Radar interference suppression and target identification integrated method and system

Abstract

The invention discloses a radar interference suppression and target identification integrated method and system, and the method comprises the steps: building an end-to-end complex valued neural network which is used for carrying out the radar target identification under radio frequency interference; training the network by using a non-interference and interference-containing complex-valued radar training data set, taking non-interference complex-valued radar image data and a target category as supervision tags, and learning anti-interference and target identification at the same time; the network structure comprises an anti-interference and target feature enhancement module and a target identification module. The anti-interference and target feature enhancement module is used for separating a target and radio frequency interference from the input complex-valued radar image, suppressing the radio frequency interference and highlighting the features of the target; and the target identification module is used for learning and extracting the features of the target from the output of the anti-interference and target feature enhancement module. The method can solve the problem that the recognition performance is reduced due to error accumulation in a cascade processing flow in the prior art, and is suitable for radar image target recognition tasks under radio frequency suppression interference.

Description

Technical Field

[0001] This invention relates to the field of radar technology, and specifically to an integrated method and system for radar interference suppression and target identification. Background Technology

[0002] Radar target identification technology has wide applications in remote sensing and traffic control. However, due to the increasing number of radiation sources, radar is likely to be subject to radio frequency interference from communication systems and other radar systems when receiving echoes. These radio frequency interference signals can severely damage the target information in the radar echo, making radar target identification difficult. Therefore, how to efficiently achieve radar target identification under radio frequency interference is an urgent problem to be solved in the field of radar target identification technology.

[0003] Currently, radar target recognition under radio frequency interference mainly adopts a cascaded processing flow: anti-interference and radar target recognition are performed separately. First, an independent anti-interference module suppresses radio frequency interference in the radar image, and then the processed image is input into the target recognition module to complete the recognition. Although the traditional cascaded processing flow has advantages such as low complexity and ease of implementation, and performs well in weak interference scenarios, it has significant drawbacks in complex electromagnetic interference environments. The processing error of the anti-interference module will be transmitted to the subsequent radar target recognition module, leading to error accumulation and ultimately a significant decline in recognition performance. At the same time, current radar target recognition technologies are mostly based on deep learning. However, radar images are essentially complex images, containing both amplitude and phase information. But most existing methods use neural networks with real weights, which only utilize the amplitude information of the radar image and ignore the target features contained in the phase information. This results in insufficient utilization of radar image information, further restricting the improvement of recognition performance. Summary of the Invention

[0004] Purpose of the invention: This invention aims to provide an integrated method and system for radar interference suppression and target recognition. It fully utilizes the amplitude and phase information in radar images while solving the problem of error accumulation in traditional cascaded processing. It realizes an end-to-end network structure that integrates anti-interference and target recognition tasks, thereby improving the performance of radar target recognition tasks under radio frequency interference.

[0005] Technical Solution: To achieve the above-mentioned objectives, the first aspect of this invention provides an integrated method for radar interference suppression and target identification, comprising the following steps:

[0006] Obtain interference-free complex-valued radar echo data, obtain or simulate interference-containing complex-valued radar datasets, retain interference-free complex-valued radar image data as supervision labels during training; segment the complex-valued radar image data to obtain interference-free and interference-containing complex-valued radar target recognition datasets.

[0007] An end-to-end complex-valued neural network is constructed for radar target recognition under radio frequency interference. The network is trained using a training dataset, and interference-free complex-valued radar image data and target categories are used as supervision labels, allowing the network to learn both anti-interference and target recognition simultaneously, resulting in a trained end-to-end complex-valued neural network. The end-to-end complex-valued neural network structure includes an anti-interference and target feature enhancement module and a target recognition module. The anti-interference and target feature enhancement module separates the target from the radio frequency interference from the input complex-valued radar image, suppresses the radio frequency interference, and highlights the target's features. The target recognition module learns and extracts target features from the complex-valued radar image output by the anti-interference and target feature enhancement module, and outputs the target recognition result.

[0008] The complex-valued radar image data to be tested is input into a trained end-to-end complex-valued neural network to obtain the radar target identification results under radio frequency interference.

[0009] Furthermore, the creation of the dataset includes:

[0010] Acquire interference-free complex-valued radar echo data Collect or obtain radio frequency interference data of different intensities and categories through simulation;

[0011] Add radio frequency interference data to the original interference-free radar echo data along the fast time dimension. In the process, radar echo data containing interference was obtained. ;

[0012] The obtained radar echo data containing interference Imaging processing is performed to obtain complex-valued radar image data containing interference. At the same time, the original, interference-free radar echo data Imaging processing is performed to obtain interference-free complex-valued radar image data. ;

[0013] Complex radar image data with and without interference were analyzed separately. and Perform the same segmentation operation to obtain complex-valued radar image data containing a single target, and then perform upsampling or downsampling processing to unify the samples to the same size, resulting in complex-valued radar target recognition datasets with and without interference; record the corresponding target category as a supervision label.

[0014] Furthermore, the anti-interference and target feature enhancement module is implemented through a complex-valued U-Net network structure. By utilizing the encoding and decoding of the complex-valued U-Net network and the skip connection structure, it can suppress radio frequency interference in the input sample and enhance the target features.

[0015] The target recognition module is implemented through a complex-valued CNN network and a complex-valued fully connected layer. The complex-valued CNN network is used to further extract target features, and then the complex-valued fully connected layer, together with the complex-valued modulo operation and the softmax function, completes the task of recognizing radar targets.

[0016] Furthermore, the complex-valued U-Net network includes complex-valued convolutional layers, complex-valued ReLU activation functions, and complex-valued max-pooling layer operations;

[0017] The formula for calculating complex convolutional layers is:

[0018] ;

[0019] in, The output of the complex convolutional layer. For complex-valued convolution kernels, For the input of the complex-valued convolutional layer, * represents a convolution operation for complex values, and * represents a convolution operation for real values. and These are operations that extract the real part and operations that extract the imaginary part, respectively.

[0020] The formula for calculating the complex-valued ReLU activation function is:

[0021] ;

[0022] The formula for calculating the complex-valued max pooling layer is:

[0023] ;

[0024] ;

[0025] in, For complex-valued characteristic matrices In position The value at that location, For complex value modulo operation, For real-valued max-pooling operations, This is the index of the position corresponding to the value obtained in the max pooling operation.

[0026] Furthermore, the complex-valued CNN network, in addition to complex-valued convolutional layers and complex-valued ReLU activation function operations, also includes complex-valued batch normalization layers, complex-valued average pooling layers, and complex-valued fully connected layers;

[0027] The formula for calculating the normalized layer of the complex-valued batch is:

[0028] ;

[0029] in, and These are the learnable scaling and translation transformation parameters, respectively. For inputting a batch of complex data The mean, For inputting a batch of complex data The variance of a complex-valued neural network, The calculation formula is:

[0030] ;

[0031] in, Calculate the covariance;

[0032] The formula for calculating the complex-valued average pooling layer is:

[0033] ;

[0034] in, For real-valued average pooling layer operations, This is the complex-valued feature matrix corresponding to the complex-valued pooling layer;

[0035] The formula for calculating complex-valued fully connected layers is:

[0036] ;

[0037] in, For the output of the complex-valued fully connected layer, Here is the complex-valued weight matrix of the complex-valued fully connected layer. For the input of the complex-valued fully connected layer, This is a complex bias.

[0038] Furthermore, the training method for the end-to-end complex-valued neural network is as follows:

[0039] Interference-laden complex-valued radar image data is used as input, while interference-free complex-valued radar image data is used to supervise the parameter learning of the anti-interference and target feature enhancement module. The target category simultaneously supervises both the anti-interference and target feature enhancement module and the target recognition module, and the network parameters are updated through the backpropagation algorithm.

[0040] Furthermore, the training of the end-to-end complex-valued neural network employs an improved loss function, which consists of a root mean square error loss term and a cross-entropy error loss term, calculated as follows:

[0041] ;

[0042] in, Hyperparameters used to control the proportions of the two terms in the loss function; , These are the functions corresponding to the anti-interference and target feature enhancement module and the target recognition module, respectively. Representative input sample The one-hot encoding of the corresponding category c; Representative complex-valued radar image samples without interference. Represents complex-valued radar images The value of the j-th pixel in the dataset; m, K, and n represent the total number of training samples, the total number of classes in the dataset, and the total number of pixels in a single input sample, respectively; The probability of the sample class output by the network is calculated using the softmax function.

[0043] According to another aspect of the present invention, the present invention also provides an integrated radar interference suppression and target recognition system for implementing the method according to the first aspect, the system comprising:

[0044] The dataset construction module is used to acquire interference-free complex-valued radar echo data, acquire or simulate complex-valued radar datasets with interference, retain interference-free complex-valued radar image data as supervision labels during training, and segment the complex-valued radar image data to obtain interference-free and interference-containing complex-valued radar target recognition datasets.

[0045] A network construction and training module is used to build an end-to-end complex-valued neural network for radar target recognition under radio frequency interference. The network is trained using a training dataset, and interference-free complex-valued radar image data and target categories are used as supervision labels, allowing the network to learn both anti-interference and target recognition simultaneously, resulting in a trained end-to-end complex-valued neural network. The end-to-end complex-valued neural network structure includes an anti-interference and target feature enhancement module and a target recognition module. The anti-interference and target feature enhancement module separates the target from radio frequency interference from the input complex-valued radar image, suppresses radio frequency interference, and highlights the target's features. The target recognition module learns and extracts target features from the complex-valued radar image output by the anti-interference and target feature enhancement module, and outputs the target recognition result.

[0046] The interference suppression and target recognition module is used to input the complex value radar image data to be tested into a trained end-to-end complex value neural network to obtain the target category recognition result.

[0047] According to another aspect of the present invention, the present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the integrated radar interference suppression and target recognition method described in the first aspect.

[0048] According to another aspect of the present invention, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the integrated radar interference suppression and target recognition method described in the first aspect.

[0049] Beneficial Effects: The integrated radar interference suppression and target recognition method and system provided by this invention are applicable to radar target recognition tasks under radio frequency suppression interference scenarios. Compared with existing technologies, its significant advantages are: by using a neural network with complex weights, the amplitude and phase information in the complex radar image can be utilized more fully; by integrating the anti-interference and target recognition modules, end-to-end training avoids error propagation in traditional cascaded processes, improving the stability of radar target recognition under complex electromagnetic interference; and by combining loss functions to constrain the interference suppression and target recognition functions respectively, ensuring that the network has excellent performance in both core tasks. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating an integrated radar interference suppression and target recognition method provided in an embodiment of the present invention.

[0051] Figure 2 This is a diagram showing the overall structure of the end-to-end complex-valued neural network used in this embodiment of the invention.

[0052] Figure 3 This is a structural diagram of the anti-interference and target feature enhancement module based on the complex-valued U-Net network in an embodiment of the present invention;

[0053] Figure 4 This is a structural diagram of the target recognition module based on a complex-valued CNN network in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will now be described more clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] like Figure 1 As shown in the figure, the integrated radar interference suppression and target recognition method disclosed in this embodiment of the invention mainly includes the following steps:

[0056] Step 1: Obtain interference-free complex-valued radar echo data, obtain or simulate interference-containing complex-valued radar datasets, retain interference-free complex-valued radar image data as supervision labels during training; segment the complex-valued radar image data to obtain interference-free and interference-containing complex-valued radar target recognition datasets.

[0057] Step 2: Build an end-to-end complex valued neural network for radar target recognition under radio frequency interference; train the network using the training dataset, and use interference-free complex valued radar image data and target categories as supervision labels to allow the network to learn both anti-interference and target recognition simultaneously, thus obtaining a trained end-to-end complex valued neural network.

[0058] Step 3: Input the complex value radar image data to be tested into the trained end-to-end complex value neural network to obtain the radar target identification results under radio frequency interference.

[0059] In step 1, in this embodiment of the invention, radio frequency interference data is obtained through simulation, and a complex-valued radar image dataset containing interference is created. The specific process of creating the dataset includes:

[0060] Step 1.1: Acquire interference-free complex-valued radar echo data Furthermore, simulations were used to obtain radio frequency interference data of different interference intensities and three types: narrowband interference, frequency modulation broadband interference, and sinusoidal modulation broadband interference, which were uniformly denoted as... ;

[0061] Step 1.2: Add radio frequency interference data to the original radar echo data along the fast time dimension. In the process, radar echo data containing interference was obtained. ;

[0062] Step 1.3: Obtain the radar echo data containing interference. Imaging processing is performed to obtain complex-valued radar image data containing interference. At the same time, the original, interference-free radar echo data Imaging processing is performed to obtain interference-free complex-valued radar image data. ;

[0063] Step 1.4: Process the complex-valued radar image data obtained in Step 1.3, including both interference-containing and interference-free images. and The image is segmented to obtain complex-valued radar image data containing a single target. Each data point in the image contains amplitude and phase information. The images are then upsampled or downsampled to unify the samples to the same size (e.g., 128×128), resulting in complex-valued radar target recognition datasets with and without interference. At the same time, the target category corresponding to each sample with and without interference is recorded as a supervision label.

[0064] In step 2, the end-to-end complex-valued neural network structure in this embodiment of the invention is as follows: Figure 2 As shown, it is divided into an anti-interference and target feature enhancement module and a target recognition module.

[0065] The anti-interference and target feature enhancement module is used to separate the target from the radio frequency interference from the input complex-valued radar image containing radio frequency interference, suppress the radio frequency interference and highlight the target's features, support subsequent target feature learning and recognition, and reduce the impact of radio frequency interference on subsequent recognition tasks.

[0066] The target recognition module is used to further learn and extract target features from the complex radar image after anti-jamming and target feature enhancement, and to complete the recognition based on the features, output the target recognition result, and realize the function of radar target recognition.

[0067] In practical applications, the anti-interference and target feature enhancement module is implemented through a complex-valued U-Net network, and the target recognition module is implemented through a complex-valued CNN network.

[0068] The Complex-Valued U-Net anti-jamming and target feature enhancement module aims to highlight target features and suppress radio frequency interference from the input complex-valued radar image containing radio frequency interference by utilizing the encoding / decoding structure and skip connection structure of the Complex-Valued U-Net network. In the encoding stage, global interference and target features are extracted through downsampling. In the decoding stage, local detail features from the encoding stage are fused through upsampling and skip connections, achieving accurate interference suppression and target feature enhancement, ensuring that the output complex-valued radar image retains complete target feature information. The Complex-Valued U-Net network specifically involves complex-valued convolutional layers, complex-valued ReLU activation functions, and complex-valued max-pooling layers. These operations are explained below:

[0069] The formula for calculating complex convolutional layers is:

[0070] ;

[0071] in The output of the complex convolutional layer. For complex-valued convolution kernels, For the input of the complex-valued convolutional layer, * represents a convolution operation for complex values, and * represents a convolution operation for real values. and These are operations that extract the real part and operations that extract the imaginary part, respectively.

[0072] The formula for calculating the complex-valued ReLU activation function is:

[0073] ;

[0074] The formula for calculating the complex-valued max pooling layer is:

[0075] ;

[0076] ;

[0077] in, For complex-valued characteristic matrices In position The value at that location, For complex value modulo operation, For real-valued max-pooling operations, This is the index of the position corresponding to the value obtained in the max pooling operation.

[0078] The complex-valued CNN target recognition module aims to further extract target features from complex-valued radar images after anti-interference and target feature enhancement, and then complete the recognition. This module first extracts deep target features through a complex-valued CNN network, then maps these features to a class space using a complex-valued fully connected layer. Finally, it calculates the class probability using a complex-valued modulo operation and a softmax function, outputting the final recognition result. The formula for calculating the class probability is shown below:

[0079] ;

[0080] in, For the network output layer The output value of each node, This represents the total number of nodes in the network output layer, and in object recognition tasks, it represents the total number of object categories in the dataset.

[0081] Complex-valued CNN networks include complex-valued convolutional layers, complex-valued ReLU activation functions, complex-valued batch normalization layers, complex-valued average pooling layers, and complex-valued fully connected layers. The following explains the complex-valued batch normalization layers, complex-valued average pooling layers, and complex-valued fully connected layers:

[0082] The formula for calculating the normalized layer of the complex-valued batch is:

[0083] ;

[0084] in, and These are the learnable scaling and translation transformation parameters, respectively. For inputting a batch of complex data The mean, For inputting a batch of complex data The variance of a complex-valued neural network, The calculation formula is:

[0085] ;

[0086] in, Calculate the covariance;

[0087] The formula for calculating the complex-valued average pooling layer is:

[0088] ;

[0089] in For real-valued average pooling layer operations, This is the complex-valued feature matrix corresponding to the complex-valued pooling layer.

[0090] The calculation formula for complex-valued fully connected layers is:

[0091] ;

[0092] in, For the output of the complex-valued fully connected layer, Here is the complex-valued weight matrix of the complex-valued fully connected layer. For the input of the complex-valued fully connected layer, This is a complex bias.

[0093] In step 2, the network training method is as follows:

[0094] Interference-laden complex-valued radar image data is used as input, while interference-free complex-valued radar image data is used to supervise the parameter learning of the anti-jamming and target feature enhancement module. The target category simultaneously supervises both the anti-jamming and target feature enhancement module and the target recognition module. Finally, the network parameters are updated through the backpropagation algorithm.

[0095] In network training, an improved loss function is used to guide the learning of network parameters. This loss function consists of two parts: the root mean square error loss term and the cross-entropy error loss term. This reflects the degree of difference between the interference-free complex-valued radar image and the interference-suppressed complex-valued radar image. By calculating the root mean square error between the output of the anti-jamming and target feature enhancement module and the interference-free data sample, the module is guided to learn the functions of interference suppression and target feature enhancement; the cross-entropy error loss term... This reflects the difference between the network's predicted category and the actual target category. By calculating the cross-entropy error loss between the network output and the actual category, the entire network is guided to learn target feature recognition.

[0096] The specific formula for calculating the loss function is shown below:

[0097] ;

[0098] in, Hyperparameters used to control the proportions of the two terms in the loss function; , These are the functions corresponding to the anti-interference and target feature enhancement module and the target recognition module, respectively. Representative input sample The one-hot encoding of the corresponding category c; This represents a sample of radar images without interference. Represents complex-valued radar images The value of the j-th pixel in the dataset; m, K, and n represent the total number of training samples, the total number of classes in the dataset, and the total number of pixels in a single input sample, respectively; The probability of the sample class output by the network is calculated using the softmax function.

[0099] For example, in step 2, the more specific steps in this embodiment of the invention are as follows:

[0100] Step 2.1: With the input data feature dimension being 1×128×128, an end-to-end complex-valued neural network is constructed. The complex-valued U-Net network for the anti-interference and target feature enhancement module uses a 3-layer U-Net structure. The encoder's first layer consists of two 3×3 complex-valued convolutional layers with padding of 1. Each convolutional layer is followed by a complex-valued ReLU activation function, and the kernel has 16 channels. Complex-valued max-pooling layers are used to reduce the feature dimension to 16×64×64. The second layer consists of two 3×3 complex-valued convolutional layers with padding of 1. Each convolutional layer is followed by a complex-valued ReLU activation function. The first layer consists of two 3×3 complex convolutional layers with 32 channels each. A complex max-pooling layer is used to reduce the feature dimension to 32×32×32. The third layer comprises two 3×3 complex convolutional layers with 1 padding. Each convolutional layer is followed by a complex ReLU activation function, and each convolutional layer has 64 channels. A complex max-pooling layer is used to reduce the feature dimension to 64×16×16. The middle layer contains two 3×3 complex convolutional layers with 1 padding. Each convolutional layer is followed by a complex ReLU activation function. The first layer of the decoder uses a complex deconvolution operation to upsample the features back to 64×32×32, and then merges them with the encoder. The output of the encoder's third layer is concatenated along the channel dimension and then input into two 3×3 complex-valued convolutional layers with padding of 1. Each convolutional layer is followed by a complex-valued ReLU activation function. The kernels of each layer have 64 channels, resulting in an output feature dimension of 64×32×32. The second layer performs upsampling through complex-valued deconvolution to restore the features to 64×64×64. This feature is then concatenated along the channel dimension with the output of the encoder's second layer and input into two 3×3 complex-valued convolutional layers with padding of 1. Each convolutional layer is followed by a complex-valued ReLU activation function. The kernels of each layer have 32 channels, resulting in an output feature dimension of 32×64×64. 4; The third layer upsamples the features to 32×128×128 using complex-valued deconvolution, then concatenates it with the output of the first layer of the encoder along the channel dimension before inputting it into two 3×3 complex-valued convolutional layers with padding of 1. Each convolutional layer is followed by a complex-valued ReLU activation function, with each kernel having 16 channels. The output feature dimension is 16×128×128. Finally, it passes through a 1×1 complex-valued convolutional layer with 1 channel and 0 padding to ensure that the output dimension matches the input dimension, which is 1×128×128. The specific structure diagram of the complex-valued U-Net network is shown below. Figure 3As shown, the target recognition module consists of a 4-layer complex-valued CNN network and a 1-layer complex-valued fully connected layer. The kernel sizes of the four complex-valued convolutional layers are 5×5, 7×7, 5×5, and 5×5, with 16, 32, 64, and 128 channels respectively, and padding of 1 for each. Each convolutional layer is followed by a complex-valued ReLU activation function and a complex-valued batch normalization layer. The first three complex-valued ReLU functions are followed by a complex-valued average pooling layer with a window size of 2×2. The dimension of the complex-valued fully connected layer is 8192×num_class, where num_class represents the total target categories. The data input to the target recognition module becomes 128×8×8 after passing through the four complex-valued convolutional layers, and then flattens to 1×8192 to fit the input of the complex-valued fully connected layer. Finally, the complex-valued fully connected layer outputs data of dimension 1×num_class, which is the recognition result. The specific structure of the complex-valued CNN network is shown below. Figure 4 As shown; finally, the output of the complex-valued fully connected layer is subjected to complex-valued modulo operation and softmax function to output the class probability, and the class with the highest probability is the recognition result of the network;

[0101] Step 2.2: Input the training set constructed in Step 1 into the network, using the interference-free complex-valued radar image as the root mean square error loss term. The supervision labels are used, with the corresponding target category as the cross-entropy error loss term. The network is trained using supervisory labels;

[0102] Step 2.3: Optimize the loss function and update the network parameters using the backpropagation algorithm to complete the end-to-end training of the network;

[0103] Step 2.4: Adjust the hyperparameters in the loss function Repeat steps 2.2 and 2.3 to select the network parameters with the best training results as the final training result.

[0104] To compare the end-to-end method of this invention with the traditional cascaded method, the embodiments of this invention perform step-by-step training of the network, that is, using the root mean square error loss function. The adversarial interference and target feature enhancement module uses a complex-valued U-Net network for supervised training, employing the cross-entropy error loss function. The complex-valued CNN network for target recognition is trained under supervision, and then the two trained networks are cascaded to obtain a complex-valued network for radar target recognition under radio frequency interference. To compare the complex-valued neural network used in this invention with the traditional real-valued neural network, the network weights of the method described in this invention are replaced with real-valued weights for comparison. To verify the performance and effectiveness of the improved loss function used in this invention, the network is trained using the unmodified loss function, namely the cross-entropy error loss function, for comparison with the method of this invention.

[0105] Tables 1, 2, and 3 show the experimental results of the method implemented in this invention compared with the traditional cascaded method, the real-valued weighted neural network method, and the network training method without improved loss function. Among them, the three datasets ①, ②, and ③ are datasets with different average interference intensities constructed according to step 1, with average interference-to-signal ratios of 12.6dB, 18.7dB, and 22.2dB, respectively. The interference type is one or more of the three types mentioned.

[0106] The results in Table 1 show that the end-to-end method of this invention outperforms the traditional cascaded method, verifying that the end-to-end training method can alleviate the problem of performance degradation of the back-end recognition task caused by error propagation in the cascaded method.

[0107] The results in Table 2 show that the complex-valued neural network of the present invention outperforms the real-valued neural network when processing complex-valued radar images, verifying that the complex-valued network can make fuller use of the information in complex-valued radar images with phase information.

[0108] The results in Table 3 show that the improved loss function of the present invention enables the network to learn both anti-jamming and target recognition tasks more effectively, while the traditional target recognition training method without improved loss function is difficult to achieve good performance on complex radar images under interference.

[0109] Table 1. Comparison of end-to-end and cascaded methods

[0110]

[0111] Table 2 Comparison results of complex-valued networks and real-valued networks

[0112]

[0113] Table 3 Comparison results before and after loss function improvement

[0114]

[0115] This invention also discloses an integrated radar interference suppression and target recognition system, used to implement the integrated radar interference suppression and target recognition method of the aforementioned embodiments. The system includes: a dataset construction module, used to acquire interference-free complex-valued radar echo data, acquire or simulate interference-containing complex-valued radar datasets, and retain interference-free complex-valued radar image data as supervision labels during training; segmenting the complex-valued radar image data to obtain interference-free and interference-containing complex-valued radar target recognition datasets; and a network construction and training module, used to build an end-to-end complex-valued neural network for radar target recognition under radio frequency interference; training the network using the training dataset, and using interference-free complex-valued radar image data and target classes. The network is trained by simultaneously learning anti-interference and target recognition, without using a supervisory label. The end-to-end complex-valued neural network structure includes an anti-interference and target feature enhancement module and a target recognition module. The anti-interference and target feature enhancement module separates the target from radio frequency interference from the input complex-valued radar image, suppresses radio frequency interference, and highlights the target's features. The target recognition module learns and extracts target features from the complex-valued radar image output by the anti-interference and target feature enhancement module, and outputs the target recognition result. The interference suppression and target recognition module inputs the complex-valued radar image data to be tested into the trained end-to-end complex-valued neural network to obtain the target category recognition result.

[0116] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the radar interference suppression and target recognition integrated method of the foregoing embodiments.

[0117] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the radar interference suppression and target recognition integrated method of the foregoing embodiments.

[0118] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for integrating radar interference suppression and target recognition, characterized in that, Includes the following steps: Obtain interference-free complex-valued radar echo data, obtain or simulate interference-containing complex-valued radar datasets, retain interference-free complex-valued radar image data as supervision labels during training; segment the complex-valued radar image data to obtain interference-free and interference-containing complex-valued radar target recognition datasets. An end-to-end complex-valued neural network is constructed for radar target recognition under radio frequency interference. The network is trained using a training dataset, and interference-free complex-valued radar image data and target categories are used as supervision labels, allowing the network to learn both anti-interference and target recognition simultaneously, resulting in a trained end-to-end complex-valued neural network. The end-to-end complex-valued neural network structure includes an anti-interference and target feature enhancement module and a target recognition module. The anti-interference and target feature enhancement module separates the target from the radio frequency interference from the input complex-valued radar image, suppresses the radio frequency interference, and highlights the target's features. The target recognition module learns and extracts target features from the complex-valued radar image output by the anti-interference and target feature enhancement module, and outputs the target recognition result. The complex-valued radar image data to be tested is input into a trained end-to-end complex-valued neural network to obtain the radar target identification results under radio frequency interference.

2. The integrated radar interference suppression and target identification method according to claim 1, characterized in that, The creation of the dataset includes: Acquire interference-free complex-valued radar echo data Collect or obtain radio frequency interference data of different intensities and categories through simulation; Add radio frequency interference data to the original interference-free radar echo data along the fast time dimension. In the process, radar echo data containing interference was obtained. ; The obtained radar echo data containing interference Imaging processing is performed to obtain complex-valued radar image data containing interference. At the same time, the original, interference-free radar echo data Imaging processing is performed to obtain interference-free complex-valued radar image data. ; Complex radar image data with and without interference were analyzed separately. and Perform the same segmentation operation to obtain complex-valued radar image data containing a single target, and then perform upsampling or downsampling processing to unify the samples to the same size, resulting in complex-valued radar target recognition datasets with and without interference; record the corresponding target category as a supervision label.

3. The integrated radar interference suppression and target recognition method according to claim 1, characterized in that: The anti-interference and target feature enhancement module is implemented through a complex-valued U-Net network structure. By utilizing the encoding and decoding of the complex-valued U-Net network and the skip connection structure, it can suppress radio frequency interference in the input sample and enhance the target features. The target recognition module is implemented through a complex-valued CNN network and a complex-valued fully connected layer. The complex-valued CNN network is used to further extract target features, and then the complex-valued fully connected layer, together with the complex-valued modulo operation and the softmax function, completes the task of recognizing radar targets.

4. The integrated radar interference suppression and target identification method according to claim 3, characterized in that, The complex-valued U-Net network includes complex-valued convolutional layers, complex-valued ReLU activation functions, and complex-valued max-pooling layer operations. The formula for calculating complex convolutional layers is: ； in, The output of the complex convolutional layer. For complex-valued convolution kernels, For the input of the complex-valued convolutional layer, * represents a convolution operation for complex values, and * represents a convolution operation for real values. and These are operations that extract the real part and operations that extract the imaginary part, respectively. The formula for calculating the complex-valued ReLU activation function is: ； The formula for calculating the complex-valued max pooling layer is: ； ； in, For complex-valued characteristic matrices In position The value at that location, For complex value modulo operation, For real-valued max-pooling operations, This is the index of the position corresponding to the value obtained in the max pooling operation.

5. The integrated radar interference suppression and target recognition method according to claim 4, characterized in that, The complex-valued CNN network, in addition to complex-valued convolutional layers and complex-valued ReLU activation function operations, also includes complex-valued batch normalization layers, complex-valued average pooling layers, and complex-valued fully connected layers. The formula for calculating the normalized layer of the complex-valued batch is: ； in, and These are the learnable scaling and translation transformation parameters, respectively. For inputting a batch of complex data The mean, For inputting a batch of complex data The variance of a complex-valued neural network, The calculation formula is: ； in, Calculate the covariance; The formula for calculating the complex-valued average pooling layer is: ； in, For real-valued average pooling layer operations, This is the complex-valued feature matrix corresponding to the complex-valued pooling layer; The formula for calculating complex-valued fully connected layers is: ； in, For the output of the complex-valued fully connected layer, Here is the complex-valued weight matrix of the complex-valued fully connected layer. For the input of the complex-valued fully connected layer, This is a complex bias.

6. The integrated radar interference suppression and target identification method according to claim 1, characterized in that, The training method for the end-to-end complex-valued neural network is as follows: Interference-laden complex-valued radar image data is used as input, while interference-free complex-valued radar image data is used to supervise the parameter learning of the anti-interference and target feature enhancement module. The target category simultaneously supervises both the anti-interference and target feature enhancement module and the target recognition module, and the network parameters are updated through the backpropagation algorithm.

7. The integrated radar interference suppression and target recognition method according to claim 6, characterized in that, The end-to-end complex-valued neural network is trained using an improved loss function, which consists of a root mean square error loss term and a cross-entropy error loss term, and is calculated as follows: ； in, Hyperparameters used to control the proportions of the two terms in the loss function; , These are the functions corresponding to the anti-interference and target feature enhancement module and the target recognition module, respectively. Representative input sample The one-hot encoding of the corresponding category c; Representative complex-valued radar image samples without interference. Represents complex-valued radar images The value of the j-th pixel in the dataset; m, K, and n represent the total number of training samples, the total number of classes in the dataset, and the total number of pixels in a single input sample, respectively; The probability of the sample class output by the network is calculated using the softmax function.

8. An integrated radar interference suppression and target recognition system, used to implement the method according to any one of claims 1-7, characterized in that, include: The dataset construction module is used to acquire interference-free complex-valued radar echo data, acquire or simulate complex-valued radar datasets with interference, retain interference-free complex-valued radar image data as supervision labels during training, and segment the complex-valued radar image data to obtain interference-free and interference-containing complex-valued radar target recognition datasets. A network construction and training module is used to build an end-to-end complex-valued neural network for radar target recognition under radio frequency interference. The network is trained using a training dataset, and interference-free complex-valued radar image data and target categories are used as supervision labels, allowing the network to learn both anti-interference and target recognition simultaneously, resulting in a trained end-to-end complex-valued neural network. The end-to-end complex-valued neural network structure includes an anti-interference and target feature enhancement module and a target recognition module. The anti-interference and target feature enhancement module separates the target from radio frequency interference from the input complex-valued radar image, suppresses radio frequency interference, and highlights the target's features. The target recognition module learns and extracts target features from the complex-valued radar image output by the anti-interference and target feature enhancement module, and outputs the target recognition result. The interference suppression and target recognition module is used to input the complex value radar image data to be tested into a trained end-to-end complex value neural network to obtain the target category recognition result.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the integrated radar interference suppression and target recognition method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the integrated radar interference suppression and target recognition method according to any one of claims 1-7.

Images