Small sample identification method and device for space targets, computing device and storage medium

Abstract

The present application relates to the technical field of data processing, and particularly relates to a small sample identification method and device for space targets, a computing device and a storage medium. The method comprises: obtaining an initial training set and a test set; inputting the initial training set into a pre-constructed neural network to obtain an initial identification model and deep dimension reduction features of training samples; for each test sample, performing: inputting the current test sample into the initial identification model to determine whether the current test sample is expanded to the initial training set based on a clustering method, the deep dimension reduction features of the training samples and deep dimension reduction features of the current test sample; if yes, performing secondary training on the initial identification model based on an identification result of the current test sample and a judgment result; if no, judging the next test sample until obtaining a final training set, a target identification model and an identification result of the test sample. The present application can effectively improve the identification accuracy of space targets in a small sample situation.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing technology, and in particular to a method, apparatus, computing device and storage medium for small sample identification of spatial targets. Background Technology

[0002] With the rapid development of space technology, space situational awareness has become a research hotspot in countries around the world. Space target identification, as an important part of space situational awareness, is one of the key technologies of space systems.

[0003] However, in practical applications, due to the angle at which the observation station and the space target pass overhead, only a limited amount of electromagnetic scattering data can usually be collected. Existing space target identification methods often suffer from low accuracy in small sample situations due to insufficient training data. Therefore, achieving efficient space target identification under limited sample conditions is of great significance.

[0004] Therefore, there is an urgent need for a small-sample identification method for spatial targets. Summary of the Invention

[0005] To address the issue of low accuracy in small sample scenarios in existing spatial target recognition methods, this invention provides a small sample recognition method for spatial targets.

[0006] In a first aspect, embodiments of the present invention provide a method for small-sample identification of spatial targets, comprising:

[0007] Obtain an initial training set and a test set; wherein the initial training set includes several training samples of spatial targets and a class label for each training sample, and the test set includes several test samples of spatial targets;

[0008] The initial training set is input into a pre-constructed neural network to perform initial training on the neural network, thereby obtaining an initial recognition model and deep dimensionality reduction features of the training samples;

[0009] For each test sample, the following steps are performed: input the current test sample into the initial recognition model, and determine whether the current test sample should be expanded into the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample.

[0010] If so, the initial recognition model is then trained a second time based on the recognition and judgment results of the current test sample;

[0011] If not, the next test sample is used as the new current training sample, and the process jumps to inputting the current test sample into the initial recognition model.

[0012] The final training set, target recognition model, and recognition results for each test sample are obtained.

[0013] Secondly, embodiments of the present invention also provide a small-sample identification device for space targets, comprising:

[0014] An acquisition unit is used to acquire an initial training set and a test set; wherein the initial training set includes several training samples of spatial targets and a category label for each training sample, and the test set includes several test samples of spatial targets.

[0015] The first training unit is used to input the initial training set into a pre-constructed neural network to perform initial training on the neural network, thereby obtaining an initial recognition model and deep dimensionality reduction features of the training samples.

[0016] An expansion unit is configured to perform the following for each test sample: input the current test sample into the initial recognition model, and determine whether the current test sample should be expanded into the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample.

[0017] The second training unit is used to perform secondary training on the initial recognition model based on the recognition and judgment results of the current test sample when the condition is met.

[0018] The jump unit is used to, if not, take the next test sample as the new current training sample and jump to execute the step of inputting the current test sample into the initial recognition model;

[0019] The output unit is used to obtain the final training set, the target recognition model, and the recognition result of each test sample.

[0020] Thirdly, embodiments of the present invention also provide a computing device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method described in any embodiment of this specification.

[0021] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the methods described in any embodiment of this specification.

[0022] This invention provides a method, apparatus, computing device, and storage medium for small-sample spatial target recognition. First, an initial training set with category labels and a test set without category labels are obtained. Then, the initial training set is input into a pre-constructed neural network for initial training, resulting in an initial recognition model and deep dimensionality reduction features of the training samples. Next, for each test sample, the following steps are performed: the current test sample is input into the initial recognition model, and based on clustering methods, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample, it is determined whether the current test sample should be expanded into the initial training set. If so, the initial recognition model is retrained based on the recognition result and judgment result of the current test sample; if not, the next test sample is selected, and this process continues until the final training set, target recognition model, and recognition result for each test sample are obtained. Therefore, this solution can effectively improve the recognition accuracy of spatial targets in small-sample situations. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a small-sample identification method for spatial targets provided in an embodiment of the present invention;

[0025] Figure 2 This is a simulation model diagram of a type of space target provided in an embodiment of the present invention;

[0026] Figure 3 This is a simulation model diagram of another type of space target provided in an embodiment of the present invention;

[0027] Figure 4 This is a simulation model diagram of another type of space target provided in an embodiment of the present invention;

[0028] Figure 5 This is a hardware architecture diagram of a computing device provided in an embodiment of the present invention;

[0029] Figure 6 This is a structural diagram of a small sample recognition device for spatial targets provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] As mentioned earlier, existing spatial target recognition methods often suffer from low recognition accuracy in small sample situations due to insufficient training samples.

[0032] To address the aforementioned technical problems, the inventors could consider combining neural networks and clustering methods to construct a deep clustering network, thereby expanding the initial training set using test samples to improve the accuracy of spatial target recognition in small sample situations.

[0033] The following describes the specific implementation of the above concept.

[0034] Please refer to Figure 1 This invention provides a method for small-sample identification of spatial targets, the method comprising:

[0035] Step 100: Obtain the initial training set and test set; wherein, the initial training set includes several training samples of spatial targets and the class label of each training sample, and the test set includes several test samples of spatial targets.

[0036] Step 102: Input the initial training set into the pre-built neural network to perform initial training on the neural network and obtain the initial recognition model and deep dimensionality reduction features of the training samples;

[0037] Step 104: For each test sample, perform the following: Input the current test sample into the initial recognition model, and determine whether the current test sample should be expanded into the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample.

[0038] Step 106: If yes, then the initial recognition model is trained a second time based on the recognition and judgment results of the current test sample.

[0039] Step 108: If not, then use the next test sample as the new current training sample and jump to the step of inputting the current test sample into the initial recognition model.

[0040] Step 110 yields the final training set, target recognition model, and recognition results for each test sample.

[0041] In this embodiment of the invention, an initial training set with category labels and a test set without category labels are first obtained. Then, the initial training set is input into a pre-constructed neural network for initial training, resulting in an initial recognition model and deep dimensionality reduction features of the training samples. Next, for each test sample, the following steps are performed: the current test sample is input into the initial recognition model, and based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample, it is determined whether the current test sample should be expanded into the initial training set. If so, the initial recognition model is retrained based on the recognition result and judgment result of the current test sample. If not, the next test sample is selected, and this process continues until the final training set, target recognition model, and recognition result for each test sample are obtained. Therefore, this scheme can effectively improve the recognition accuracy of spatial targets in small sample situations.

[0042] For step 100:

[0043] In electromagnetic scattering space target recognition technology, input data can be mainly divided into three categories: narrowband RCS sequence data, broadband HRRP data, and broadband two-dimensional ISAR imaging data. With the rapid development of radar sensors, space target situational awareness based on high-resolution imaging radar has become a mainstream development trend. Under the scattering point model, high-resolution one-dimensional range profiles (HRRP data) represent the amplitude of the sum of echoes from the target's scattering center within a relevant time interval. They contain very rich target range-direction structural information, allowing the extraction of parameters such as the number, location, intensity, and radial length of scattering centers, making them an important source of features for radar target recognition. Furthermore, compared to two-dimensional ISAR images, HRRP offers advantages in easier data acquisition and processing, thus demonstrating good feasibility and practicality in radar space target recognition.

[0044] Therefore, in this embodiment, both the training samples and the test samples are HRRP data of spatial targets. Furthermore, the training samples in the initial training set all have class labels, while the test samples in the test set do not. Thus, the number of training samples in the initial training set is relatively small, which is the small sample case described in this invention.

[0045] Regarding step 102:

[0046] In this embodiment of the invention, the pre-constructed neural network includes a three-layer sparse autoencoder network and a classifier. The recognition result is the output of the classifier, and the deep dimensionality reduction features are the output of the last layer of the sparse autoencoder network.

[0047] The training samples from the initial training set are input into the neural network. A three-layer sparse autoencoder network performs dimensionality reduction on the features of each training sample. The classifier classifies and identifies each training sample based on the deep dimensionality reduction features output by the last layer of the sparse autoencoder network. Based on the identification result and the class label of each training sample, an initial recognition model can be trained and generated. Therefore, when each training sample is input into the initial recognition model, the last layer of the sparse autoencoder network outputs the deep dimensionality reduction features of each training sample.

[0048] Regarding step 104:

[0049] In this embodiment of the invention, test samples from the test set are sequentially input into the initial recognition model. Then, for each input test sample, it is determined whether the current test sample can be expanded into the initial training set. If it can, the current test sample is used to train the initial recognition model. When the next test sample is input, the determination is made again, and so on in a loop.

[0050] It is understandable that when the current test sample is input into the initial recognition model, the deep dimensionality reduction features of the current test sample can be obtained from the output of the last layer of the sparse autoencoder network.

[0051] In this embodiment of the invention, the step "determining whether the current test sample is expanded to the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample" may include the following steps S1-S3:

[0052] Step S1: Based on the clustering method, cluster the deep dimensionality reduction features of the training samples to obtain the location of the center point of each category and the average distance of each category; wherein, the average distance is the average distance of all training samples in the category to the center point of the category.

[0053] In this embodiment of the invention, the k-means algorithm is used to cluster the deep dimensionality reduction features of the training samples in the initial training set, which can obtain the position C(k) of the center point of each category and the average distance AvS(k) of all training samples in each category to the center point of that category, where k is the category, k = 1, ... K, and K is the total number of categories.

[0054] Specifically, the distance from each training sample to the center point of that category is calculated using the following formula:

[0055]

[0056] In the formula, X is the deep dimensionality reduction feature of the training sample, N is the data point dimension of the deep dimensionality reduction feature of the training sample, X(n) is each data point of the training sample, and C is the center point position of the corresponding category of the training sample.

[0057] As can be seen from the above formula, the distance R(X,C) from each training sample to the center point of its class can be calculated.

[0058] Then, by averaging the distances of all training samples in each category, we can obtain the average distance AvS(k) for each category.

[0059] Step S2: Based on the deep dimensionality reduction features of the current test sample, calculate the distance from the current test sample to the center point of each category.

[0060] In this embodiment, the distance R(X,C) from the training sample to the center point of its class is calculated in the same way as in step S1. The distance R from the current test sample to the center point of each class can be calculated. C (i,k), where i is the test sample identifier and k is the category.

[0061] Step S3: Based on the average distance of each category and the distance of the current test sample to the center point of each category, determine whether the current test sample should be expanded to the initial training set.

[0062] In this embodiment of the invention, step S3 may include the following steps H1-H3:

[0063] Step H1: Compare the distances of the current test sample to the center point of each category, and determine the category corresponding to the smallest distance as the target category of the current test sample;

[0064] Step H2: Determine the distance threshold for the target category based on the average distance corresponding to the target category;

[0065] Step H3: When the distance from the current test sample to the center point of the target category is less than or equal to the distance threshold, generate the target label of the current test sample based on the target category, and expand the current test sample and the target label into the initial training set.

[0066] In this embodiment, the distance threshold is determined by the following formula:

[0067] Th(k) = a*AvS(k)

[0068] In the formula, k is the category, Th is the distance threshold, AvS is the average distance, and a is the threshold coefficient.

[0069] Based on the above formula for calculating the distance threshold, the distance threshold for each category can be calculated based on the average distance of each category. The threshold coefficient a∈[0,1]. Since the average distance of each category is different, the distance threshold for different categories will also be different.

[0070] Then, you can determine whether the current test sample has been expanded to the initial training set in the following way:

[0071]

[0072] When the minimum distance from the current test sample to the centroid of each category (i.e., the distance from the current test sample to the centroid of the target category) is less than or equal to the distance threshold of the target category, a target label is generated for the current test sample based on the target category. The current test sample and the target label are then added to the initial training set. In this case, the judgment result for the current test sample is its target label. When the minimum distance from the current test sample to the centroid of each category is greater than the distance threshold of the target category, the current test sample cannot be added to the initial training set, and no label is generated for the current test sample.

[0073] It should be noted that the threshold coefficient setting affects the final target recognition model, and further affects the model's accuracy in recognizing test samples. If the threshold coefficient is set too high, too many misclassified test samples will be introduced into the training set; conversely, if the threshold coefficient is set too low, the training set will be too small, resulting in low accuracy due to insufficient training samples. Therefore, in this embodiment, through multiple experiments, the threshold coefficient that yields the highest recognition accuracy is determined as the final target threshold coefficient to ensure the accuracy of the target recognition model.

[0074] Regarding step 106:

[0075] When the current test sample is expanded to the initial training set, the initial recognition model can be retrained based on the recognition result of the current test sample and the target label of the current test sample determined in step H3. Then, the next test sample is obtained as the new current training sample, and the process jumps to step 104 to "input the current test sample into the initial recognition model".

[0076] It is understandable that if the previous test sample was expanded into the initial training set, then when determining whether the next test sample should be expanded into the initial training set, the previous test sample has already become a training sample. Therefore, the determination of the next test sample will take into account the test sample that has been expanded into the initial training set.

[0077] Regarding steps 108 and 110:

[0078] If the current test sample is not expanded into the initial training set, then the next test sample is obtained as the new current training sample, and the process jumps to step 104 to "input the current test sample into the initial recognition model".

[0079] Thus, after judging each test sample in the test set, and using each test sample expanded to the training set to perform secondary training on the initial recognition model, the final training set and target recognition model can be obtained. Then, by inputting each test sample in the test set into the target recognition model, the recognition result of each test sample can be obtained.

[0080] In this embodiment of the invention, the broadband radar features of three typical space targets are identified using the scheme of the present invention. The simulation models of the three typical space targets are as follows: Figure 2 , Figure 3 and Figure 4 As shown.

[0081] The simulation conditions for broadband data of three types of space targets are: center frequency 11 GHz; bandwidth 2 GHz; frequency interval 5 MHz; HH polarization. The azimuth angle is -10 to 190° with an angle interval of 0.02°, and the elevation angle is 30°. After HRRP imaging processing, HRRP history maps of the three types of space targets can be obtained. Each target dataset consists of 10,000 HRRP sets, each with a dimension of 1024, for a total of 30,000 HRRP data sets across the three target types.

[0082] To make the simulation more realistic, 10dB of Gaussian white noise was added to the HRRP data.

[0083] Four classification and recognition experiments were conducted with different threshold coefficient settings. 163 sets of data were randomly selected as training samples, accounting for approximately 0.54% of the total samples, while the remaining 99.46% of the HRRP data were used as test samples. The recognition results of the traditional Sparse Autoencoder (SAE) network and the deep clustering network proposed in this scheme are compared in Table 1 below:

[0084] Table 1

[0085]

[0086] As can be seen from Table 1: (1) When the threshold coefficient a = 0.1, this scheme has the best recognition accuracy, which can be improved by about 3.5% compared with the traditional SAE network; (2) When the threshold coefficient a = 0.1, this scheme can effectively expand the number of training samples, with an increment of 328%, which effectively improves the recognition accuracy of the target recognition model; (3) The threshold coefficient is not set as large as possible. Although a larger value of a can increase the number of training samples, setting the threshold coefficient too large will also introduce too many incorrect label samples, which will mislead the training network and affect the recognition accuracy.

[0087] like Figure 5 , Figure 6As shown, this embodiment of the invention provides a small-sample recognition device for spatial targets. The device embodiment can be implemented through software, hardware, or a combination of both. From a hardware perspective, as... Figure 5 The diagram shown is a hardware architecture diagram of a computing device housing a small-sample recognition device for spatial targets provided in an embodiment of the present invention. (Except for...) Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, the computing device in the embodiment may also include other hardware, such as a forwarding chip responsible for processing packets. Taking software implementation as an example, such as... Figure 6 As shown, a device in a logical sense is formed by the CPU of its computing device reading the corresponding computer program from non-volatile memory into memory and running it. This embodiment provides a small-sample recognition device for spatial targets, comprising:

[0088] The acquisition unit 601 is used to acquire an initial training set and a test set; wherein, the initial training set includes several training samples of spatial targets and the category label of each training sample, and the test set includes several test samples of spatial targets.

[0089] The first training unit 602 is used to input the initial training set into the pre-built neural network to perform initial training on the neural network and obtain the initial recognition model and deep dimensionality reduction features of the training samples.

[0090] The expansion unit 603 is used to perform the following for each test sample: input the current test sample into the initial recognition model, and determine whether the current test sample should be expanded into the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples and the deep dimensionality reduction features of the current test sample.

[0091] The second training unit 604 is used to perform secondary training on the initial recognition model based on the recognition and judgment results of the current test sample when the condition is met.

[0092] Jump unit 605 is used to, if not, take the next test sample as the new current training sample and jump to execute the input of the current test sample into the initial recognition model;

[0093] Output unit 606 is used to obtain the final training set, target recognition model and recognition results for each test sample.

[0094] In one embodiment of the present invention, the neural network in the first training unit 602 includes a three-layer sparse autoencoder network and a classifier, the recognition result is the output of the classifier, and the deep dimensionality reduction features are the output of the last layer of the sparse autoencoder network.

[0095] In one embodiment of the present invention, the expansion unit 603 is used to perform:

[0096] Based on the clustering method, the deep dimensionality reduction features of the training samples are clustered to obtain the location of the centroid of each category and the average distance of each category; where the average distance is the average distance of all training samples in that category to the centroid of that category;

[0097] Based on the deep dimensionality reduction features of the current test sample, calculate the distance from the current test sample to the center point of each category;

[0098] Based on the average distance of each category and the distance of the current test sample to the center point of each category, determine whether the current test sample should be expanded into the initial training set.

[0099] In one embodiment of the present invention, the distance from the training sample in the augmentation unit 603 to the center point of the category is calculated using the following formula:

[0100]

[0101] In the formula, X is the deep dimensionality reduction feature of the training sample, N is the data point dimension of the deep dimensionality reduction feature of the training sample, X(n) is each data point of the training sample, and C is the center point position of the corresponding category of the training sample.

[0102] In one embodiment of the present invention, when the expansion unit 603 determines whether the current test sample should be expanded to the initial training set based on the average distance of each category and the distance from the current test sample to the center point of each category, it performs the following:

[0103] Compare the distances from the current test sample to the center point of each category, and determine the category corresponding to the smallest distance as the target category of the current test sample;

[0104] Based on the average distance corresponding to the target category, determine the distance threshold for the target category;

[0105] When the distance from the current test sample to the center point of the target category is less than or equal to the distance threshold, the target label of the current test sample is generated based on the target category, and the current test sample and the target label are expanded to the initial training set.

[0106] In one embodiment of the present invention, the distance threshold in the expansion unit 603 is determined by the following formula:

[0107] Th(k) = a*AvS(k)

[0108] In the formula, k is the category, Th is the distance threshold, AvS is the average distance, and a is the threshold coefficient.

[0109] In one embodiment of the present invention, the judgment result of the current test sample in the second training unit 604 is the target label of the current test sample.

[0110] It is understood that the structures illustrated in the embodiments of the present invention do not constitute a specific limitation on a small sample identification device for space targets. In other embodiments of the present invention, a small sample identification device for space targets may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0111] The information interaction and execution process between the modules in the above-mentioned device are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description of the method embodiment of the present invention, and will not be repeated here.

[0112] This invention also provides a computing device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements a small-sample recognition method for spatial targets according to any embodiment of this invention.

[0113] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform a small-sample recognition method for spatial targets according to any embodiment of this invention.

[0114] Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage medium.

[0115] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0116] Examples of storage media used to provide program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0117] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0118] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion module connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion module execute some and all of the actual operations, thereby realizing the function of any of the above embodiments.

[0119] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0120] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying spatial targets using small samples, characterized in that, include: Obtain an initial training set and a test set; wherein the initial training set includes several training samples of spatial targets and a class label for each training sample, and the test set includes several test samples of spatial targets; The initial training set is input into a pre-constructed neural network to perform initial training on the neural network, thereby obtaining an initial recognition model and deep dimensionality reduction features of the training samples; For each test sample, the following steps are performed: input the current test sample into the initial recognition model, and determine whether the current test sample should be expanded into the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample; If so, the initial recognition model is then trained a second time based on the recognition and judgment results of the current test sample; If not, the next test sample will be used as the new current test sample, and the process will jump to inputting the current test sample into the initial recognition model. The final training set, target recognition model, and recognition results for each test sample are obtained. The step of determining whether the current test sample should be expanded to the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample includes: Based on the clustering method, the deep dimensionality reduction features of the training samples are clustered to obtain the position of the center point of each category and the average distance of each category; wherein, the average distance is the average distance of all training samples in that category to the center point of that category; Based on the deep dimensionality reduction features of the current test sample, calculate the distance from the current test sample to the center point of each category; Based on the average distance of each category and the distance of the current test sample to the center point of each category, determine whether the current test sample should be expanded to the initial training set; The step of determining whether the current test sample should be expanded into the initial training set based on the average distance of each category and the distance from the current test sample to the center point of each category includes: Compare the distances from the current test sample to the center point of each category, and determine the category corresponding to the smallest distance as the target category of the current test sample; Based on the average distance corresponding to the target category, a distance threshold for the target category is determined; When the distance from the current test sample to the center point of the target category is less than or equal to the distance threshold, a target label for the current test sample is generated based on the target category, and the current test sample and the target label are expanded to the initial training set.

2. The method according to claim 1, characterized in that, The neural network comprises a three-layer sparse autoencoder network and a classifier. The recognition result is the output of the classifier, and the deep dimensionality reduction feature is the output of the last layer of the sparse autoencoder network.

3. The method according to claim 1, characterized in that, The distance from the training sample to the center point of the category is calculated using the following formula: In the formula, The deep dimensionality reduction features of the training samples, The data point dimension of the deep dimensionality reduction features of the training samples. For each data point of the training samples, This represents the center point location of the corresponding category for the training samples.

4. The method according to claim 1, characterized in that, The distance threshold is determined by the following formula: In the formula, As a category, Distance threshold The average distance, This is the threshold coefficient.

5. The method according to claim 1, characterized in that, The judgment result of the current test sample is the target label of the current test sample.

6. A small-sample identification device for spatial targets, used to perform the method as described in any one of claims 1-5, characterized in that, include: An acquisition unit is used to acquire an initial training set and a test set; wherein the initial training set includes several training samples of spatial targets and a category label for each training sample, and the test set includes several test samples of spatial targets. The first training unit is used to input the initial training set into a pre-constructed neural network to perform initial training on the neural network, thereby obtaining an initial recognition model and deep dimensionality reduction features of the training samples. An expansion unit is configured to perform the following for each test sample: input the current test sample into the initial recognition model, and determine whether the current test sample should be expanded into the initial training set based on the clustering method, the deep dimensionality reduction features of the training samples, and the deep dimensionality reduction features of the current test sample. The second training unit is used to perform secondary training on the initial recognition model based on the recognition and judgment results of the current test sample when the condition is met. The jump unit is used to, if not, take the next test sample as the new current test sample and jump to execute the step of inputting the current test sample into the initial recognition model; The output unit is used to obtain the final training set, the target recognition model, and the recognition result of each test sample.

7. A computing device comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, implements the method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-5.

Images