Hierarchical clustering self-distillation radar target identification method, system and device

Abstract

The invention relates to the technical field of radar target identification, and provides a hierarchical clustering self-distillation radar target identification method, system and device. The method comprises the following steps: obtaining a radar target identification data set, and preprocessing the data set to obtain preprocessed identification data; inputting the preprocessed identification data into a CSD clustering model to obtain a CSD weight; according to the CSD weight, training the CSD clustering model by using small-batch supervision loss and additional CSD loss to obtain an optimal CSD clustering model and a corresponding optimal soft label; and inputting a to-be-identified radar heat map and the corresponding optimal soft tag into the optimal CSD clustering model to obtain a target identification result. According to the method, knowledge distillation is carried out on the current prediction result, adaptive clustering selection is realized by using a clustering evaluation method, and compared with other methods, the method does not depend on any hyper-parameter, so that the learning ability of the radar target recognition network is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target recognition, and provides a radar target recognition method, system and device based on hierarchical clustering self-distillation. Background Art

[0002] In recent years, intelligent optical image processing has made great progress. However, the effect of optical images is affected by time and meteorological conditions. Radar uses electromagnetic waves with lower frequencies than visible light for target detection, which makes radar more robust to different environments and greatly increases the detectable distance. Therefore, radar has a wide range of applications in the fields of autonomous driving, missile guidance, remote sensing, underground exploration, etc. Research on radar target recognition technology can effectively improve the accuracy and automation of radar target recognition, and promote the development of different fields.

[0003] Current mainstream radar target recognition technologies learn from labeled data. Most of the latest and most widely used achievements are based on the supervised learning paradigm, including ResNet, Bert, Swin Transformer, Pathways, etc. Most of the work on radar target recognition uses One-hot labels. One-hot encoding is a method for converting discrete categorical labels into vectors, where each category is represented by a binary vector with only one element being 1 and the other elements being 0. Many researchers have studied how to convert One-hot labels back into soft labels with similarity information, which is called label enhancement. Subsequently, knowledge distillation was proposed. Knowledge distillation is a technique for training smaller and lighter neural network models (usually called student models) to learn knowledge from a larger and more complex model (usually called teacher model). The main idea of knowledge distillation is to transfer the knowledge of the large model (teacher model) to the small model (student model) so that the student model can better generalize and perform under the same task. This technique uses the teacher model to also convert One-hot labels back into soft labels with similarity information, i.e., label enhancement, to improve the performance of the small model. Knowledge distillation can be divided into three categories: offline distillation, online distillation, and self-distillation according to the differences in the distillation scheme. In offline distillation, the student model is trained using the soft labels of the trained teacher model. In online distillation, the student model and the teacher model are trained simultaneously. Online distillation improves the training efficiency of knowledge distillation. Different from other distillation schemes, there is no teacher model in self-distillation. In self-distillation, the distillation process is synchronized with the training process. Therefore, self-distillation can be regarded as a special online distillation. Self-distillation takes into account the efficiency of online distillation and overcomes the disadvantages of online distillation.

[0004] With the development of deep learning, various self-distillation algorithms have been proposed: self-distillation with the number of epochs as the time span; self-distillation with the batch size as the time span; and OSKD (Online Subclass Knowledge Distillation) which introduces the concept of subclasses and performs contrastive learning based on labels for the current batch. These techniques have improved the ability of radar target recognition, but there are still some problems. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a method, system and device for radar target recognition based on hierarchical clustering self-distillation, which avoids the complexity and inaccuracy caused by setting hyperparameters and simplifies the use of self-distillation.

[0006] The present invention provides a method for radar target recognition based on hierarchical clustering self-distillation, comprising the following steps: S1: Obtain a radar target recognition data set, and preprocess the data set to obtain the recognition data after preprocessing; S2: Input the recognition data after preprocessing into a CSD (Clustering Self-Distillation) clustering model to obtain CSD weights; S3: According to the CSD weights, train the CSD clustering model using mini-batch supervised loss and additional CSD loss to obtain the optimal CSD clustering model and the corresponding optimal soft labels; S4: Input the radar heat map to be recognized and the corresponding optimal soft labels into the optimal CSD clustering model to obtain the target recognition result.

[0007] According to the method for radar target recognition based on hierarchical clustering self-distillation provided by the present invention, step S1 includes: S11: Perform FFT (Fast Fourier Transform) transformation on the radar target recognition data set to obtain a radar heat map to be trained; Wherein, the radar target recognition data set is a signal after analog-to-digital conversion, along the sample, chirp signal and antenna dimensions; S12: Convert the radar heat map to be trained into a labeled data set , , wherein, is the input vector, is the hard label, is the vector ordinal number, , is the total number of vectors, and the set of the input vectors is the input set , and the set of the hard labels is the output set ; S13: For the input set and the output set , define the neural network mapping such that ; where the weight parameters of the neural network are , , where is the weight parameter of the th layer, is the layer ordinal number of the neural network, is the total number of layers of the neural network, .

[0008] According to a hierarchical clustering self-distillation radar target recognition method provided by the present invention, step S2 includes: S21: Calculate the prediction probability according to the input set, the weight parameters of the neural network, and the neural network mapping, and determine whether the prediction probability is available to obtain an available prediction probability; S22: Cluster the available prediction probabilities to obtain the final CSD clustering model result; S23: Evaluate the final CSD clustering model result by EB (elbow method) to obtain the CSD weight.

[0009] According to a hierarchical clustering self-distillation radar target recognition method provided by the present invention, step S21 includes: S211: Calculate the prediction probability of the th vector: where is the th element of the prediction probability of the th vector, is the th element of the score, is the distillation temperature hyperparameter, is the element ordinal number of the hard label, is the total number of elements of the hard label, , is the th element of the score; S212: Determine whether the prediction probability is available: When the prediction probability satisfies: at this time The predicted probability is available; if the condition is not met, it is not available, and the available predicted probability is obtained. , where is the maximum value function, is the available probability ordinal number, is the total number of available probabilities, , is the th element of the th hard label in the

[0010] According to a hierarchical clustering self-distillation radar target recognition method provided by the present invention, step S22 includes: S221: Calculate and : Where is the clustering set of , is the th prediction result of the rd set in is the th clustering soft label of the in the rd set, is the center distance between two clustering soft labels; is the th clustering soft label of the in the th set, , , are all set ordinals; S222: Update the clustering set: Where is the clustering set of , is the th set in is the th set in is the in the set, is the center distance of the soft label of set clustering, is in the set, is in the set, is in the set, is the center distance of the soft label of set clustering, , , are all set ordinals; , ; S223: Traverse all clustering sets to obtain the final CSD clustering model result .

[0011] According to a hierarchical clustering self-distillation radar target recognition method provided by the present invention, step S23 includes: S231: Calculate the sum of squared errors of the final CSD clustering model result: wherein, is the sum of squared errors of, is average value of; S232: Calculate the second derivative of, and judge the best CSD clustering model result : wherein, is the second derivative of, is sum of squared errors of, is sum of squared errors of; S233: Calculate the CSD weight : wherein, is center distance of the soft label of clustering, For the cluster purity of the is based on the number of categories obtained from the hard labels within the

[0012] According to a hierarchical clustering self-distillation radar target recognition method provided by the present invention, step S3 includes: training the model using a traditional mini-batch supervised loss and an additional CSD loss, and the loss function of the model is: wherein, is a mini-batch sample, is the matrix transpose of the predicted probability of the is the total number of elements of is the corresponding optimal clustering soft label, is the matrix transpose of is the the prediction result of the

[0013] The present invention also provides a hierarchical clustering self-distillation radar target recognition system, including: a data preprocessing module: obtaining a radar target recognition data set, and preprocessing the data set to obtain preprocessed recognition data; a model training module: inputting the preprocessed recognition data into a CSD clustering model to obtain CSD weights; according to the CSD weights, training the CSD clustering model using a mini-batch supervised loss and an additional CSD loss to obtain an optimal CSD clustering model and corresponding optimal soft labels; a target recognition module: inputting a radar heat map to be recognized and the corresponding optimal soft labels into the optimal CSD clustering model to obtain a target recognition result.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps of the hierarchical clustering self-distillation radar target recognition method as described in any one of the above.

[0015] One or more of the above technical solutions in the embodiments of the present invention have at least one of the following technical effects: A method, system and device for radar target recognition based on hierarchical clustering self-distillation provided by the present invention perform knowledge distillation on the current prediction results by using knowledge distillation and clustering evaluation, and have the following advantages: 1. Compared with other radar target recognition algorithms based on soft labels, the knowledge extracted by this algorithm comes from the latest weights, which ensures the correctness of the extracted knowledge and provides an idea for subsequent algorithms for self-distillation based on the current network.

[0016] 2. The present invention obtains the soft labels of subclasses through an adaptive clustering algorithm. Compared with self-distillation algorithms that span time and have complex hyperparameters, this algorithm avoids the complexity and inaccuracy brought by setting hyperparameters, and simplifies the use of self-distillation.

[0017] 3. The present invention obtains the inter-class similarity based on both similar samples and dissimilar samples at the same time, making the inter-class similarity information of self-distillation more complete. This not only improves the performance of self-distillation, but also helps to more completely analyze the potential subclass semantic information in the deep learning network for radar target recognition.

[0018] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. Description of the Drawings

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 is a schematic flow chart of a method for radar target recognition based on hierarchical clustering self-distillation provided by the present invention.

[0021] Figure 2 is a structural block diagram of a device for radar target recognition based on hierarchical clustering self-distillation provided by the present invention.

[0022] Figure 3 is a curve graph of the average test accuracy of different methods.

[0023] Figure 4 is Figure 3 an enlarged view of part A.

[0024] Figure 5 is a schematic structural diagram of an electronic device provided by the present invention.

[0025] Reference Signs: 101. Data preprocessing module; 102. Model training module; 103. Target recognition module; 810. Processor; 820. Communication interface; 830. Memory; 840. Communication bus. Detailed implementation manners

[0026] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention. The following embodiments are used to illustrate the present invention but cannot be used to limit the scope of the present invention.

[0027] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without conflict, those skilled in the art can combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0028] The following combines Figures 1 to 5 to describe the present invention.

[0029] Embodiment As Figure 1 shown, Figure 1 is a schematic flowchart of a radar target recognition method based on hierarchical clustering self-distillation provided by the present invention. Specifically, it includes the following steps: S1: Obtain a radar target recognition data set, and preprocess the data set to obtain the recognized data after preprocessing; S2: Input the recognized data after preprocessing into the CSD clustering model to obtain the CSD weights; S3: According to the CSD weights, use the mini-batch supervised loss and the additional CSD loss to train the CSD clustering model to obtain the optimal CSD clustering model and the corresponding optimal soft labels; S4: Input the radar heat map to be recognized and the corresponding optimal soft labels into the optimal CSD clustering model to obtain the target recognition result.

[0030] In the radar target recognition based on hierarchical clustering distillation, hierarchical clustering is performed on the prediction results of the current batch, and the results of hierarchical clustering are evaluated based on a clustering evaluation method to obtain the optimal clustering results. According to the smoothness assumption, the instances corresponding to the predictions within each cluster are considered to belong to the same subclass. A class may be divided into multiple clusters. These clusters have different characteristics similar to other classes. These similar characteristics are the inter-class similarity information. At this time, the inter-class similarity from the same-class samples is obtained. Similarly, the instances within the same cluster may belong to different classes, which represents the inter-class similarity generated between different-class samples. The soft labels of the instances within the subclass are the cluster centers. The final training loss is targeted at both the conventional one-hot labels and the soft labels. The radar target recognition of the present invention is knowledge extraction based on a clustering method. It divides the sample predictions randomly distributed in the feature space into several clusters. Each cluster represents a subclass. The subclass contains inter-class similarity information. This similarity information is learned by the model to enhance the learning ability of the model. Briefly speaking, the radar target recognition based on hierarchical clustering distillation is a processing method for the predictions of the current batch. Next, the detailed process of radar target recognition based on hierarchical clustering distillation is described.

[0031] Specifically, step S1 includes: S11: Perform FFT transformation on the radar target recognition data set to obtain the radar heat map to be trained; Among them, the radar target recognition data set is a signal after analog-to-digital conversion, including information such as samples, chirp signals, and antenna dimensions; S12: Convert the radar heat map to be trained into a labeled data set , , where is the input vector, is the hard label, is the vector ordinal number, , is the total number of vectors, and the set of input vectors is the input set , and the set of hard labels is the output set ; S13: For the input set and the output set , define the neural network mapping , such that ; Among them, the weight parameter of the neural network is , , where is the weight parameter of the th layer, is the layer ordinal number of the neural network, is the total number of layers of the neural network, 。

[0032] The core of the CSD algorithm proposed by the present invention is to use the clustering center as a soft label. Therefore, the potential clustering objects are the scores (logits) or predictions related to the labels. The distributions of logits can vary greatly, which results in similar samples possibly having a greater distance. Therefore, a more appropriate clustering object is the prediction result of the neural network. Among them, step S2 is the full step of the CSD clustering model, including: calculating the prediction probability according to the input set, the weight parameters of the neural network, and the neural network mapping, and determining whether the prediction probability is available to obtain an available prediction probability; clustering the available prediction probabilities to obtain the final CSD clustering model result; evaluating the final CSD clustering model result using the EB (elbow method) to obtain the CSD weight.

[0033] Specifically, the specific implementation steps of the CSD clustering model are as follows: S21: Calculate the prediction probability according to the input set, the weight parameters of the neural network, and the neural network mapping, and determine whether the prediction probability is available to obtain an available prediction probability.

[0034] Specifically, S211: Calculate the prediction probability of the th vector : Among them, is the th element of the prediction probability of the th vector, is the th element of the score, is the distillation temperature hyperparameter, is the element ordinal number of the hard label, is the total number of elements of the hard label, , is the th element of the score; S212: Determine whether the prediction probability is available: When the prediction probability satisfies: at this time Among them, is the maximum value function, and the prediction probability is available. If it does not meet the condition, it is not available, and the available prediction probability is obtained, where is the available probability ordinal number, is the total available probability, , is the -th -dimensional element of the hard label.

[0035] S22: Cluster the available prediction probabilities to obtain the final CSD clustering model result; Specifically, S221: Calculate and : Among them, is the clustering set, is the -th prediction result of the -th set, is the -th clustering soft label of the -th set, is the number of elements in the -th set, is the center distance between two clustering soft labels; is the -th clustering soft label of the -th set, is the , , are all set ordinals; S222: Update the clustering set: Among them, is the clustering set, is the -th set, is the -th set, is the -th set, is the -th center distance of the clustering soft label of the is the the set, is the set, is the set, is the center distance of the set clustering soft label, , , are all set ordinals; , ; S223: Traverse all clustering sets to obtain the final CSD clustering model result .

[0036] S23: Perform EB evaluation on the final CSD clustering model result to obtain the CSD weight.

[0037] Specifically, S231: Calculate the sum of squared errors of the final CSD clustering model result: Among them, is the sum of squared errors of, is the average value of; S232: Calculate the second derivative of, and judge the best CSD clustering model result : Among them, is the second derivative of, is the sum of squared errors of, is Sum of squared errors; The sum of squared errors is an indicator that measures the clustering error of all samples and is used to evaluate the quality of the clustering effect. The elbow method is a technique that helps determine the optimal number of clusters in a dataset. As the number of clusters increases, the sample partitioning becomes more refined, and the degree of aggregation of each cluster gradually increases, resulting in a decrease in the sum of squared errors. If the number of clusters is less than the actual number of clusters, the sum of squared errors will decrease significantly because increasing the number of clusters will greatly improve the degree of aggregation of each cluster. Once the number of clusters reaches the true number of clusters, as the number of clusters continues to increase, the improvement in the degree of cluster aggregation will become rapidly smaller, and the decrease in the sum of squared errors will slow down sharply. Then, as the number of clusters continues to increase, the sum of squared errors will tend to level off. This relationship between the sum of squared errors and the number of clusters can be visualized graphically, and its shape is elbow-shaped. The inflection point of the elbow represents the true number of clusters in the data, and this point is defined as the position where the change in the sum of squared errors decreases the most.

[0038] S233: Calculate the CSD weight : where is the central distance of the clustering soft label of is the cluster purity of the is based on the number of categories obtained from the hard labels within the

[0039] The larger the number of samples in the same category, the stronger the regularity of the subclass features expressed by the cluster and the more likely it is to conform to the actual situation. In addition, the value range of the cluster purity is [0,1]. Therefore, the average cluster purity can be regarded as 0.5.

[0040] Specifically, step S3 includes: training the model using the traditional mini-batch supervised loss and the additional CSD loss. The loss function of the model is: where is the mini-batch sample, is the matrix transpose of is the predicted probability of the th vector, is the total number of elements of is the corresponding optimal clustering soft label, is the matrix transpose of is the th prediction result of the set.

[0041] In the embodiments of the present invention, the radar target recognition data is the dataset Deep Open Space Segmentation using Automotive Radar (SCORP). The vehicle carrying the data collected by SCORP is a car equipped with side view radars and cameras, and the purpose is the drivable open space in the parking lot. SCORP contains radar output data and camera output data. This dataset uses a linear frequency modulated continuous wave radar with a frequency of 76Ghz, and uses 8 virtual channels composed of 2 Tx elements transmitting sequentially and 4 Rx elements receiving coherently. In the multiple input multiple output mode, environmental radar observation data is collected. The radar data is a signal after analog-to-digital conversion. Applying the fast Fourier transform along the sample, Chirps (linear frequency modulated signals), and antenna dimensions can obtain the range-Doppler azimuth representation, that is, the radar heat map, which is the input of this implementation. The camera output data is used to annotate the radar data. The annotation is divided into two categories, one is the drivable space and the other is the non-drivable space. The resulting dataset consists of 3913 frames and is collected in 11 driving sequences. The RDA (Radar Data Array) data of each frame contains a 256×256×64 matrix, representing the intensity of each azimuth, range, and speed. Since the 256×256×64 matrix as the input of the network will greatly occupy computing resources, the 256×256×64 matrix is compressed into a 128×128×32 matrix through bilinear interpolation. The test set is 10% of SCORP randomly selected, and the other data except the test set is used as the training set.

[0042] All experiments are carried out using the PyTorch framework. The mini-batch gradient descent method is used for network training. The settings of the hyperparameters are as follows: in the convolutional layers with 64 and 128 output channels, the dropout probability is set to 0.1; in the convolutional layers with 256 and 512 output channels, the dropout probability is set to 0.2; in the convolutional layer with 1024 output channels, the dropout probability is set to 0.3. The batch size of all experiments is set to 32. All networks are trained until the 10th epoch. During the training process, the learning rate of the optimizer is set to 0.1 and the momentum is set to 0.9. L2 regularization is adopted, and its weight decay is 0.0001. In order to save computing resources, the embodiments of the present invention use 16-bit floating-point numbers for training.

[0043] Input the radar heat map into the trained radar target recognition network to obtain the result of radar target recognition. Apply the method of the present invention and the radar target recognition methods based on other knowledge distillation methods to perform segmentation experiments on the images in the radar target dataset respectively, and evaluate the respective performances of these methods from the accuracy of the segmentation results, the integrity of the detailed information, and the clarity of the edge boundaries. The results are shown in Table 1 and Figure 3 and Figure 4 shown as follows.

[0044] Table 1 Comparison table of radar target recognition between hierarchical clustering distillation and other self-distillation methods

[0045] Among them, W / o Self-Distillation (no distillation method), LS (Label Smooth Loss), SD (The Snapshot Distillation), PS-KD (Progressive Self-knowledge Distillation), DLB (The Self-Distillation from Last Mini-Batch), OSKD (Online Subclass Knowledge Distillation) are other radar target recognition methods, and CSD is the method proposed by the present invention. As shown in Table 1, the method proposed by the present invention is superior to the radar target recognition using other self-distillation methods.

[0046] As Figure 3 and Figure 4 shown, it can be intuitively seen the performance improvement of the radar target recognition proposed by the present invention on the baseline model and the comparison results with other methods, which clearly describes the regularization effect of this method. Therefore, the proposed method of radar target recognition based on hierarchical clustering self-distillation can improve the performance in most cases.

[0047] As Figure 2 shown, Figure 2 is the structural block diagram of a radar target recognition device based on hierarchical clustering self-distillation provided by the present invention, including: Data preprocessing module 101: Obtain the radar target recognition dataset, and preprocess the dataset to obtain the preprocessed recognition data; Model training module 102: Input the preprocessed recognition data into the CSD clustering model to obtain the CSD weights; according to the CSD weights, use the mini-batch supervised loss and the additional CSD loss to train the CSD clustering model to obtain the optimal CSD clustering model and the corresponding optimal soft labels; Target recognition module 103: Input the radar heat map to be recognized and the corresponding best soft label into the best CSD clustering model to obtain the target recognition result.

[0048] Figure 5 An example of the physical structure diagram of an electronic device is shown as Figure 5 shown. The electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840. Among them, the processor 810, the communication interface 820, and the memory 830 complete mutual communication through the communication bus 840. The processor 810 can call the logical instructions in the memory 830 to execute a radar target recognition method based on hierarchical clustering self-distillation. The method includes: S1: Obtain a radar target recognition data set, and preprocess the data set to obtain the preprocessed recognition data; S2: Input the preprocessed recognition data into the CSD clustering model to obtain the CSD weight; S3: According to the CSD weight, use the mini-batch supervised loss and the additional CSD loss to train the CSD clustering model to obtain the best CSD clustering model and the corresponding best soft label; S4: Input the radar heat map to be recognized and the corresponding best soft label into the best CSD clustering model to obtain the target recognition result.

[0049] In addition, when the logical instructions in the above-mentioned memory 830 are implemented in the form of software functional units and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0050] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative work.

[0051] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course also by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0052] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

[0053] It should be noted that the embodiments of the present disclosure can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic: the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code is provided on a programmable memory or a data carrier such as an optical or electronic signal carrier.

[0054] In addition, although the operations of the methods of the present disclosure are depicted in the drawings in a particular order, this is not a requirement or implication that the operations must be performed in that particular order, or that all of the illustrated operations must be performed to achieve the desired result. On the contrary, the steps depicted in the flowcharts may be altered in their order of execution. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step and executed, and / or one step may be decomposed into multiple steps and executed. It should also be noted that the features and functions of two or more devices according to the present disclosure may be embodied in one device. Conversely, the features and functions of one device described above may be further divided and embodied by multiple devices.

[0055] Although the present disclosure has been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the specific embodiments disclosed. The present disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A radar target recognition method based on hierarchical clustering self-distillation, characterized in that: The following steps are involved: S1: Obtain a radar target recognition data set, and preprocess the data set to obtain recognition data after preprocessing; S2: Input the preprocessed recognition data into the CSD clustering model to obtain the CSD weights, including: S21: Calculate the prediction probability according to the input set, the weight parameters of the neural network and the neural network mapping, and determine whether the prediction probability is available to obtain the available prediction probability; S22: clustering the available prediction probabilities to obtain a final CSD clustering model result; S23: performing EB evaluation on the final CSD clustering model result to obtain a CSD weight; S3: According to the CSD weight, the CSD clustering model is trained using a small batch supervision loss and an additional CSD loss to obtain an optimal CSD clustering model and a corresponding optimal soft label; S4: Inputting the radar heat map to be identified and the corresponding best soft label into the best CSD clustering model to obtain a target recognition result.

2. The radar target recognition method based on hierarchical clustering self-distillation according to claim 1, characterized in that: Step S1 includes: S11: performing FFT transformation on the radar target recognition data set to obtain a radar heat map to be trained; The radar target recognition data set is a signal that has been converted from analog to digital, including samples, linear frequency modulation signals, and antenna dimensions; S12: Convert the radar heat map to be trained into a labeled data set , ,in, is the input vector, For hard tags, is the vector ordinal, , is the total number of vectors, the set of input vectors is the input set , the set of hard labels is the output set ; S13: For the input set and output set , define the neural network mapping , so that ; Among them, the weight parameter of the neural network is , ,in, For the The weight parameters of the layer, is the number of layers of the neural network, is the total number of layers in the neural network, .

3. The radar target recognition method based on hierarchical clustering self-distillation according to claim 2, characterized in that: Step S21 includes: S211: Calculate the The predicted probability of a vector : in, For the The predicted probability of the vector Dimensional elements, For the score Dimensional elements, is the distillation temperature hyperparameter, is the element ordinal of the hard label, is the total number of hard-labeled elements, , For the score Dimensional elements; S212: Determine whether the predicted probability is available: When the predicted probability satisfies: hour The predicted probability is available, if not satisfied then unavailable, obtain the available predicted probability ,in, To obtain the maximum value function, is the available probability ordinal, is the total number of available probabilities, , For the Hard tag Dimensional elements.

4. The radar target recognition method based on hierarchical clustering self-distillation according to claim 3, characterized in that: Step S22 includes: S221: Calculation and : in, for The clustering set of for The The collection The prediction results, for The The cluster soft labels of the set, for The the number of elements in the collection, is the center distance between two cluster soft labels; for The The cluster soft labels of the set, for The The cluster soft labels of the set, , , All are set ordinals; S222: Update clustering set: in, for The clustering set of for The gather, for The gather, for The gather, For the The collection The center distance of the soft labels of the clustering, for The gather, for The gather, for The gather, For the The collection The center distance of the soft labels of the clustering, , , All are set ordinals; , ; S223: Traverse all cluster sets and obtain the final CSD clustering model result .

5. The radar target recognition method based on hierarchical clustering self-distillation according to claim 4, characterized in that: Step S23 includes: S231: Calculate the sum of squared errors of the final CSD clustering model results: in, for The sum of squared errors, for The average value of S232: Calculation The second-order derivative of and judge the best CSD clustering model results : in, for The second derivative of for The sum of squared errors, for The sum of squared errors; S233: Calculate CSD weight : in, for The center distance of the cluster soft label, for No. The cluster purity of the set, Based on No. The number of categories obtained by hard labels in the set.

6. The radar target recognition method based on hierarchical clustering self-distillation according to claim 5, characterized in that: Step S3 includes: training the model using the traditional small batch supervision loss and the additional CSD loss, and the loss function of the model is for: in, is a small batch of samples, for The matrix transpose of For the The predicted probability of a vector, for The total number of elements of is the corresponding best clustering soft label, for The matrix transpose of for No. The collection prediction results.

7. A radar target recognition system based on hierarchical clustering self-distillation, used to execute a radar target recognition method based on hierarchical clustering self-distillation as claimed in any one of claims 1 to 6, characterized in that: include: Data preprocessing module: obtains radar target recognition data set, preprocesses the data set to obtain recognition data after preprocessing; Model training module: input the preprocessed recognition data into the CSD clustering model to obtain the CSD weight; According to the CSD weights, the CSD clustering model is trained using a mini-batch supervision loss and an additional CSD loss to obtain an optimal CSD clustering model and a corresponding optimal soft label; Target recognition module: input the radar heat map to be recognized and the corresponding best soft label into the best CSD clustering model to obtain the target recognition result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the radar target recognition method based on hierarchical clustering self-distillation as described in any one of claims 1 to 6 are implemented.

Images