Method for constructing HRRP recognition database based on prototype generation

Abstract

The present invention discloses a method for constructing an HRRP recognition database based on prototype generation, comprising: obtaining a training data set; constructing a polymorphic function f(x, h); wherein the operation rule of the polymorphic function f(x, h) is to divide each sample x in the prototype set into h subsamples and adjust the subsamples after division to the size of the original sample; based on the training data set, using the polymorphic function and single-step parameter matching to perform prototype generation to obtain a prototype set; and processing the prototype set using the polymorphic function to obtain an HRRP recognition database. Compared with traditional methods, this method can ensure lower computational complexity while ensuring that the most valuable information in the HRRP data is mined to construct a high-quality HRRP recognition database, thereby saving storage and computing resources.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a method for constructing an HRRP recognition database based on prototype generation. Background Art

[0002] Radar (Radio Detecting and Ranging) is a radio detection device with all-day, all-weather, and long-range detection capabilities. During operation, a radar antenna transmits electromagnetic waves. These waves are reflected by a target and received by a radar receiver. Finally, by processing and analyzing the received radar echo signals, the radar's presence and location can be detected. A high-resolution range profile (HRRP) is the vector sum of the target's scattered point echoes, acquired by the high-resolution radar, projected along the radar ray. HRRP data is a one-dimensional vector that contains rich structural and characteristic information about the target. It is easy to acquire, store, and process, making it widely used for target model identification. With the rapid development of radar software and hardware technology, the amount of radar HRRP data has increased dramatically. While this massive amount of data provides ample room for data-driven machine learning methods to extract valuable information, high dimensionality, excessive redundancy, and high noise are inherent characteristics of this vast and complex data. Prototype generation is an effective method for eliminating data redundancy, discovering data structure, and improving data quality. By finding a prototype set to represent the target set, data can be reduced from the sample space, thereby enhancing data availability and improving the execution efficiency of the machine learning algorithm.

[0003] In their 2018 paper, "Dataset Distillation," Wang, Zhu, et al. formulated network parameters as functions of a prototype set and formulated the prototype generation task as a two-stage optimization problem. Specifically, their approach focuses on optimizing the prototype set to minimize the loss of a deep network model trained on the prototype set on the original large dataset. However, this approach involves a nested optimization process, requiring the recursive computation graph of the prototype set to be expanded across multiple optimization steps of the network parameters. This leads to expensive computational costs and limited performance.

[0004] In their paper "Dataset Condensation with Gradient Matching," published at the 2021 International Conference on Learning Representations, Zhao et al. proposed matching the gradients of network parameters from both the original dataset and the prototype set to address the problem of recursive graph expansion. However, this approach has several drawbacks. First, the gradient matching method employed is ineffective. Second, while the computational cost of prototype set optimization is somewhat lower than that proposed by Wang, it remains high. Finally, the optimized prototype set is limited in its representativeness. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a method for constructing an HRRP recognition database based on prototype generation. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] A method for constructing an HRRP recognition database based on prototype generation includes:

[0007] Step 1: Get the training dataset;

[0008] Step 2: Construct a polymorphic function f(x, h);

[0009] The operation rule of the polymorphic function f(x, h) is to divide each sample x in the prototype set into h subsamples, and adjust the subsamples after division to the size of the original sample;

[0010] Step 3: Based on the training data set, prototype generation is performed using the multi-form function and single-step parameter matching to obtain a prototype set;

[0011] Step 4: Process the prototype set using the polymorphic function to obtain the HRRP recognition database.

[0012] In one embodiment of the present invention, step 1 includes:

[0013] Several high-resolution range images are extracted from the radar echo database, and the spectrum of each high-resolution range image is calculated using short-time Fourier transform. The obtained spectrums are combined into a training data set Tr.

[0014] In one embodiment of the present invention, step 3 includes:

[0015] 31) Initialize the prototype set using Gaussian distribution And set the maximum number of outer layer iterations K and the maximum number of inner layer iterations T;

[0016] in, represents the prototype set of the tth inner iteration in the kth outer iteration. Initially, t and k are both 0;

[0017] 32) Initialize the deep neural network and pre-train the initialized deep neural network for one iteration using a randomized learning rate on the training data set to obtain the network parameter θ of the current kth outer layer iteration k ;

[0018] 33) From the prototype set And extract a batch of samples from each category in the training data set Tr and And use the polymorphic function to Process and generate sample sets

[0019] Among them, c represents the category label;

[0020] 34) and After inputting the deep neural network respectively, the network parameters θ are obtained using the back propagation algorithm. k The first-order gradient and

[0021] 35) Based on the first-order gradient and Calculate the loss function using the single-step parameter matching formula;

[0022] 36) Using the loss function to calculate the current prototype set Perform back propagation to obtain the prototype set of the t+1th inner iteration in the kth outer iteration

[0023] 37) Input the deep neural network and train it several times to update the network parameters θ k ;

[0024] 38) Let t = t + 1, return to step 33) and perform the next round of inner iteration until the maximum number of inner iterations T is reached, and the prototype set of the T-1 inner iteration in the kth outer iteration is obtained.

[0025] 39) Let k = k + 1, and let the prototype set of the first inner iteration in the k + 1th outer iteration be Return to step 32) and start the next round of outer layer iteration until the maximum number of outer layer iterations K is reached to obtain the final prototype set.

[0026] In one embodiment of the present invention, in step 35), the calculation formula of the loss function is:

[0027]

[0028] Among them, C represents the number of categories, and the value of λ is the number of categories C. is represented as follows:

[0029]

[0030] Among them, A l ,B l Represents the gradient of the network layer l, L is the total number of layers of the deep neural network, out represents the number of output channels of the network layer l, Represents the gradient of the network parameters contained in the i-th output channel of the l-th layer of the network.

[0031] In one embodiment of the present invention, in step 36), the loss function is used to calculate the current prototype set. The expression for back propagation is:

[0032]

[0033] Among them, η M To optimize the learning rate of the prototype set, is the result of the tth iteration, Indicates the use of loss on the prototype set Find the first derivative.

[0034] Beneficial effects of the present invention:

[0035] 1. The method provided by the present invention addresses the situation where there are too many samples in the radar HRRP database. It constructs a polymorphic function and uses it to match single-step parameters to generate a prototype set. This method takes into account the data regularity of the original dataset. When there are too many samples in the radar HRRP database, compared with traditional methods, this method can mine the most valuable information in the HRRP data while maintaining lower computational complexity to obtain a high-quality HRRP recognition database, saving storage and computing resources.

[0036] 2. The method provided by the present invention does not require the expansion of expensive recursive computation graphs, and therefore does not require too high computational costs; and the method utilizes the early model of the network to describe the complete feature space, which not only saves computational resources but also enhances the diversity of the prototype set.

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 1 is a flow chart of a method for constructing an HRRP recognition database based on prototype generation provided by an embodiment of the present invention;

[0039] Figure 2 The present invention provides a flow chart of a radar target recognition method. DETAILED DESCRIPTION

[0040] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0041] Example 1

[0042] See Figure 1 , Figure 1 FIG. 1 is a flow chart of a method for constructing an HRRP recognition database based on prototype generation provided by an embodiment of the present invention, which includes:

[0043] Step 1: Get the training dataset.

[0044] Specifically, a number of high-resolution range images can be extracted from the radar echo database, and the spectrum of each high-resolution range image can be calculated using short-time Fourier transform, and the obtained spectrum images can be combined into a training data set Tr.

[0045] For example, in this embodiment, first, N high-resolution range images can be extracted as samples from the radar echo database of ten types of civil aircraft; wherein the value of N can be an integer greater than 60,000, and the number of HRRPs of the ten types of civil aircraft is an integer greater than 6,000.

[0046] Then, the spectrogram of each sample is calculated using short-time Fourier transform to generate the training dataset.

[0047] It should be noted that the above method can also be used to generate a test data set.

[0048] For example, M high-resolution range images can be extracted from the HRRP data of ten types of civil aircraft received by the radar to generate a test data set, where M is an integer greater than 0.

[0049] Step 2: Construct a multi-dimensional function.

[0050] Specifically, the polymorphic function constructed in this embodiment is represented as f(x, h), where x represents a sample in the prototype set and h represents the number of information samples contained in each sample. The specific operation rule of the polymorphic function f(x, h) is to divide each sample x in the prototype set into h subsamples and adjust the subsamples to the size of the original sample.

[0051] For example, in this embodiment, h=2 can be taken, and the processing flow of the multi-form function is:

[0052] The sample x in the prototype set is divided into two samples by splitting it from the middle part of the sample; then, the two split samples are adjusted to the size of the original sample by the bilinear interpolation method in upsampling.

[0053] Step 3: Based on the training data set, prototype generation is performed using the multi-form function and single-step parameter matching to obtain a prototype set.

[0054] In this embodiment, step 3 specifically includes:

[0055] 31) Initialize the prototype set using Gaussian distribution And set the maximum number of outer iterations K and the maximum number of inner iterations T; where, It represents the prototype set of the tth inner iteration in the kth outer iteration. Initially, t and k are both 0.

[0056] In this embodiment, the number of samples of each type in the prototype set is IPC, the maximum number of outer layer iterations K can be set to 300, and the maximum number of inner layer iterations T can be set to 100.

[0057] 32) Initialize the deep neural network and pre-train the initialized deep neural network for one iteration using a randomized learning rate on the training data set to obtain the network parameter θ of the current kth outer layer iteration k .

[0058] In this embodiment, the deep neural network can adopt any existing network structure, and this embodiment does not limit this.

[0059] 33) From the prototype set And extract a batch of samples from each category in the training data set Tr and And use the polymorphic function to Process and generate sample sets Among them, c represents the category label.

[0060] Specifically, the prototype set The sample drawn from Input multi-form function f(x, h) to get new samples in, The number of samples is Twice the number of samples.

[0061] 34) and Input deep neural network θ respectively kAfterwards, the back propagation algorithm is used to obtain the network parameters θ k The first-order gradient and

[0062] For the specific algorithm process, reference may be made to the existing related technologies, and this embodiment will not be described in detail here.

[0063] 35) Based on the first-order gradient and The loss function is calculated using the single-step parameter matching formula.

[0064] In this embodiment, the calculation formula of the loss function is:

[0065]

[0066] Among them, C represents the number of categories, and the value of λ is the number of categories C. is represented as follows:

[0067]

[0068] Among them, A l ,B l Represents the gradient of the network layer l, L is the total number of layers of the deep neural network, out represents the number of output channels of the network layer l, Represents the gradient of the network parameters contained in the i-th output channel of the l-th layer of the network.

[0069] 36) Using the loss function to calculate the current prototype set Perform back propagation to obtain the prototype set of the t+1th inner iteration in the kth outer iteration The formula is:

[0070]

[0071] Among them, η M To optimize the learning rate of the prototype set, is the result of the tth iteration, Indicates the use of loss on the prototype set Find the first derivative.

[0072] 37) Input the deep neural network and train it several times to update the network parameters θ k .

[0073] Optionally, in this embodiment, the deep neural network training can be iterated 10 times to update the network parameters θ k .

[0074] 38) Let t = t + 1, return to step 33) and perform the next round of inner iteration until the maximum number of inner iterations T is reached, and the prototype set of the T-1 inner iteration in the kth outer iteration is obtained.

[0075] 39) Let k = k + 1, and let the prototype set of the first inner iteration in the k + 1th outer iteration be Return to step 32) and start the next round of outer iteration until the maximum number of outer iterations K is reached to obtain the final prototype set

[0076] First, this embodiment does not need to expand the expensive recursive calculation graph, so it does not require too high a computational cost; secondly, it takes into account the data regularity of the original data set and generates more prototype set samples under the same storage budget; secondly, it optimizes the prototype generation matching formula to generate a more informative prototype set; finally, the present invention does not use thousands of networks to splice out a large sample feature space, but uses the early model of the network to describe the complete feature space, which not only saves computing resources but also enhances the diversity of the prototype set.

[0077] Step 4: Process the prototype set using the polymorphic function to obtain the HRRP recognition database.

[0078] The method provided by the present invention addresses the situation where there are too many samples in the radar HRRP library. A polymorphic function is constructed and the polymorphic function is used to match the single-step parameters to generate a prototype set. The method takes into account the data regularity of the original data set. When there are too many samples in the radar HRRP library, compared with traditional methods, the method can, while ensuring lower computational complexity, mine the most valuable information in the HRRP data to obtain a high-quality HRRP recognition database, thereby saving storage resources and computing resources.

[0079] Example 2

[0080] Based on the above embodiment 1, this embodiment also provides a radar target recognition method. Figure 2 , Figure 2 1 is a flow chart of a radar target recognition method provided by an embodiment of the present invention, which includes:

[0081] S1: Build an HRRP recognition database and extract training sample sets and test sample sets from it;

[0082] S2: Using the training sample set to train the deep neural network until the network converges;

[0083] S3: Input the test sample set into the trained deep neural network to obtain a recognition result.

[0084] In step S1, the HRRP recognition database is constructed using the method provided in the above embodiment 1. After the HRRP recognition database is constructed, samples can be selected from the database to form a training sample set and a test sample set.

[0085] Therefore, the radar target recognition method provided by this embodiment can also save storage resources and computing resources, and has a good target recognition effect.

[0086] Example 3

[0087] The beneficial effects of the present invention are verified and explained below through simulation experiments.

[0088] 1. Test conditions:

[0089] This experiment uses ten types of civil aircraft physical parameters, as shown in Table 1, and sets the size of the prototype set to 0.1%, 1%, and 5% of the training dataset size.

[0090] Table 1 Physical parameters of ten types of civil aircraft

[0091] serial number Aircraft model Wingspan (m) Total length of fuselage (m) Overall height of fuselage (m) 1 Airbus A319 34.09 33.84 11.76 2 Airbus A320 34.09 37.57 11.76 3 Airbus A321 34.09 44.51 11.76 4 Airbus A330-200 60.30 58.82 17.39 5 Airbus A330-300 60.30 63.69 16.83 6 Airbus A350-941 64.75 66.80 17.05 7 Boeing 737-700 34.30 33.60 12.50 8 Boeing 737-800 34.30 39.50 12.50 9 Boeing 747-89L 59.64 70.60 19.30 10 Bombardier CRJ-900 24.85 36.19 7.57

[0092] 2. Experimental content and results analysis

[0093] Based on the above aircraft information, three prototype sets were selected from the training data set using the method of the present invention and the traditional method, respectively. These sets were input into the deep neural network for training. The test set was then fed into the trained deep neural network to obtain three sets of recognition results using the method of the present invention and three sets of recognition results using the traditional method. The results are shown in Table 2.

[0094] Table 2 Comparison of classification accuracy between traditional method and the method of the present invention

[0095]

[0096] It can be seen from Table 2 that the method of the present invention has a higher classification accuracy than the traditional method.

[0097] The aforementioned performance differences stem primarily from three key differences. First, the present invention considers the data regularity of the original dataset, generating more prototype set samples within the same storage budget. Second, the present invention optimizes the prototype generation and matching formula to generate a more informative prototype set. Finally, rather than using thousands of networks to stitch together a large sample feature space, the present invention leverages the network's early models to describe the complete feature space, saving computational resources and increasing the diversity of the prototype set. Therefore, the prototype-generation-based HRRP recognition database construction method can generate the most valuable samples from the HRRP database when the radar HRRP library is overpopulated, thus conserving both storage and computational resources.

[0098] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for constructing an HRRP recognition database based on prototype generation, characterized in that: include: Step 1: Get the training dataset; Step 2: Construct a multi-dimensional function ; Wherein, the multi-form function The operation rule is to convert each sample in the prototype set Divided into Sub-samples are generated and the sub-samples are adjusted to the size of the original sample; Step 3: Based on the training data set, prototype generation is performed using the multi-form function and single-step parameter matching to obtain a prototype set; Step 4: Processing the prototype set using the polymorphic function to obtain a HRRP recognition database; Wherein, step 3 includes: 31) Initialize the prototype set using Gaussian distribution , and set the maximum number of outer iterations K and the maximum number of inner iterations T ; in, Indicates the k In the outer iteration t The prototype set of the inner iteration, initially, t and k All are 0; 32) Initialize the deep neural network and pre-train the initialized deep neural network for one iteration using a randomized learning rate on the training data set to obtain the current k Network parameters for the next outer iteration ; 33) From the prototype set And extract a batch of samples from each category in the training data set Tr and , and use the polymorphic function to Process and generate sample sets ; in, c Represents the category label; 34) and After inputting the deep neural network respectively, the network parameters are obtained using the back propagation algorithm. The first-order gradient and ; 35) Based on the first-order gradient and Calculate the loss function using the single-step parameter matching formula; 36) Using the loss function to calculate the current prototype set Perform back propagation to obtain k In the outer iteration t +1 inner iteration prototype set ; 37) Input deep neural network and train it several times to update network parameters ; 38) Order t = t +1, return to step 33) and perform the next round of inner layer iteration until the maximum number of inner layer iterations is reached T , get the k In the outer iteration T -1 prototype set of inner iterations ; 39) Order k = k +1, and make the k +1 prototype set of the first inner iteration in the first outer iteration , return to step 32) to start the next round of outer layer iteration until the maximum number of outer layer iterations is reached K , get the final prototype set ; Wherein, in step 35), the calculation formula of the loss function is: ； in, C Indicates the number of categories, The value is the number of categories C , The expression is as follows: ； in, Indicates the network The gradient of the layer, L is the total number of layers in the deep neural network, out Indicates the network The number of layer output channels, Indicates the network Tier i The gradient of the network parameters contained in the output channels; Among them, in step 36), the current prototype set is The expression for back propagation is: ； in, To optimize the learning rate of the prototype set, For the t The result of the iterations, Indicates use Prototype Set Find the first derivative.

2. The method for constructing an HRRP recognition database based on prototype generation according to claim 1, characterized in that: Step 1 includes: Several high-resolution range images are extracted from the radar echo database, and the spectrum of each high-resolution range image is calculated using short-time Fourier transform. The obtained spectrums are combined into a training data set Tr.

3. A radar target recognition method, characterized in that: include: Adopt the method described in any one of claims 1-2 to construct an HRRP recognition database, and extract a training sample set and a test sample set from it; Training a deep neural network using the training sample set until the network converges; The test sample set is input into the trained deep neural network to obtain the recognition result.

Images