A semi-supervised hrrp noise label filtering method based on early learning guidance
The semi-supervised HRRP noise label filtering method guided by early learning, utilizing the generalized cross-entropy loss function and data augmentation techniques, solves the performance degradation problem caused by noise labels in radar target recognition, achieving higher recognition accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2024-03-29
- Publication Date
- 2026-06-23
AI Technical Summary
Existing HRRP-based radar target recognition methods suffer from drastic performance degradation in real-world environments due to noise tags, making them unsuitable for handling target recognition problems in complex environments.
We employ a semi-supervised HRRP noise label filtering method based on early learning guidance. We use the generalized cross-entropy loss function for pre-training and combine data augmentation techniques with a semi-supervised learning framework to reduce the impact of noise labels through data partitioning and label reshaping.
It improves the stability and robustness of radar target recognition, reduces the impact of noise labels on the model, enhances recognition performance and generalization ability, and simplifies the implementation process.
Smart Images
Figure CN118395295B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar technology, specifically relating to a semi-supervised HRRP noise label filtering method based on early learning guidance. Background Technology
[0002] Radar target recognition refers to determining the type of a target using its radar echo signals. A High Resolution Range Profile (HRRP) is the vector sum of the projections of the target's scattered point echoes onto the radar line of sight, obtained through a broadband signal. It contains information about the target's structural dimensions and structure. HRRPs are easy to acquire, store, and process, making HRRP-based radar target recognition a current research hotspot. Existing HRRP-based target recognition methods all assume that the dataset labels are completely correct; however, this assumption is difficult to satisfy in real-world environments.
[0003] In real-world environments, the acquisition of HRRP data for non-cooperative targets is influenced by numerous factors. First, electronic jamming and deception techniques severely impact HRRP data acquisition. Electronic jamming causes deviations in radar system target tracking, while electronic deception generates false radar echo signals; both contribute to noise labels in HRRP data. Second, the complex real-world environment is another significant factor affecting HRRP data acquisition. In environments with dense target concentrations or obstructed line of sight, radar systems may misinterpret signals from different targets as a single target or mix signals from multiple targets, resulting in noise labels. Furthermore, the opponent's camouflage and deception strategies can make it difficult for radar systems to obtain accurate target information. By using camouflage to conceal the real target, radar systems can collect erroneous HRRP data. Finally, human factors cannot be ignored. During the acquisition and labeling of HRRP data, operator misjudgments or errors can also lead to noise labels.
[0004] In summary, real-world data acquisition scenarios are complex, and broadband echo data acquisition of non-cooperative targets is subject to mislabeling, incorrect labeling, and unknown labels, which will lead to a sharp degradation or even failure of the performance of existing HRRP target identification methods. Summary of the Invention
[0005] To address the aforementioned problems in existing technologies, this invention provides a semi-supervised HRRP noise label filtering method based on early learning guidance. The technical problem to be solved by this invention is achieved through the following technical solution:
[0006] This invention provides a semi-supervised HRRP noise label filtering method based on early learning guidance, comprising:
[0007] S1: Construct an HRRP training dataset with noise labels and a deep neural network model;
[0008] S2: Input the HRRP training dataset into the deep neural network model for multiple rounds of pre-training to obtain the pre-trained network model;
[0009] S3: Use the pre-trained network model to predict the HRRP training dataset, and divide the HRRP training dataset into two datasets X and U according to the confidence of the model prediction results;
[0010] S4: Construct a known-label dataset and an unknown-label dataset using the datasets X and U to form a semi-supervised training dataset, and perform data augmentation on the data in the semi-supervised training dataset;
[0011] S5: Use the data augmented to perform semi-supervised training on the pre-trained network model;
[0012] S6: Generate labels for the unknown label dataset using the semi-supervised trained network model, and perform fully supervised training on the semi-supervised trained network model to obtain a fully supervised trained network model.
[0013] S7: Use the fully supervised trained network model to perform target recognition on the HRRP data to be identified.
[0014] In one embodiment of the present invention, constructing an HRRP training dataset with noise labels includes:
[0015] Construct an HRRP training dataset with noise labels. Where N is the total number of data in the HRRP training dataset, x i This represents the i-th HRRP data entry. Representing data x i Corresponding target information tags, tags Whether it is correct or not is unknown, and some of the labels in the HRRP training dataset are incorrect labels.
[0016] In one embodiment of the present invention, the loss function used in the pre-training process is the generalized cross-entropy loss function, the expression of which is:
[0017]
[0018] in, f represents the generalized cross-entropy loss. y (x; θ) represents the predicted probability of the label category obtained by inputting the data sample x into the deep neural network model, θ represents the model parameter, and q is a preset constant.
[0019] In one embodiment of the present invention, S3 includes:
[0020] The pre-trained network model is used to predict the HRRP training dataset. Based on the model prediction results, the HRRP training dataset is divided into two datasets, X and U. Data with a model prediction confidence greater than the confidence threshold p are assigned to dataset X, and samples with a model prediction confidence less than the confidence threshold p are assigned to dataset U.
[0021] Update the labels of the samples in dataset X to obtain the updated label for each data point in dataset X:
[0022]
[0023] in, For data sample x i Updated tags Represents data sample x in dataset X i The model predicts the probability distribution. This indicates that the pre-trained deep neural network model will use data samples x i The probability of predicting it as class K, where argmax{} represents the target type label corresponding to the maximum predicted probability.
[0024] In one embodiment of the present invention, a known-label dataset and an unknown-label dataset are constructed using the datasets X and U to form a semi-supervised training dataset, including:
[0025] The dataset X with updated labels is used as the known labeled dataset for the semi-supervised learning training set, denoted as X.
[0026] Remove the labels from dataset U, and use dataset U as the unlabeled dataset for the semi-supervised learning training set, denoted as D. unlabeled ={x|x∈U}.
[0027] In one embodiment of the present invention, data augmentation is performed on the HRRP data in the semi-supervised training dataset, including:
[0028] For a known labeled dataset D labeled Perform weak data augmentation to obtain the augmented dataset. For the unknown label dataset D unlabeled Weak data augmentation and strong data augmentation were performed separately to obtain the augmented dataset. and Wherein, weak data augmentation refers to performing random cropping, translation, and random flipping operations on the data in sequence, while strong data augmentation refers to performing random cropping, translation, random flipping, noise addition, and random overlay operations on the data in sequence.
[0029] In one embodiment of the present invention, S5 includes:
[0030] The enhanced dataset and Input the pre-trained deep neural network model, and perform semi-supervised training using a semi-supervised loss function to obtain the semi-supervised trained network model. The semi-supervised loss function is expressed as follows:
[0031] L = L Q +λL P ,
[0032] Among them, L P L represents the consistency regularization loss. Q This represents the cross-entropy loss, where λ represents a preset coefficient.
[0033]
[0034]
[0035] Where |·| represents the modulo operation, This represents finding the sum of squares. This represents the data after the model augments the weak data. The predicted probability distribution This indicates the data after the model has been augmented with strong data. The predicted probability distribution This represents the cross-entropy loss function.
[0036] In one embodiment of the present invention, S6 includes:
[0037] Using a semi-supervised trained network model for the unknown label dataset D unlabeled Data generation labels in the middle:
[0038]
[0039] Where, p x This represents the model's predicted probability for data x;
[0040] Obtain the dataset after generating labels Merge datasets and D labeled As the new training set D new ;
[0041] Using the new training set D newThe semi-supervised network model is then trained using the cross-entropy loss function until the model converges, thus obtaining the fully supervised network model.
[0042] Another aspect of the present invention provides a storage medium storing a computer program for performing the steps of the semi-supervised HRRP noise label filtering method based on early learning guidance described in any of the above embodiments.
[0043] Another aspect of the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor, when invoking the computer program in the memory, performs the steps of the semi-supervised HRRP noise label filtering method based on early learning guidance as described in any of the above embodiments.
[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0045] 1. This invention aims to address the problem of noise labeling in radar target recognition, proposing a semi-supervised HRRP noise label learning method based on early learning guidance, filling a gap in the field of HRRP noise label target recognition. By using the generalized cross-entropy loss function as the loss function for early model training, the quality of data partitioning is improved, providing strong data support for semi-supervised learning. A set of data augmentation schemes suitable for HRRP is proposed, and the noise labels are reshaped using a semi-supervised learning framework, significantly reducing the impact of noise labels on the model. Compared with existing technologies, this invention can better reduce the impact of noise labels on the model, while possessing better stability and feasibility.
[0046] 2. Existing methods all use the cross-entropy loss function in the early learning stage of the model. This loss function has low robustness and is sensitive to noisy labels. This invention uses the generalized cross-entropy loss function in the early learning and training stage of the model. This loss function has higher robustness, is not sensitive to noisy labels, and greatly improves the reliability of data partitioning.
[0047] 3. This invention combines weak and strong data augmentation techniques to enable the model to provide consistent predictions under varying degrees of perturbation. This consistency regularization strategy significantly improves the model's generalization ability. The model not only needs to perform well on the original data but also needs to maintain stable predictions on data after various transformations, making the trained model more robust. When generating pseudo-labels, these pseudo-labels are only used for training when the model is very confident in its predictions (i.e., the prediction probability exceeds a certain threshold). This mechanism reduces the introduction of erroneous pseudo-labels, thereby reducing the risk of the model learning incorrect information. In this way, it can more robustly handle unlabeled data, improving the overall performance of the model. The algorithm is relatively simple and easy to implement, requiring no complex network structure or additional training stages. This simplicity makes it easy to apply to practical engineering projects.
[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0049] Figure 1 This is a flowchart of a semi-supervised HRRP noise label filtering method based on early learning guidance provided in an embodiment of the present invention. Detailed Implementation
[0050] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of the semi-supervised HRRP noise label filtering method based on early learning guidance proposed in accordance with the present invention is provided in conjunction with the accompanying drawings and specific embodiments.
[0051] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.
[0052] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element.
[0053] To address the performance degradation of classic HRRP target recognition algorithms in noisy label scenarios, this invention proposes a semi-supervised HRRP noise label filtering method based on early learning guidance. This method improves the data partitioning quality of early learning guidance by using a generalized cross-entropy loss function as the loss function for early guidance. Simultaneously, a data augmentation scheme for HRRP is proposed, introducing a semi-supervised learning framework to solve the problems caused by noise labels in the HRRP target recognition domain.
[0054] Please see Figure 1 , Figure 1 This is a flowchart of a semi-supervised HRRP noise label filtering method based on early learning guidance provided by an embodiment of the present invention. The specific implementation steps of this method are as follows:
[0055] S1: Construct the HRRP training dataset and deep neural network model with noise labels.
[0056] Construct an HRRP training dataset with noise labels. Where N is the total number of data in the HRRP training dataset, x i This represents the i-th HRRP data entry. Representing data x i Corresponding target information tags, tags Whether it is correct or not is unknown, and some of the labels in the HRRP training dataset are incorrect labels.
[0057] Subsequently, a deep neural network model f(θ) is constructed. Any existing deep network model can be used, such as VGG13, VGG16, ResNet18, ResNet50, etc.
[0058] S2: Input the HRRP training dataset into the deep neural network model for multiple rounds of pre-training to obtain the pre-trained deep neural network model.
[0059] The pre-training phase uses the generalized cross-entropy loss function, which is expressed as follows:
[0060]
[0061] in, f represents the generalized cross-entropy loss. y (x; θ) represents the predicted label category probability obtained by inputting data sample x into the deep neural network model, θ represents the model parameters, and q is a preset constant. In this embodiment, q = 0.9. T is set between 10 and 50.
[0062] S3: Use the pre-trained deep neural network model to predict the HRRP training dataset, and divide the HRRP training dataset into two datasets X and U based on the model prediction results.
[0063] Specifically, data with model prediction confidence scores greater than the confidence threshold p are assigned to dataset X, and samples with model prediction confidence scores less than the confidence threshold p are assigned to dataset U. Simultaneously, the sample labels in dataset X are updated to obtain the updated label for each data point in dataset X:
[0064]
[0065] in, For data sample x i The updated label, i.e. the target type label corresponding to the maximum probability, Represents data sample x in dataset X i The model predicts the probability distribution. This indicates that the pre-trained deep neural network model will use data samples x i The probability of predicting the class K is given, and argmax{} represents the target type label corresponding to the maximum predicted probability. The model prediction confidence is the maximum probability value in the vector obtained by the model prediction. Preferably, the confidence threshold p is set between 0.85 and 0.99.
[0066] S4: Construct a known-label dataset and an unknown-label dataset using datasets X and U to form a semi-supervised training dataset, and perform data augmentation on the data in the semi-supervised training dataset.
[0067] The semi-supervised learning training set in this embodiment comprises two parts: a dataset with known labels and a dataset with unknown labels. The semi-supervised dataset is constructed based on the two datasets X and U obtained in step S3. Specifically, the dataset X with updated labels is used as the known label dataset of the semi-supervised learning training set, denoted as... Remove the labels from dataset U, and use dataset U as the unlabeled training set for semi-supervised learning, denoted as D. unlabeled ={x|x∈U}.
[0068] Subsequently, data augmentation was performed on the HRRP data in the semi-supervised training dataset.
[0069] Basic enhancement schemes for HRRP data include:
[0070] (1) Random cropping: Randomly select a distance cell interval of HRRP data and replace the amplitude information in the interval with random noise.
[0071] (2) Random flip: Flip the HRRP data according to the distance unit.
[0072] (3) Translation: The entire HRRP data is cyclically shifted to the left or right along the distance dimension. For example, if the 1*200 signal is cyclically shifted 5 bits to the right, the 200th distance unit data is shifted to the 1st position, the 1st position is shifted to the 2nd position, and so on. This process is repeated five times, and finally the 200th position is shifted to the 4th position.
[0073] (4) Random overlay: Set a portion of the distance cells in the HRRP data to zero.
[0074] (5) Add noise: Add Gaussian white noise to the HRRP data.
[0075] Furthermore, HRRP's weak and strong data augmentation are derived from a series of basic data augmentation combinations, as detailed below:
[0076] HRRP Weak Data Augmentation: {Random Pruning, Shifting, Random Flipping}
[0077] HRRP Strong Data Augmentation: {Random pruning, translation, random flipping, adding noise, random overlay}
[0078] For a known labeled dataset D labeled Perform weak data augmentation to obtain the augmented dataset. For the unknown label dataset D unlabeled Weak data augmentation and strong data augmentation were performed separately to obtain the augmented dataset. and Weak data augmentation involves performing random cropping, translation, and random flipping operations on the data in sequence, while strong data augmentation involves performing random cropping, translation, random flipping, noise addition, and random overlay operations on the data in sequence.
[0079] S5: Use data augmented data to perform semi-supervised training on the pre-trained network model.
[0080] Based on the data augmented dataset and Semi-supervised training is performed on the pre-trained deep neural network model. Specifically, the augmented dataset is used... and Input the pre-trained network model and perform semi-supervised training using a semi-supervised loss function, which is expressed as follows:
[0081] L = L Q +λL P ,
[0082] Among them, L PThe consistency regularization loss is based on the core idea that the model should produce similar outputs on the same input data after slight perturbations. This means that even when the input data changes slightly, the model's predictions should remain consistent. Q λ represents the cross-entropy loss, which makes the model's predictions closer to the sample labels. λ represents a set coefficient, preferably λ = 1.
[0083] Furthermore, the consistency regularization loss L P and cross-entropy loss L Q The expression is:
[0084]
[0085]
[0086] Where |·| represents the modulo operation, This represents finding the sum of squares. This represents the data after the model augments the weak data. The predicted probability distribution This indicates the data after the model has been augmented with strong data. The predicted probability distribution Let l represent the cross-entropy loss function. ce (p,y)=-log(p y ), .
[0087] S8: Use the semi-supervised trained network model to generate labels for the unknown label dataset, and then perform fully supervised training on the semi-supervised trained network model to obtain the fully supervised trained network model.
[0088] Using a semi-supervised trained network model for the unknown label dataset D unlabeled Data generation labels in the middle:
[0089]
[0090] Where, p x This represents the model's predicted probability for data x.
[0091] Obtain the dataset after generating labels Merge datasets and D labeled As the new training set D new ;
[0092] Using the new training set D new The semi-supervised network model is trained under full supervision until the model converges, resulting in a fully supervised network model. The cross-entropy loss function is used in this stage.
[0093] S9: Use the fully supervised trained network model to identify targets in the HRRP data to be identified.
[0094] Specifically, by inputting the HRRP data to be identified into the fully supervised network model, the corresponding target type identification result, i.e., the probability distribution of the target type, can be output.
[0095] The following simulation experiments further illustrate the effectiveness of the semi-supervised HRRP noise label filtering method based on early learning guidance proposed in this embodiment of the invention.
[0096] (1) Simulation experimental conditions:
[0097] This invention utilizes HRRP (High-Resolution Rate) measured data from a certain X-band broadband radar for ten categories of civil aircraft. The ten categories are: Airbus A319, Airbus A320, Airbus A321, Airbus A330-200, Airbus A330-300, Airbus A350-941, Boeing 737-700, Boeing 737-800, Boeing 747-89L, and Bombardier CRJ-900. The model and dimensional parameters of these ten aircraft are shown in Table 1.
[0098] Table 1. Information on Ten Types of Aircraft Models and Dimensions
[0099] Aircraft model Length (m) Width (m) Height (m) Airbus A319 33.84 34.09 11.76 Airbus A320 37.57 34.09 11.76 Airbus A321 44.51 34.09 11.76 Airbus A330-200 58.82 60.30 17.39 Airbus A330-300 63.69 60.30 16.83 Airbus A350-941 66.80 64.75 17.05 Boeing 737-700 33.60 34.30 12.50 Boeing 737-800 39.50 34.30 12.50 Boeing 747-89L 70.60 59.64 19.30 Bombardier CRJ-900 36.19 24.85 7.57
[0100] Experimental data setup: Training dataset: 10,000 HRRP data points per class, totaling 100,000; Test dataset: 2,000 HRRP data points per class, totaling 20,000; Small-scale clean-label dataset: 100 data points per class, totaling 1,000.
[0101] Classic HRRP Target Recognition Methods: This invention compares four classic HRRP target recognition methods: Random Forest, Support Vector Machine, Convolutional Neural Network, and Recurrent Neural Network. Specific parameter settings are as follows:
[0102] 1) Random Forest (RF): The number of decision trees in RF is set to 200. To ensure the diversity of decision trees, the depth of decision trees is not constrained in the experiment.
[0103] 2) Support Vector Machine (SVM): The model parameters are optimized using the grid search method. Based on the optimization results, the kernel function is set to the radial basis function, the penalty factor C is set to 32, and the kernel function coefficient G is set to 1.
[0104] 3) Convolutional Neural Network (CNN): The network architecture used is ResNet18. During the training phase, the optimization algorithm for the CNN is set to the Adam algorithm, the number of training iterations is set to 200, and the learning rate is set to 0.02.
[0105] 4) Recurrent Neural Network (RNN): The required input is a subsequence of multiple time steps. Therefore, the HRRP samples need to be preprocessed by sliding window slicing to obtain the corresponding HRRP subsequences. Specifically, in the experiment, the sliding window length was set to 32, the step size was 16, and the overlap rate of adjacent windows was 0.5. Finally, each 200-dimensional HRRP sample can obtain a subsequence data with a dimension of 12*32.
[0106] (I) Symmetrical noise experiment:
[0107] The true labels in the training dataset were uniformly flipped to all other categories. Experiments were conducted at four noise-to-label ratios of 0.2, 0.4, 0.6, and 0.8. The experimental results are shown in Table 2. As can be seen from Table 2, the existing classic HRRP recognition method exhibits a sharp decline in recognition performance under noisy label scenarios, failing to meet the requirements of the recognition task. The method proposed in this invention demonstrates the highest recognition performance at different label noise ratios. Notably, at label noise ratios of 0.2, 0.4, and 0.6, the method maintains a recognition accuracy of over 88%; and at a label noise ratio of 0.8, the method shows a 12.27% performance improvement compared to the suboptimal method.
[0108] Table 2 Performance of the present invention and the classic radar HRRP target recognition method in a symmetric noise tag scenario.
[0109]
[0110] (II) Asymmetric noise experiment:
[0111] The real labels in the training dataset were randomly flipped to another class, and experiments were conducted at four noise label ratios of 0.1, 0.2, 0.3, and 0.4. The experimental results are shown in Table 3. As can be seen from Table 3, the existing classic HRRP recognition method exhibits a sharp decline in recognition performance under noisy label scenarios, failing to meet the requirements of the recognition task. The method proposed in this invention achieves the highest recognition performance at different label noise ratios. Notably, when the label noise is 0.2, 0.4, 0.6, and 0.8, the method of this invention maintains a recognition accuracy of over 88%.
[0112] Table 3. Performance of the classic radar HRRP target recognition method in asymmetric noise tag scenarios
[0113]
[0114] This invention aims to address the problem of noise labeling in radar target recognition, proposing a semi-supervised HRRP noise label learning method based on early learning guidance, filling a gap in the field of HRRP noise label target recognition. By using the generalized cross-entropy loss function as the loss function for early model training, the quality of data partitioning is improved, providing strong data support for semi-supervised learning. A set of data augmentation schemes suitable for HRRP is proposed, using a semi-supervised learning framework to reshape noise labels, significantly reducing the impact of noise labels on the model. Compared to existing technologies, this invention can better reduce the impact of noise labels on the model, while possessing better stability and feasibility. This invention uses the generalized cross-entropy loss function in the early model learning and training stage, which has higher robustness and is insensitive to noise labels, greatly improving the reliability of data partitioning.
[0115] This invention combines weak and strong data augmentation techniques to enable the model to make consistent predictions under varying degrees of perturbation. This consistency regularization strategy significantly improves the model's generalization ability. The model not only needs to perform well on the original data but also maintain stable predictions on data after various transformations, resulting in a more robust model. When generating pseudo-labels, these pseudo-labels are only used for training when the model is very confident in its predictions (i.e., the prediction probability exceeds a certain threshold). This mechanism reduces the introduction of erroneous pseudo-labels, thereby lowering the risk of the model learning incorrect information. In this way, it can more robustly handle unlabeled data, improving the overall performance of the model. The algorithm is relatively simple and easy to implement, requiring no complex network structure or additional training stages. This simplicity makes it easily applicable to practical engineering projects.
[0116] Another embodiment of the present invention provides a storage medium storing a computer program for executing the steps of the semi-supervised HRRP noise label filtering method based on early learning guidance described in the above embodiments. A further aspect of the present invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the semi-supervised HRRP noise label filtering method based on early learning guidance as described in the above embodiments. Specifically, the integrated modules implemented as software functional modules can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0117] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A semi-supervised HRRP noise label filtering method based on early learning guidance, characterized in that, include: S1: Construct an HRRP training dataset with noise labels and a deep neural network model; S2: Input the HRRP training dataset into the deep neural network model for multiple rounds of pre-training to obtain the pre-trained network model; S3: Use the pre-trained network model to make predictions on the HRRP training dataset, and divide the HRRP training dataset into two datasets based on the confidence level of the model prediction results. and ; S4: Utilize the dataset and A known-label dataset and an unknown-label dataset are constructed to form a semi-supervised training dataset, and data augmentation is performed on the data in the semi-supervised training dataset; S5: Use the data augmented to perform semi-supervised training on the pre-trained network model; S6: Generate labels for the unknown label dataset using the semi-supervised trained network model, and perform fully supervised training on the semi-supervised trained network model to obtain a fully supervised trained network model. S7: Use the fully supervised trained network model to perform target recognition on the HRRP data to be identified; The loss function used during pre-training is the generalized cross-entropy loss function, the expression of which is: in, Represents the generalized cross-entropy loss. To sample the data Input the label category prediction probability obtained from the deep neural network model. Indicates model parameters, This is a preset constant; S3 includes: The pre-trained network model is used to predict the HRRP training dataset, and the HRRP training dataset is divided into two datasets based on the model prediction results. and Among them, models with prediction confidence greater than a confidence threshold are... Data is incorporated into the dataset If the model's prediction confidence is less than the confidence threshold, then... The samples were included in the dataset. ; For dataset Update the sample labels in the dataset to obtain the dataset. Labels for each updated data point: ， in, For data samples Updated tags Represents the dataset Data samples The model predicts the probability distribution. , This indicates that the pre-trained deep neural network model will use data samples Predicted as the first K The probability of a class This indicates that the target type label corresponding to the maximum predicted probability is selected.
2. The semi-supervised HRRP noise label filtering method based on early learning guidance according to claim 1, characterized in that, Construct an HRRP training dataset with noisy labels, including: Construct an HRRP training dataset with noise labels. ,in, The total number of data points in the HRRP training dataset. Indicates the first i One HRRP data point, Representing data Corresponding target information tags, tags Whether it is correct or not is unknown, and some of the labels in the HRRP training dataset are incorrect labels.
3. The semi-supervised HRRP noise label filtering method based on early learning guidance according to claim 2, characterized in that, Using the dataset and Construct a semi-supervised training dataset by using known-label and unknown-label datasets, including: Dataset with updated labels The known labeled dataset, which serves as the semi-supervised training dataset, is denoted as... ; Remove dataset The tags will be used to define the dataset. The unlabeled dataset, which is the semi-supervised training dataset, is denoted as... .
4. The semi-supervised HRRP noise label filtering method based on early learning guidance according to claim 3, characterized in that, Data augmentation is performed on the HRRP data in the semi-supervised training dataset, including: For known labeled datasets Perform weak data augmentation to obtain the augmented dataset. For unknown label datasets Weak data augmentation and strong data augmentation were performed separately to obtain the augmented dataset. and The weak data augmentation refers to performing random cropping, translation, and random flipping operations on the data in sequence, while the strong data augmentation refers to performing random cropping, translation, random flipping, noise addition, and random overlay operations on the data in sequence.
5. The semi-supervised HRRP noise label filtering method based on early learning guidance according to claim 4, characterized in that, S5 includes: The enhanced dataset , and Input the pre-trained deep neural network model, and perform semi-supervised training using a semi-supervised loss function to obtain the semi-supervised trained network model. The semi-supervised loss function is expressed as follows: ， in, This represents the consistency regularization loss. Represents cross-entropy loss, Indicates the preset coefficient. in, This indicates a modulo operation. This represents finding the sum of squares. This represents the data after the model augments the weak data. The predicted probability distribution This indicates the data after the model has been augmented with strong data. The predicted probability distribution This represents the cross-entropy loss function.
6. The semi-supervised HRRP noise label filtering method based on early learning guidance according to claim 5, characterized in that, S6 includes: Using a semi-supervised trained network model for an unknown label dataset Data generation labels in the middle: in, Representing data x The model predicts probabilities; Obtain the dataset after generating labels Merge datasets and As a new training set ; Using the new training set The semi-supervised network model is then trained using the cross-entropy loss function until the model converges, thus obtaining the fully supervised network model.
7. A storage medium storing a computer program for performing the steps of the semi-supervised HRRP noise label filtering method based on early learning guidance as described in any one of claims 1 to 6.
8. An electronic device comprising a memory and a processor, the memory storing a computer program, wherein the processor, when invoking the computer program in the memory, performs the steps of the semi-supervised HRRP noise label filtering method based on early learning guidance as described in any one of claims 1 to 6.