SAR (Synthetic Aperture Radar) target identification method and device based on projection features and improved self-learning

By using projection features and improved self-learning methods, and leveraging the SVM classifier and the local geometric properties of unlabeled samples, the problem of deep learning's dependence on labeled samples is solved, thereby improving the performance and accuracy of SAR target recognition.

CN121330381APending Publication Date: 2026-01-13LEIHUA ELECTRONICS TECH RES INST AVIATION IND OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511528545.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-24
Publication Date
2026-01-13

AI Technical Summary

Technical Problem

Deep learning requires a large number of labeled samples for SAR target recognition. Manual labeling is inefficient and inaccurate, resulting in a small number of samples. Furthermore, unlabeled samples are not fully utilized, which limits the improvement of recognition performance.

Method used

We employ a projection feature-based and improved self-learning approach. By using ensemble learning, we project samples onto a new prototype space and utilize classification posterior probability and local geometric properties, combined with an SVM classifier, to extract valuable information from unlabeled samples and optimize the classifier.

Benefits of technology

It effectively improves the performance of SAR target recognition with a small number of labeled samples, makes full use of unlabeled samples, and improves classification confidence and recognition accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121330381A_ABST
    Figure CN121330381A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of radar automatic target recognition, and relates to an SAR target recognition method and device based on projection features and improved self-learning. The method comprises the following steps: respectively processing each sample parameter of a training sample by a classifier T respectively trained by each prototype set to obtain integrated projection features of the sample parameters; initializing an SVM classifier based on the labeled sample feature set; taking out part of samples from the unlabeled sample feature set, and constructing a temporary sample pool; classifying the samples in the temporary sample pool by using an SVM classifier; evaluating the classification confidence of the samples in the temporary sample pool by using the classification posterior probability; adding a plurality of samples with high confidence into the labeled sample feature set, and retraining the SVM classifier; repeating for multiple times until an optimal SVM classifier is obtained; and target identification is carried out based on the optimal SVM classifier. The target recognition performance under the condition of a small number of labeled samples and a large number of unlabeled samples is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of automatic radar target recognition, specifically relating to a SAR target recognition method and apparatus based on projection features and improved self-learning. Background Technology

[0002] Due to its ability to acquire high-resolution images regardless of weather and lighting conditions, SAR (Synthetic Aperture Radar) is widely used in various target recognition fields, such as urban planning, natural resource monitoring, and reconnaissance. Among these, SAR automatic target recognition aims to effectively acquire and identify ground vehicle targets from complex terrain scenes. It can quickly and accurately complete specific target recognition tasks using computers, freeing humans from sifting through massive amounts of data. Therefore, since its inception in 1993, it has received widespread attention from researchers both domestically and internationally.

[0003] Deep neural networks have been applied in SAR target recognition and have achieved great breakthroughs. However, the following problems exist in practical applications:

[0004] Question 1: The excellent recognition performance of deep learning requires the support of a large dataset. Due to the special nature and complexity of SAR images, they usually need to be labeled and analyzed by professionals. However, in complex environments, the efficiency and accuracy of manual labeling are relatively low, resulting in very few samples that can be used for training. This severely restricts the improvement of SAR target recognition performance and reduces the functionality of radar.

[0005] Question 2: A massive amount of data has been accumulated during the research, development, testing, and use of equipment. This data lacks annotation by professionals and consists of unlabeled samples. However, deep learning only utilizes labeled samples, failing to fully leverage the large number of unlabeled samples, resulting in a significant waste of data. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a SAR target recognition method and apparatus based on projection features and improved self-learning.

[0007] The first aspect of this application provides a SAR target recognition method based on projection features and improved self-learning, mainly including:

[0008] Step S1: Extract multiple prototype sets from the training sample set. The classifier T trained by each prototype set processes each sample parameter of the training sample to obtain the integrated projection features of the sample parameters. The training sample set consists of a labeled sample feature set and an unlabeled sample feature set.

[0009] Step S2: Initialize the SVM classifier F based on the labeled sample feature set;

[0010] Step S3: Extract a portion of samples from the unlabeled sample feature set to construct a temporary sample pool;

[0011] Step S4: Use SVM classifier F to classify the samples in the temporary sample pool;

[0012] Step S5: Evaluate the classification confidence of samples in the temporary sample pool using the posterior probability of classification.

[0013] Step S6: Add multiple samples with high confidence to the labeled sample feature set and retrain the SVM classifier F;

[0014] Step S7: Repeat steps S3-S6 multiple times to obtain the optimal SVM classifier F;

[0015] Step S8: Perform target recognition based on the optimal SVM classifier F.

[0016] Preferably, in step S1, extracting multiple prototype sets from the training sample set includes:

[0017] Step S11: Randomly select multiple samples from the training sample set to form set V, and calculate the total error of set V;

[0018] Step S12: When the total error exceeds the error setting value, the total error is used as the new error setting value, and set V is used as the new skeleton set, where the initial error setting value is 0 and the initial skeleton set is an empty set.

[0019] Step S13: Repeat the above steps multiple times to obtain the final skeleton set;

[0020] Step S14: For each sample in the skeleton set, select multiple nearest neighbor samples from the training set and add them to the skeleton set to form a prototype set;

[0021] Step S15: Repeat the above steps multiple times to form multiple prototype sets.

[0022] Preferably, in step S11, the total error is the sum of the Euclidean distances between any two samples in set V.

[0023] Preferably, in step S3, after extracting a portion of the samples from the unlabeled sample feature set, the extracted samples are deleted from the unlabeled sample feature set.

[0024] Preferably, after step S7, the method further includes:

[0025] Step S71: Obtain the integrated projection features of the test sample;

[0026] Step S72: Classify the ensemble projection features of the test samples using the optimal SVM classifier F. Determine the classification ability of the optimal SVM classifier F based on the category to which the output test sample belongs and the actual category of the test sample.

[0027] The second aspect of this application provides a SAR target recognition device based on projection features and improved self-learning, mainly comprising:

[0028] An integrated projection feature acquisition module is used to extract multiple prototype sets from the training sample set. The classifier T trained by each prototype set processes each sample parameter of the training sample to obtain the integrated projection features of the sample parameters. The training sample set consists of a labeled sample feature set and an unlabeled sample feature set.

[0029] The SVM classifier initialization module is used to initialize the SVM classifier F based on the labeled sample feature set.

[0030] The temporary sample pool construction module is used to extract a portion of samples from the unlabeled sample feature set and construct a temporary sample pool.

[0031] The sample classification module is used to classify samples in a temporary sample pool using an SVM classifier F.

[0032] The classification confidence assessment module is used to assess the classification confidence of samples in the temporary sample pool using the posterior probability of classification.

[0033] The SVM classifier retraining module is used to add multiple samples with high confidence to the labeled sample feature set and retrain the SVM classifier F.

[0034] The optimal SVM classifier F determination module is used to repeatedly call the temporary sample pool construction module, sample classification module, classification confidence evaluation module, and classification confidence evaluation module to obtain the optimal SVM classifier F.

[0035] The target recognition module is used to perform target recognition based on the optimal SVM classifier F.

[0036] Preferably, the integrated projection feature acquisition module includes:

[0037] The total error calculation unit is used to randomly select multiple samples from the training sample set to form a set V, and calculate the total error of set V.

[0038] The skeleton set update unit is used to use the total error as the new error setting value and the set V as the new skeleton set when the total error exceeds the error setting value, wherein the initial error setting value is 0 and the initial skeleton set is an empty set.

[0039] The skeleton set determination unit is used to repeatedly call the total error calculation unit and the skeleton set update unit to obtain the final skeleton set;

[0040] The prototype set update unit is used to select multiple nearest neighbor samples from the training set for each sample in the skeleton set and add them to the skeleton set to form a prototype set.

[0041] The prototype set determination unit is used to repeatedly call the skeleton set determination unit and the prototype set update unit to form multiple prototype sets.

[0042] Preferably, the total error is the sum of the Euclidean distances between any two samples in set V.

[0043] Preferably, in the temporary sample pool construction module, after taking some samples from the unlabeled sample feature set, the taken samples are deleted from the unlabeled sample feature set.

[0044] Preferably, the device further includes:

[0045] The test sample integrated projection feature acquisition module is used to obtain the integrated projection features of the test sample;

[0046] The classification ability determination module is used to classify the ensemble projection features of the test samples using the optimal SVM classifier F, and to determine the classification ability of the optimal SVM classifier F based on the category to which the output test sample belongs and the actual category of the test sample.

[0047] The advantages of this application include:

[0048] On the one hand, this application makes full use of the intra-class compactness and inter-class separability of training samples, and draws on the idea of ​​ensemble learning to project the samples into a new prototype space to obtain new feature representations of the samples; on the other hand, this application improves the traditional self-learning algorithm by using classification posterior probability and local geometric properties to enhance the classification confidence of unlabeled samples.

[0049] This application effectively improves the SAR target recognition performance under conditions of a small number of labeled samples and a large number of unlabeled samples by extracting highly separable features and improving the self-learning algorithm. Attached Figure Description

[0050] Figure 1 This is a flowchart of a preferred embodiment of the SAR target recognition method based on projection features and improved self-learning in this application.

[0051] Figure 2 This is a schematic diagram of integrated projection feature extraction. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are only some, not all, of the embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0053] The first aspect of this application provides a SAR target recognition method based on projection features and improved self-learning, such as... Figure 1 As shown, it mainly includes:

[0054] Step S1: Extract multiple prototype sets from the training sample set. The classifier T trained by each prototype set processes each sample parameter of the training sample to obtain the integrated projection features of the sample parameters. The training sample set consists of a labeled sample feature set and an unlabeled sample feature set.

[0055] Step S2: Initialize the SVM classifier F based on the labeled sample feature set;

[0056] Step S3: Extract a portion of samples from the unlabeled sample feature set to construct a temporary sample pool;

[0057] Step S4: Use SVM classifier F to classify the samples in the temporary sample pool;

[0058] Step S5: Evaluate the classification confidence of samples in the temporary sample pool using the posterior probability of classification.

[0059] Step S6: Add multiple samples with high confidence to the labeled sample feature set and retrain the SVM classifier F;

[0060] Step S7: Repeat steps S3-S6 multiple times to obtain the optimal SVM classifier F;

[0061] Step S8: Perform target recognition based on the optimal SVM classifier F.

[0062] This application first extracts the integrated projection features of each sample in the training set in step S1.

[0063] In some alternative implementations, such as Figure 2 As shown, in step S1, extracting multiple prototype sets from the training sample set includes:

[0064] Step S11: Randomly select multiple samples from the training sample set to form set V, and calculate the total error of set V;

[0065] Step S12: When the total error exceeds the error setting value, the total error is used as the new error setting value, and set V is used as the new skeleton set, where the initial error setting value is 0 and the initial skeleton set is an empty set.

[0066] Step S13: Repeat the above steps multiple times to obtain the final skeleton set;

[0067] Step S14: For each sample in the skeleton set, select multiple nearest neighbor samples from the training set and add them to the skeleton set to form a prototype set;

[0068] Step S15: Repeat the above steps multiple times to form multiple prototype sets.

[0069] In some alternative implementations, in step S11, the total error is the sum of the Euclidean distances between any two samples in set V.

[0070] In this embodiment, the input training sample set ,in, This indicates that there is a labeled sample set. Indicates the first One labeled sample, Indicates the first The labeling of each sample Indicates the category of the sample. This represents the number of labeled samples. Number of target categories; This represents the unlabeled sample set. Indicates the first One unlabeled sample, This represents the number of unlabeled samples.

[0071] Next, extract each sample from the training set. The integrated projection features specifically include:

[0072] Initial distance and skeleton set ;

[0073] Randomly selected from the training set The samples constitute a set Calculate the set Total error :

[0074]

[0075] in, Indicates sample and Euclidean distance;

[0076] like Then let , ;

[0077] Repeat the above steps Next, will be obtained As the final skeleton sample;

[0078] For skeleton samples For each sample in the dataset, select the nearest neighbor from the training set that is the skeleton sample. The prototype set consists of one sample, the skeleton sample set, and the selected samples. ;

[0079] Repeat the above steps Next, get prototype set ;

[0080] In each prototype set Training the classifier To obtain the classifier set ;

[0081] For each training sample and Projecting onto each classifier , to obtain the projection vector The projected vectors from all classifiers are concatenated to obtain the sample. Integrated projection features And obtain samples Integrated projection features ;

[0082] For the training set For each sample, ensemble projection features are extracted according to the steps described above to form a training sample feature set. ,in , .

[0083] Then, in step S2, the labeled sample feature set is used. Initialize the SVM classifier .

[0084] Next, in step S3, a temporary sample pool is constructed.

[0085] In some alternative implementations, in step S3, after taking a portion of the samples from the unlabeled sample feature set, the taken samples are deleted from the unlabeled sample feature set.

[0086] In this embodiment, from the unlabeled sample feature set A temporary sample pool is formed by randomly selecting samples from the middle. And these samples were extracted from the unlabeled sample feature set. Removed from the middle.

[0087] Then, in step S4, the classifier is used. Temporary sample pool The samples in the dataset are classified, and their posterior probabilities are obtained. In step S5, the temporary sample pool is optimized using the posterior probability of classification and the local geometric information of the samples. Classification confidence of medium samples Conduct an assessment

[0088] .

[0089] In the formula, Indicates sample Among the nearest neighbor samples, they are divided into the first... The number of samples in the class.

[0090] Then, in step S6, from Select the category with the highest confidence level. Each sample is added to the labeled sample feature set. In the middle; utilizing the expanded labeled sample feature set Retrain the classifier .

[0091] Then, in step S7, the above steps are repeated until the number of iterations is reached. The final classifier The optimal classifier.

[0092] In some alternative implementations, after step S7, the following is further included:

[0093] Step S71: Obtain the integrated projection features of the test sample;

[0094] Step S72: Classify the ensemble projection features of the test samples using the optimal SVM classifier F. Determine the classification ability of the optimal SVM classifier F based on the category to which the output test sample belongs and the actual category of the test sample.

[0095] In this embodiment, features are extracted from the test samples using the same method to obtain their integrated projection features; and the optimal classifier obtained in step S7 is used. The integrated projection features of the test samples are classified to obtain the category to which the test samples belong.

[0096] Finally, in step S8, target recognition is performed using the optimal SVM classifier F.

[0097] A second aspect of this application provides a SAR target recognition device based on projection features and improved self-learning, corresponding to the above-described method, mainly comprising:

[0098] An integrated projection feature acquisition module is used to extract multiple prototype sets from the training sample set. The classifier T trained by each prototype set processes each sample parameter of the training sample to obtain the integrated projection features of the sample parameters. The training sample set consists of a labeled sample feature set and an unlabeled sample feature set.

[0099] The SVM classifier initialization module is used to initialize the SVM classifier F based on the labeled sample feature set.

[0100] The temporary sample pool construction module is used to extract a portion of samples from the unlabeled sample feature set and construct a temporary sample pool.

[0101] The sample classification module is used to classify samples in a temporary sample pool using an SVM classifier F.

[0102] The classification confidence assessment module is used to assess the classification confidence of samples in the temporary sample pool using the posterior probability of classification.

[0103] The SVM classifier retraining module is used to add multiple samples with high confidence to the labeled sample feature set and retrain the SVM classifier F.

[0104] The optimal SVM classifier F determination module is used to repeatedly call the temporary sample pool construction module, sample classification module, classification confidence evaluation module, and classification confidence evaluation module to obtain the optimal SVM classifier F.

[0105] The target recognition module is used to perform target recognition based on the optimal SVM classifier F.

[0106] In some alternative implementations, the integrated projection feature acquisition module includes:

[0107] The total error calculation unit is used to randomly select multiple samples from the training sample set to form a set V, and calculate the total error of set V.

[0108] The skeleton set update unit is used to use the total error as the new error setting value and the set V as the new skeleton set when the total error exceeds the error setting value, wherein the initial error setting value is 0 and the initial skeleton set is an empty set.

[0109] The skeleton set determination unit is used to repeatedly call the total error calculation unit and the skeleton set update unit to obtain the final skeleton set;

[0110] The prototype set update unit is used to select multiple nearest neighbor samples from the training set for each sample in the skeleton set and add them to the skeleton set to form a prototype set.

[0111] The prototype set determination unit is used to repeatedly call the skeleton set determination unit and the prototype set update unit to form multiple prototype sets.

[0112] In some alternative implementations, the total error is the sum of the Euclidean distances between any two samples in set V.

[0113] In some alternative implementations, in the temporary sample pool construction module, after taking a portion of the samples from the unlabeled sample feature set, the taken samples are deleted from the unlabeled sample feature set.

[0114] In some alternative embodiments, the apparatus further includes:

[0115] The test sample integrated projection feature acquisition module is used to obtain the integrated projection features of the test sample;

[0116] The classification ability determination module is used to classify the ensemble projection features of the test samples using the optimal SVM classifier F, and to determine the classification ability of the optimal SVM classifier F based on the category to which the output test sample belongs and the actual category of the test sample.

[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A SAR target recognition method based on projection features and improved self-learning, characterized in that, include: Step S1: Extract multiple prototype sets from the training sample set. The classifier T trained by each prototype set processes each sample parameter of the training sample to obtain the integrated projection features of the sample parameters. The training sample set consists of a labeled sample feature set and an unlabeled sample feature set. Step S2: Initialize the SVM classifier F based on the labeled sample feature set; Step S3: Extract a portion of samples from the unlabeled sample feature set to construct a temporary sample pool; Step S4: Use SVM classifier F to classify the samples in the temporary sample pool; Step S5: Evaluate the classification confidence of samples in the temporary sample pool using the posterior probability of classification; Step S6: Add multiple samples with high confidence to the labeled sample feature set and retrain the SVM classifier F; Step S7: Repeat steps S3-S6 multiple times to obtain the optimal SVM classifier F; Step S8: Perform target recognition based on the optimal SVM classifier F.

2. The SAR target recognition method based on projection features and improved self-learning as described in claim 1, characterized in that, In step S1, extracting multiple prototype sets from the training sample set includes: Step S11: Randomly select multiple samples from the training sample set to form set V, and calculate the total error of set V; Step S12: When the total error exceeds the error setting value, the total error is used as the new error setting value, and set V is used as the new skeleton set, where the initial error setting value is 0 and the initial skeleton set is an empty set. Step S13: Repeat the above steps multiple times to obtain the final skeleton set; Step S14: For each sample in the skeleton set, select multiple nearest neighbor samples from the training set and add them to the skeleton set to form a prototype set; Step S15: Repeat the above steps multiple times to form multiple prototype sets.

3. The SAR target recognition method based on projection features and improved self-learning as described in claim 2, characterized in that, In step S11, the total error is the sum of the Euclidean distances between any two samples in set V.

4. The SAR target recognition method based on projection features and improved self-learning as described in claim 1, characterized in that, In step S3, after extracting some samples from the unlabeled sample feature set, the extracted samples are deleted from the unlabeled sample feature set.

5. The SAR target recognition method based on projection features and improved self-learning as described in claim 1, characterized in that, Following step S7, the following further includes: Step S71: Obtain the integrated projection features of the test sample; Step S72: Classify the ensemble projection features of the test samples using the optimal SVM classifier F. Determine the classification ability of the optimal SVM classifier F based on the category to which the output test sample belongs and the actual category of the test sample.

6. A SAR target recognition device based on projection features and improved self-learning, characterized in that, include: An integrated projection feature acquisition module is used to extract multiple prototype sets from the training sample set. The classifier T trained by each prototype set processes each sample parameter of the training sample to obtain the integrated projection features of the sample parameters. The training sample set consists of a labeled sample feature set and an unlabeled sample feature set. The SVM classifier initialization module is used to initialize the SVM classifier F based on the labeled sample feature set. The temporary sample pool construction module is used to extract a portion of samples from the unlabeled sample feature set and construct a temporary sample pool. The sample classification module is used to classify samples in a temporary sample pool using an SVM classifier F. The classification confidence assessment module is used to assess the classification confidence of samples in the temporary sample pool using the posterior probability of classification. The SVM classifier retraining module is used to add multiple samples with high confidence to the labeled sample feature set and retrain the SVM classifier F. The optimal SVM classifier F determination module is used to repeatedly call the temporary sample pool construction module, sample classification module, classification confidence evaluation module, and classification confidence evaluation module to obtain the optimal SVM classifier F. The target recognition module is used to perform target recognition based on the optimal SVM classifier F.

7. The SAR target recognition method based on projection features and improved self-learning as described in claim 6, characterized in that, The integrated projection feature acquisition module includes: The total error calculation unit is used to randomly select multiple samples from the training sample set to form a set V, and calculate the total error of set V. The skeleton set update unit is used to use the total error as the new error setting value and the set V as the new skeleton set when the total error exceeds the error setting value, wherein the initial error setting value is 0 and the initial skeleton set is an empty set. The skeleton set determination unit is used to repeatedly call the total error calculation unit and the skeleton set update unit to obtain the final skeleton set; The prototype set update unit is used to select multiple nearest neighbor samples from the training set for each sample in the skeleton set and add them to the skeleton set to form a prototype set. The prototype set determination unit is used to repeatedly call the skeleton set determination unit and the prototype set update unit to form multiple prototype sets.

8. The SAR target recognition method based on projection features and improved self-learning as described in claim 7, characterized in that, The total error is the sum of the Euclidean distances between any two samples in set V.

9. The SAR target recognition method based on projection features and improved self-learning as described in claim 6, characterized in that, In the temporary sample pool construction module, after taking some samples from the unlabeled sample feature set, the taken samples are deleted from the unlabeled sample feature set.

10. The SAR target recognition method based on projection features and improved self-learning as described in claim 6, characterized in that, The device further includes: The test sample integrated projection feature acquisition module is used to obtain the integrated projection features of the test sample; The classification ability determination module is used to classify the ensemble projection features of the test samples using the optimal SVM classifier F, and to determine the classification ability of the optimal SVM classifier F based on the category to which the output test sample belongs and the actual category of the test sample.