Radar robust target recognition method based on channel feature pyramid model
By extracting multi-scale features using a channel feature pyramid model, the problem of performance degradation caused by repetition frequency changes in radar target recognition is solved, and robust target recognition under different repetition frequency conditions is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2023-03-29
- Publication Date
- 2026-06-12
AI Technical Summary
Existing radar target recognition methods exhibit decreased performance when pulse repetition frequency changes, and traditional methods suffer from severe spectral aliasing at low repetition frequencies, resulting in poor classification robustness.
We employ a channel feature pyramid model, extracting multi-scale features through an average pooling module and a channel pyramid component. We also combine a global average pooling layer to address the issue of varying input data length, thus constructing a robust recognition model.
It maintains good recognition performance and robustness when the radar repetition rate changes, improving the recognition accuracy, and performs exceptionally well under low repetition rate conditions.
Smart Images
Figure CN116699603B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar technology, specifically relating to a robust radar target recognition method based on a channel feature pyramid model. Background Technology
[0002] Radar target recognition refers to determining the type of a target using its radar echo signal. Ideally, to satisfy the Nyquist sampling theorem, the pulse repetition frequency of the radar complex signal should be greater than twice the bandwidth to avoid aliasing. Furthermore, the higher the pulse repetition frequency, the richer the information contained in the echo, and the better the recognition performance. However, in reality, to address issues such as range ambiguity, the pulse repetition frequency may change. If the trained recognition model does not include this repetition frequency information, a mismatch will occur during the testing phase, leading to a decrease in recognition performance. To address this problem, existing recognition methods mainly train different templates for training samples at different repetition frequencies during the training phase, and then match the template with the same repetition frequency as the current test sample during the testing phase. However, training a recognition model at new repetition frequencies is also limited by factors such as data scale and training cost. Therefore, achieving robust recognition when the pulse repetition frequency changes is an urgent problem to be solved.
[0003] Xu Yijian et al. proposed a method based on spatial pyramid pooling networks to address radar target recognition problems under pulse repetition frequency mismatch during the training and testing phases. This method replaces traditional pooling layers with spatial pyramid pooling layers, allowing the network to extract features of different scales from radar echoes and fix the output dimension. This solves the problem of input data length variations caused by pulse repetition frequency changes, significantly improving radar target recognition performance in scenarios with varying pulse repetition frequencies. However, spatial pyramid pooling layers only use pooling layers of different sizes to process data, resulting in inferior feature extraction compared to convolution. When the repetition frequency of test samples is low, spectral aliasing is severe, leading to low target recognition accuracy and poor classification robustness, failing to achieve the expected results. Summary of the Invention
[0004] To address the aforementioned problems in the existing technology, this invention provides a robust radar target recognition method based on a channel feature pyramid model. The technical problem to be solved by this invention is achieved through the following technical solution:
[0005] This invention provides a robust radar target recognition method based on a channel feature pyramid model. The method utilizes a trained channel feature pyramid model to identify targets in the measured echo data, and includes the following steps:
[0006] Acquire the echo data to be tested;
[0007] The echo data to be tested is input into the first average pooling module, the second average pooling module, the third average pooling module and the first convolution module to obtain the first sampling feature, the second sampling feature, the third sampling feature and the convolution feature;
[0008] The convolutional features are concatenated with the first sampling features, and the first concatenated features are input into the first downsampling module and the first channel pyramid part that are connected in sequence to obtain the first downsampling features and the first channel pyramid output.
[0009] The first downsampling feature, the first channel pyramid output, and the second sampling feature are concatenated, and the second concatenated feature is input into the second downsampling module and the second channel pyramid part that are connected in sequence to obtain the second downsampling feature and the second channel pyramid output.
[0010] The second downsampling feature, the second channel pyramid output, and the third sampling feature are concatenated, and the third concatenated feature is input into the second convolution module and the recognition and classification module that are connected in sequence to obtain the recognition and classification result.
[0011] In one embodiment of the present invention, the first average pooling module includes an average pooling layer;
[0012] The second average pooling module includes two consecutive average pooling layers;
[0013] The third average pooling module includes three consecutive average pooling layers.
[0014] In one embodiment of the present invention, the output length of the average pooling layer is:
[0015]
[0016] Among them, L in is the length of the input data, p is the padding size, k is the kernel length, and s is the stride size.
[0017] In one embodiment of the present invention, both the first downsampling module and the second downsampling module include a first convolutional layer and a max-pooling layer connected in sequence, wherein,
[0018] When the number of input channels of the first downsampling module or the second downsampling module is greater than the number of output channels, the number of output channels of the first convolutional layer is the same as the number of output channels of the first downsampling module or the second downsampling module.
[0019] When the number of input channels of the first downsampling module or the second downsampling module is less than the number of output channels, the number of output channels of the first convolutional layer is the difference between the number of output channels and the number of input channels of the first downsampling module or the second downsampling module. The output result of the first convolutional layer is then concatenated with the output result of the max pooling layer to obtain the output of the first downsampling module or the second downsampling module.
[0020] In one embodiment of the present invention, the first channel pyramid portion includes two consecutive channel pyramid modules;
[0021] The second channel pyramid section comprises six consecutive channel pyramid modules.
[0022] In one embodiment of the present invention, the channel pyramid module includes a second convolutional layer, a first feature pyramid channel, a second feature pyramid channel, a third feature pyramid channel, a fourth feature pyramid channel, and a third convolutional layer, wherein,
[0023] The output of the second convolutional layer is simultaneously input into the first feature pyramid channel, the second feature pyramid channel, the third feature pyramid channel, and the fourth feature pyramid channel;
[0024] The output data of the first feature pyramid channel is added to the output data of the second feature pyramid channel to obtain a first addition result. The first addition result is added to the output data of the third feature pyramid channel to obtain a second addition result. The second addition result is added to the output data of the fourth feature pyramid channel to obtain a third addition result. The concatenation result of the first feature pyramid channel output data, the first addition result, the second addition result and the third addition result is used as the input data of the third convolutional layer.
[0025] The output data of the third convolutional layer is added to the input data of the first convolutional layer to obtain the output data of the channel pyramid module.
[0026] In one embodiment of the present invention, the first feature pyramid channel, the second feature pyramid channel, the third feature pyramid channel, and the fourth feature pyramid channel each include three sequentially connected dilated convolutional layers.
[0027] Specifically, the output results of the first dilated convolutional layer, the second dilated convolutional layer, and the third dilated convolutional layer are concatenated to obtain the output data of the first feature pyramid channel, the second feature pyramid channel, the third feature pyramid channel, or the fourth feature pyramid channel.
[0028] In one embodiment of the present invention, the dilation rates of the dilated convolutions in the first, second, third, and fourth feature pyramid channels are 1, 1, and 1, respectively. And d+1, where d is the expansion factor of the corresponding channel pyramid module.
[0029] In one embodiment of the present invention, the identification and classification module includes a global average pooling layer and a SoftMax layer connected in sequence.
[0030] In one embodiment of the present invention, the training method of the trained channel feature pyramid model includes the following steps:
[0031] Generate a training dataset, which includes a training echo dataset and a category label for each training sample. The training echo dataset includes echo signals of several categories with a single pulse repetition frequency.
[0032] The training echo dataset is input into the channel feature pyramid model, and the predicted identification label is output.
[0033] Using the cross-entropy loss function, the cross-entropy value between the predicted identification label and the category label corresponding to the training sample is calculated. Then, the network parameters are iteratively updated using the backpropagation algorithm until the cross-entropy loss function converges, resulting in the trained channel feature pyramid model.
[0034] The cross-entropy loss function is:
[0035]
[0036] Where H represents the cross-entropy loss function, Q represents the predicted class label of the channel feature pyramid network, P represents the true class label of the training samples in the training dataset, x = 1, 2, ... X, x represents the class index of the training samples in the training dataset, X represents the total number of classes of the training samples in the training dataset, log represents the logarithm operation to the base 10, and N represents the batch size in a forward computation and backpropagation process.
[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0038] 1. In the target recognition method of the present invention, the channel feature pyramid model is used to recognize the target in the echo data to be measured. The echo data to be measured is processed by the average pooling module for feature extraction, and then input into the downsampling module and the channel pyramid part for feature extraction in sequence. In the channel pyramid part, different receptive fields are obtained when the convolution kernel size is constant, and multi-scale features are extracted through parallel structure. It still has good recognition effect and robust recognition performance when the radar repetition frequency changes.
[0039] 2. In the target recognition method of the present invention, the global average pooling layer is used in the recognition and classification module instead of the traditional pooling layer, which solves the problem of different input data lengths caused by the repetition frequency variation. Attached Figure Description
[0040] Figure 1 A flowchart illustrating a robust radar target recognition method based on a channel feature pyramid model, provided for an embodiment of the present invention;
[0041] Figure 2 A structural block diagram of a channel feature pyramid model provided in an embodiment of the present invention;
[0042] Figure 3 This is a schematic diagram of the structure of a channel pyramid module provided in an embodiment of the present invention;
[0043] Figure 4 This is a schematic diagram of a characteristic pyramid channel provided in an embodiment of the present invention;
[0044] Figure 5 The simulation experiment results are shown in the figure provided for the embodiments of the present invention. Detailed Implementation
[0045] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0046] Example 1
[0047] Please see Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating a robust radar target recognition method based on a channel feature pyramid model, provided by an embodiment of the present invention. Figure 2 This is a structural block diagram of a channel feature pyramid model provided in an embodiment of the present invention.
[0048] like Figure 2 As shown, the channel feature pyramid model includes a first average pooling module, a second average pooling module, a third average pooling module, a first convolution module, a first downsampling module, a first channel pyramid part, a second downsampling module, a second channel pyramid part, a second convolution module, and a recognition and classification module. Specifically, the first, second, and third average pooling modules are used to downsample the input data; the first and second downsampling modules are used to downsample the input data; the first and second channel pyramid parts utilize the characteristics of the Inception model and the dilated convolution model to extract features at different scales from the input module data; and the recognition and classification module predicts and recognizes the multi-scale features of the input and outputs the recognition results.
[0049] This embodiment presents a robust radar target recognition method based on a channel feature pyramid model. The method utilizes a trained channel feature pyramid model to identify targets in the measured echo data. Specifically, the training method for the trained channel feature pyramid model includes the following steps:
[0050] S1. Generate a training dataset, which includes a training echo dataset and a category label corresponding to each training sample. The training echo dataset includes echo signals of several categories with a single pulse repetition frequency.
[0051] Specifically, firstly, echo signals with single-pulse repetition frequencies belonging to X categories are obtained as a training echo dataset. Each category contains Z echo signals, where X ≥ 3 and Z ≥ 1000. Then, a corresponding category label is added to each training sample.
[0052] S2. The channel feature pyramid model is trained using the training echo dataset and the category labels corresponding to the training samples to obtain the trained channel feature pyramid model. Specifically, this includes:
[0053] S21: Input the training echo dataset into the channel feature pyramid model and output the predicted recognition label.
[0054] S22. Using the cross-entropy loss function, calculate the cross-entropy value between the predicted identification label and the category label corresponding to the training sample. Then, use the backpropagation algorithm to iteratively update the network parameters until the cross-entropy loss function converges, thus obtaining the trained channel feature pyramid model. The cross-entropy loss function is:
[0055]
[0056] Where H represents the cross-entropy loss function, Q represents the predicted class label of the channel feature pyramid network, P represents the true class label of the training samples in the training dataset, x = 1, 2, ... X, x represents the class index of the training samples in the training dataset, X represents the total number of classes of the training samples in the training dataset, log represents the logarithm operation to the base 10, and N represents the batch size in a forward computation and backpropagation process.
[0057] The radar robust target recognition method based on the channel feature pyramid model specifically includes the following steps:
[0058] S1. Obtain the echo data to be measured from the radar system.
[0059] S2. Input the echo data to be tested into the first average pooling module, the second average pooling module, the third average pooling module and the first convolution module to obtain the first sampling feature, the second sampling feature, the third sampling feature and the convolution feature.
[0060] In this embodiment, the first average pooling module, the second average pooling module, and the third average pooling module downsample the input data through average pooling.
[0061] Specifically, the first average pooling module includes one average pooling layer; the second average pooling module includes two consecutive average pooling layers; and the third average pooling module includes three consecutive average pooling layers.
[0062] In one specific embodiment, in the first average pooling module, the pooling kernel size of the average pooling layer is set to 1×3, and the pooling step size is set to 1×2. In the second average pooling module, the pooling kernel size of both average pooling layers is set to 1×3, and the pooling step size of both is set to 1×2. In the third average pooling module, the pooling kernel size of all three average pooling layers is set to 1×3, and the pooling step size of each is set to 1×2.
[0063] Furthermore, the output length of each average pooling layer is:
[0064]
[0065] Among them, L in is the length of the input data, p is the padding size, k is the kernel length, and s is the stride size.
[0066] Specifically, the first convolutional module includes three convolutional layers, which are connected sequentially, as follows: Figure 2 As shown. The number of output channels C of the first convolutional layer. out The number of output channels in the second convolutional layer is 8, the kernel size is 1×3, and the stride is 1×2; out The number of output channels in the third convolutional layer is 16, the kernel size is 1×3, and the stride is 1×1; out The value is 32, the kernel size is 1×3, and the stride is 1×1.
[0067] Specifically, the echo data to be tested is input into the first average pooling module, the second average pooling module, the third average pooling module, and the first convolution module. The first average pooling module outputs the first sampling feature, the second average pooling module outputs the second sampling feature, the third average pooling module outputs the third sampling feature, and the first convolution module outputs the convolution feature.
[0068] S3. The convolutional features are concatenated with the first sampling features, and the first concatenated features are input into the first downsampling module and the first channel pyramid part that are connected in sequence to obtain the first downsampling features and the first channel pyramid output.
[0069] Specifically, the convolutional features are concatenated with the first sampling features to obtain the first concatenated features; the first concatenated features are input into the first downsampling module and the first channel pyramid part connected in sequence, the first downsampling module outputs the first downsampling features, and the first channel pyramid part obtains the first channel pyramid output.
[0070] S4. The first downsampling feature, the first channel pyramid output, and the second sampling feature are concatenated, and the second concatenated feature is input into the second downsampling module and the second channel pyramid part that are connected in sequence to obtain the second downsampling feature and the second channel pyramid output.
[0071] Specifically, the first downsampling feature, the first channel pyramid output, and the second sampling feature are concatenated to obtain the second concatenated feature. The second concatenated feature is then input into the second downsampling module and the second channel pyramid part, which are connected in sequence. The second downsampling module outputs the second downsampling feature, and the first channel pyramid part outputs the second channel pyramid output.
[0072] Specifically, both the first downsampling module and the second downsampling module include a first convolutional layer and a max pooling layer connected in sequence. When the number of input channels of the first downsampling module or the second downsampling module is greater than the number of output channels, only the first convolutional layer is used to downsample the input data, and the output of the first convolutional layer is used as the output of the first downsampling module or the second downsampling module. In this case, the number of output channels of the first convolutional layer is the same as the number of output channels of the first downsampling module or the second downsampling module. When the number of input channels of the first downsampling module or the second downsampling module is less than the number of output channels, the output of the first convolutional layer is concatenated with the output of the max pooling layer to obtain the output of the first downsampling module or the second downsampling module. In this case, the number of output channels of the first convolutional layer is the difference between the number of output channels and the number of input channels of the first downsampling module or the second downsampling module.
[0073] Specifically, the length of the output data after each pass through the first downsampling module or the second downsampling module is reduced to half that of the input data.
[0074] In one specific embodiment, in the first downsampling module, the output dimension of the first convolutional layer is 31, the kernel size is set to 1×3, the kernel stride is set to 1×2, and the kernel size and stride of the max pooling layer are both set to 1×2. In the second downsampling module, the output dimension of the first convolutional layer is 128, the kernel size is set to 1×3, the kernel stride is set to 1×2, and no max pooling layer is used.
[0075] It is understandable that each downsampling module includes a first convolutional layer and a max pooling layer, and the output of the first convolutional layer can be determined based on the relationship between the number of input channels and the number of output channels.
[0076] Specifically, the first channel pyramid section includes two consecutive channel pyramid modules; the second channel pyramid section includes six consecutive channel pyramid modules.
[0077] Please see Figure 3 , Figure 3 This is a schematic diagram of a channel pyramid module provided in an embodiment of the present invention. Each channel pyramid module includes a second convolutional layer, a first feature pyramid channel, a second feature pyramid channel, a third feature pyramid channel, a fourth feature pyramid channel, and a third convolutional layer.
[0078] Specifically, the output of the second convolutional layer is simultaneously input into the first feature pyramid channel, the second feature pyramid channel, the third feature pyramid channel, and the fourth feature pyramid channel. FP1 The first feature pyramid channel output data is used to output the first feature pyramid channel output data o. FP1 The first summation result is obtained by adding the data from the second feature pyramid channel output. FP2 The first summation result o FP2 The second summation result is obtained by adding the output data of the third feature pyramid channel. FP3 The second sum result o FP3 The output data of the fourth feature pyramid channel is added together to obtain the third addition result. FP4 The first feature pyramid channel output data o FP1 The first sum result o FP2 The result of the second addition is o FP3 The result of adding the third one is o FP4 The concatenation result is used as the output O1 = [o FP1 ,o FP2 ,o FP3 ,o FP4The output O1 of the feature pyramid channel is used as the input data of the third convolutional layer. Then, the output data of the third convolutional layer is added to the input data of the first convolutional layer (i.e., the input data of the channel pyramid module) to obtain the output data of the channel pyramid module.
[0079] Please see Figure 4 , Figure 4 This is a schematic diagram of a feature pyramid channel provided in an embodiment of the present invention. The first, second, third, and fourth feature pyramid channels each include three sequentially connected dilated convolutional layers. The outputs of the first, second, and third dilated convolutional layers are concatenated to obtain the output data for the first, second, third, or fourth feature pyramid channel. In one specific embodiment, the original kernel size of each dilated convolutional layer is set to 1×3.
[0080] Specifically, the dilation rates of the dilated convolutions in the first, second, third, and fourth feature pyramid channels are 1, 1, and 1, respectively. And d+1, where d is the expansion factor of the corresponding channel pyramid module.
[0081] In one specific embodiment, in the first channel pyramid section, the dilation factors of two consecutive channel pyramid modules are 2 and 2, respectively; the output dimension of the second convolutional layer of each channel pyramid module is 16, the kernel size is set to 1×1, and the kernel stride is set to 1×1; the output dimension of the third convolutional layer is 64, the kernel size is set to 1×1, and the kernel stride is set to 1×1. In the second channel pyramid section, the dilation factors of six consecutive channel pyramid modules are 4, 4, 8, 8, 16, and 16, respectively; the output dimension of the second convolutional layer of each channel pyramid module is 32, the kernel size is set to 1×1, and the kernel stride is set to 1×1; the output dimension of the third convolutional layer is 128, the kernel size is set to 1×1, and the kernel stride is set to 1×1.
[0082] S5. The second downsampling feature, the second channel pyramid output, and the third sampling feature are concatenated, and the third concatenated feature is input into the second convolution module and the recognition and classification module connected in sequence to obtain the recognition and classification result.
[0083] Specifically, the second downsampling feature, the second channel pyramid output, and the third sampling feature are concatenated to obtain the third concatenated feature. Then, the third concatenated feature is input into the second convolution module, and the output of the second convolution module is input into the recognition and classification module, which outputs the recognition and classification result.
[0084] Specifically, the second convolutional module includes a convolutional layer with C output channels. out The value is 3, the kernel size is 1×3, and the stride is 1×1.
[0085] Specifically, the recognition and classification module includes a global average pooling layer and a SoftMax layer connected in sequence. The pooling kernel size of the global average pooling layer can be set to 1. The SoftMax layer uses the SoftMax activation function to calculate the probability that the input echo signal is recognized as each category, thus obtaining the recognition and classification result.
[0086] The target recognition method of the present invention constructs a channel feature pyramid model. In the recognition and classification module, the model uses a global average pooling layer instead of a traditional pooling layer, which solves the problem of different input data lengths caused by changes in repetition frequency.
[0087] The target recognition method of this invention constructs a channel feature pyramid model to identify targets from the echo data to be measured. The echo data to be measured is processed by an average pooling module for feature extraction, and then sequentially input into a downsampling module and a channel pyramid part for feature extraction. The channel feature pyramid model includes a channel feature pyramid module, which combines the advantages of the Inception model and the dilated convolution method. It obtains different receptive fields when the convolution kernel size is fixed, and extracts multi-scale features through a parallel structure. It still has good recognition effect and robust recognition performance when the radar repetition rate changes.
[0088] In summary, this embodiment employs a channel feature pyramid model, enabling the network to extract classification features at different scales from radar echo signals. This allows for robust radar target recognition under varying repetition rates without retraining a new classification model during the testing phase. Compared to traditional algorithms, the extracted features at different scales are more separable and maintain good recognition performance even when radar repetition rates change, demonstrating superior target recognition robustness.
[0089] Example 2
[0090] Based on Example 1, this example demonstrates the effectiveness of the radar robust target recognition method based on the channel feature pyramid model through simulation experiments.
[0091] I. Experimental Conditions
[0092] The simulation experiment in this embodiment uses radar echo signals from a training dataset with a radar operating frequency of 8 GHz, a dwell time of 100 ms, and a pulse repetition frequency of 6 kHz. Each aircraft model generates 500 radar echo signals, resulting in a total of 2000 radar echo signals per aircraft type, and the training set contains 6000 radar echo signals. Gaussian white noise with a signal-to-noise ratio of 0 dB is added to the training dataset during the experiment.
[0093] The radar echo signals used in the simulation experiment of this embodiment were generated at a radar operating frequency of 8 GHz and a dwell time of 100 ms. Ten different test datasets were created based on different pulse repetition frequencies: 1 kHz, 2 kHz, 3 kHz, 4 kHz, 5 kHz, 6 kHz, 7 kHz, 8 kHz, 9 kHz, and 10 kHz. Each sub-dataset contained 3000 radar echo signals. Gaussian white noise with a signal-to-noise ratio of 0 dB was added to the training dataset during the experiment.
[0094] The hardware platform for the simulation experiment in this embodiment is: Intel(R) Core(TM) i7-10700 CPU@2.90GHZ processor and 32GB of memory.
[0095] The software platform for the simulation experiment in this embodiment is Windows 10 operating system and Python 3.7.
[0096] The simulation experiment in this embodiment uses data generated by the electromagnetic simulation software CST, which includes 12 models of three types of aircraft: helicopters, propellers, and jets. There are four models of each type of aircraft, and the specific rotor physical parameters are shown in the table below.
[0097] Table 1 Physical Parameters of Aircraft Rotors
[0098] Aircraft target model Number of leaves <![CDATA[L1(m)]]> <![CDATA[L2(m)]]> Rotational speed (r / min) BK17 helicopter 4 0 5.5 383 Mi-17 helicopter 5 0 10.645 185 AS350 helicopter 3 0 5.345 394 Bell 212 helicopter 2 0 7.315 324 SAAB2000 propeller 6 0.28 1.905 950 L-420 propeller 5 0.12 1.15 1650 L-610G propeller 4 0.23 1.675 1150 F406 propeller 3 0.23 1.18 1690 Jet A 30 0.3 1.0 3000 Jet B 38 0.38 1.1 3520 Jet C 27 0.18 0.51 8615 Jet D 33 0.2 0.6 5000
[0099] II. Experimental Content and Results
[0100] The simulation experiment in this embodiment uses the method of Embodiment 1 and the convolutional neural network recognition method to identify the test dataset and obtain the recognition results.
[0101] In the simulation experiment, the convolutional neural network recognition method adopts the radar micro-Doppler narrowband target classification method proposed in "Radar Target Doppler Image Classification and Recognition Method Based on One-Dimensional Convolutional Neural Network".
[0102] The test dataset is classified using the method described in Example 1. First, a training echo dataset is generated. Then, a channel feature pyramid model is constructed according to the channel feature pyramid model structure described in Example 1, and the model is trained to obtain a trained channel feature pyramid model. The 10 test datasets are then used for recognition tests to obtain the recognition results for the test datasets. The recognition accuracy of this method on test data under different repetition frequency conditions is then statistically analyzed.
[0103] The test dataset was identified using a convolutional neural network recognition method. The recognition method described in the patent document was used to perform recognition tests on the test dataset, and the recognition results of the test data were obtained. The recognition accuracy of the method under different repetition frequency conditions was statistically analyzed.
[0104] The classification accuracy of the method in Example 1 and the convolutional neural network recognition method on test data under different repetition frequency conditions was plotted as a curve, as shown in the figure. Figure 5 As shown, Figure 5 The simulation experiment results are shown in the figure provided for the embodiments of the present invention. Figure 5 The horizontal axis represents the repetition rate of the test data, which are 1KHz, 2KHz, 3KHz, 4KHz, 5KHz, 6KHz, 7KHz, 8KHz, 9KHz, and 10KHz respectively. The vertical axis represents the average classification accuracy of the test data under different repetition rate conditions.
[0105] Depend on Figure 5 It is evident that, under the ideal condition of establishing classification templates for all test repetition frequencies, the channel feature pyramid network model outperforms the traditional convolutional neural network. When the test repetition frequencies differ from the training repetition frequencies, the classification performance of all models decreases compared to the ideal situation, with a more significant performance drop as the difference between the test and training repetition frequencies increases. Using time-domain echo signals as model input, the channel feature pyramid network model and the convolutional neural network model show no significant performance decline when the test repetition frequencies are 5kHz and 7kHz. At a test repetition frequency of 4kHz, the performance declines by approximately four and six percentage points, respectively. The channel feature pyramid network model exhibits a relatively smaller performance decline at 1kHz and 4kHz. This is because the channel feature pyramid network model can extract more details from the echo signals by extracting features at different scales, resulting in robust classification even with varying repetition frequencies.
[0106] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A robust radar target recognition method based on a channel feature pyramid model, characterized in that, The trained channel feature pyramid model is used to identify targets in the test echo data, including the following steps: Acquire the echo data to be tested; The echo data to be measured is input into a first average pooling module, a second average pooling module, a third average pooling module, and a first convolution module in parallel to obtain a first sampling feature, a second sampling feature, a third sampling feature, and a convolution feature; the first average pooling module, the second average pooling module, and the third average pooling module have different scales; The convolutional features are concatenated with the first sampling features, and the first concatenated features are input into the first downsampling module and the first channel pyramid part that are connected in sequence to obtain the first downsampling features and the first channel pyramid output. The first downsampling feature, the first channel pyramid output, and the second sampling feature are concatenated, and the second concatenated feature is input into the second downsampling module and the second channel pyramid part, which are connected in sequence, to obtain the second downsampling feature and the second channel pyramid output. Both the first and second downsampling modules include a first convolutional layer and a max-pooling layer connected in sequence. The first channel pyramid part includes two consecutive channel pyramid modules. The second channel pyramid part includes six consecutive channel pyramid modules. Each channel pyramid module includes a second convolutional layer, a first feature pyramid channel, a second feature pyramid channel, a third feature pyramid channel, a fourth feature pyramid channel, and a third convolutional layer, wherein the output of the second convolutional layer is simultaneously input into the second channel pyramid output. The system comprises a first feature pyramid channel, a second feature pyramid channel, a third feature pyramid channel, and a fourth feature pyramid channel. The output data of the first feature pyramid channel is added to the output data of the second feature pyramid channel to obtain a first addition result. The first addition result is added to the output data of the third feature pyramid channel to obtain a second addition result. The second addition result is added to the output data of the fourth feature pyramid channel to obtain a third addition result. The concatenation result of the first feature pyramid channel output data, the first addition result, the second addition result, and the third addition result is used as the input data of the third convolutional layer. The output data of the third convolutional layer is added to the input data of the first convolutional layer to obtain the output data of the channel pyramid module. The second downsampling feature, the second channel pyramid output, and the third sampling feature are concatenated, and the third concatenated feature is input into the second convolution module and the recognition and classification module that are connected in sequence to obtain the recognition and classification result.
2. The radar robust target recognition method based on the channel feature pyramid model according to claim 1, characterized in that, The first average pooling module includes an average pooling layer; The second average pooling module includes two consecutive average pooling layers; The third average pooling module includes three consecutive average pooling layers.
3. The radar robust target recognition method based on the channel feature pyramid model according to claim 2, characterized in that, The output length of the average pooling layer is: in, For the length of the input data, For fill size, The kernel length is 1. This represents the step size.
4. The radar robust target recognition method based on the channel feature pyramid model according to claim 1, characterized in that, When the number of input channels of the first downsampling module or the second downsampling module is greater than the number of output channels, the number of output channels of the first convolutional layer is the same as the number of output channels of the first downsampling module or the second downsampling module. When the number of input channels of the first downsampling module or the second downsampling module is less than the number of output channels, the number of output channels of the first convolutional layer is the difference between the number of output channels and the number of input channels of the first downsampling module or the second downsampling module. The output result of the first convolutional layer is then concatenated with the output result of the max pooling layer to obtain the output of the first downsampling module or the second downsampling module.
5. The radar robust target recognition method based on the channel feature pyramid model according to claim 1, characterized in that, The first feature pyramid channel, the second feature pyramid channel, the third feature pyramid channel, and the fourth feature pyramid channel each consist of three sequentially connected dilated convolutional layers. Specifically, the output results of the first dilated convolutional layer, the second dilated convolutional layer, and the third dilated convolutional layer are concatenated to obtain the output data of the first feature pyramid channel, the second feature pyramid channel, the third feature pyramid channel, or the fourth feature pyramid channel.
6. The radar robust target recognition method based on the channel feature pyramid model according to claim 1, characterized in that, The dilation rates of the dilated convolutions in the first, second, third, and fourth feature pyramid channels are 1, 1, and 1, respectively. , and ,in, This is the expansion factor for the corresponding channel pyramid module.
7. The radar robust target recognition method based on the channel feature pyramid model according to claim 1, characterized in that, The identification and classification module includes a global average pooling layer and a SoftMax layer connected in sequence.
8. The radar robust target recognition method based on the channel feature pyramid model according to claim 1, characterized in that, The training method for the completed channel feature pyramid model includes the following steps: Generate a training dataset, which includes a training echo dataset and a category label for each training sample. The training echo dataset includes echo signals of several categories with a single pulse repetition frequency. The training echo dataset is input into the channel feature pyramid model, and the predicted identification label is output. Using the cross-entropy loss function, the cross-entropy value between the predicted identification label and the corresponding category label of the training sample is calculated. Then, the network parameters are iteratively updated using the backpropagation algorithm until the cross-entropy loss function converges, resulting in the trained channel feature pyramid model. The cross-entropy loss function is: in, Represents the cross-entropy loss function. This represents the predicted category label of the channel feature pyramid network. This represents the true class label of the training samples in the training dataset. , This represents the class index of the training samples in the training dataset. This represents the total number of classes in the training samples in the training dataset. This represents a logarithmic operation with base 10. This indicates the batch size during a forward computation and backward propagation process.