention weight matrix, which includes: The channel attention matrix is subjected to maximum pooling to obtain a spatial maximum value response vector; The channel attention matrix is subjected to average pooling to obtain a spatial average value response vector; The spatial maximum value response vector and the spatial average value response vector are spliced to obtain the spatial information matrix; The spatial information matrix is subjected to spatial dimension convolution to obtain a spatial fusion feature map; A sigmod activation function is used to map the value of each element in the spatial fusion feature map to the open interval (0, 1) to obtain the spatial attention weight matrix.

6. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 1, characterized in that, The interference feature extraction network includes a plurality of independent down-sampling modules; each down-sampling module includes a first channel attention unit, a first spatial attention unit and two convolution units connected in sequence; each convolution unit includes a first 3*3 convolution layer, a first BatchNorm normalization layer and a first ReLU activation layer connected in sequence.

7. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 1, characterized in that, The interference feature recovery network includes a plurality of independent down-sampling modules; each down-sampling module includes a second channel attention unit, a second spatial attention unit, a 2*2 deconvolution layer and two convolution units connected in sequence; each convolution unit includes a second 3*3 convolution layer, a second BatchNorm normalization layer and a second ReLU activation layer connected in sequence.

8. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 1, characterized in that, The method includes: Based on different radar jamming echo generation principles, a plurality of radar jamming mathematical models under different jamming conditions are established; wherein the jamming conditions at least include: suppression jamming, deception jamming and composite jamming; The plurality of radar jamming mathematical models are used to generate corresponding two-dimensional time-frequency graphs.

9. The SAR active jamming segmentation method based on hybrid attention and multi-loss constraint network according to claim 3, characterized in that, The expression of the loss function is: ； ； ； ； in, It is the cross-entropy loss function. It is a set similarity loss function. It is the reconstruction error loss function. The value is 0.

7. The value is 1e-3. It is a vector composed of the real interference segmentation labels of all pixels. It is a vector composed of the predicted interference segmentation results of all pixels. This represents the number of segmentation categories in the interference segmentation task. It is an interference segmentation of the true label of the image. This is the result of interference segmentation prediction. This represents the total number of pixels in the spectrogram for each interference event. Representing the Real interference segmentation labels for each pixel, Representing the Predicted interference segmentation results for each pixel. To prevent extremely small constants with a denominator of 0.

Images