Noise label-oriented SAR target recognition model construction method and device

Abstract

The invention relates to a noise label-oriented SAR target recognition model construction method and device, and the method comprises the steps: extracting attribute scattering center features while carrying out feature extraction, and combining the attribute scattering center features with image features, thereby enabling the model classification basis to have more reference information. During sample separation, a plurality of binary Gaussian mixture models are used for modeling loss distribution of each category and separating clean samples and noise samples, so that the situation that distribution obtained by uniformly performing distribution modeling on all categories of data in a traditional label noise learning algorithm is inconsistent with actual distribution of partial categories of data is avoided. In addition, in the semi-supervised learning process of the first fusion network and the second fusion network after sample separation, the updating of the noise label is calibrated by a joint distribution alignment strategy. Through a joint distribution alignment strategy of clean labels and guessed labels, a distribution offset phenomenon caused by excessive accumulation of deviations in a label guessing result is reduced, and training precision is improved.

Description

Technical Field

[0001] This invention relates to the field of synthetic aperture radar target recognition technology, and in particular to a method and apparatus for constructing a SAR target recognition model oriented towards noise tags. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active imaging system with all-weather, all-day sensing capabilities, providing accurate and rich reference information for missions such as detection and rescue. In addition to imagery, Attributed Scattering Centers (ASCs) can represent SAR targets from another perspective; ASCs contain inherent physical characteristics in the frequency domain.

[0003] Synthetic Aperture Radar (SAR) target recognition algorithms can automatically detect, classify, and identify targets of interest from complex SAR images. Generally, SAR target recognition algorithms borrow from visible light image target recognition methods, employing convolutional neural networks (CNNs) trained on SAR image datasets. However, the performance of the trained neural network largely depends on the quality of the data labels. Due to the complexity of the imaging mechanism, SAR data annotation requires specialized expert knowledge, making it almost inevitable that mislabeled samples (noise labels) will be present in real-world scenarios. Overfitting to noisy labels severely degrades the generalization performance of the neural network model. Therefore, identifying and correcting these mislabeled SAR data in the presence of noisy labels, and ensuring that the model can learn correct patterns and features from noisy data to improve its robustness and generalization ability, has become a pressing challenge.

[0004] Since synthetic aperture radar (SAR) target recognition algorithms are typically trained using visible light image target recognition methods, noise label learning algorithms for visible light images are existing solutions for target recognition when SAR data contains noise labels. Regarding the noise label learning problem, to achieve noise-robust training of SAR target recognition neural networks, extensive research has been conducted from four perspectives: model architecture, regularization techniques, loss functions, and sample selection strategies. Among these, sample selection strategies are a key focus of noise label learning research. The DivideMix algorithm is a widely accepted and high-performance algorithm in sample selection strategies. The core idea of ​​the DivideMix algorithm is to train two neural networks simultaneously. After initial training using all data in the warm-up phase, a Gaussian mixture model (GMM) is used to dynamically divide the samples into clean and noisy samples, distinguishing between these two types of samples based on a set threshold. Then, co-refinement and co-guessing techniques are used to relabel clean samples and assign new labels to noisy samples, improving label quality and reducing the impact of noise. Subsequently, the processed samples were mixed with the original samples using the MixMatch data augmentation strategy to generate new training data, which was then optimized by combining cross-entropy loss and mean squared error loss.

[0005] However, in the DivideMix noisy label learning algorithm, the Gaussian mixture model is generally used for sample selection across all categories. But during training, the neural network learns at different paces for each category; categories with obvious features train faster, while those with unclear or indistinct features train slower. This results in different distributions of loss values ​​for each category. Consequently, the strategy of selecting samples across the entire dataset cannot find a suitable threshold to separate clean and noisy labeled data for all categories, leading to inaccurate data separation. Furthermore, in DivideMix's semi-supervised training, when the proportion of noisy labels in the dataset is high, the gap between the number of clean and noisy labeled samples in a particular category becomes increasingly smaller. This can cause the feature patterns learned by the neural network model in that category to be severely influenced or even dominated by noisy labeled samples. During training, the accumulation of errors leads to distribution shifts, severely impacting the performance of the noisy label learning algorithm and reducing training accuracy. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and apparatus for constructing a SAR target recognition model for noise labels to address the above-mentioned technical problems. This method can accurately separate clean samples from noise samples and correct the distribution offset problem during model training, thereby improving the training accuracy of the SAR target recognition model under noise labels.

[0007] In a first aspect, the present invention provides a method for constructing a SAR target recognition model oriented towards noise labels, comprising the following steps: The original input data is determined based on the hybrid SAR data, which includes the SAR target's ASC data and image data; The original input data is fed into the first fusion network and the second fusion network respectively. Both the first fusion network and the second fusion network are composed of an image feature extraction network and an ASC feature extraction network. The ASC features and image features obtained from the first fusion network and the second fusion network are respectively spliced ​​together to obtain the first fusion feature and the second fusion feature; Using a multi-class Gaussian mixture model, the mixed SAR data is divided into a first clean subset and a first noisy subset based on the first fusion feature, and then divided into a second clean subset and a second noisy subset based on the second fusion feature. A semi-supervised learning strategy with joint distribution alignment is adopted. The sample labels of the first clean subset and the first noisy subset are updated using the second fusion network, and the sample labels of the second clean subset and the second noisy subset are updated using the first fusion network. Based on the Mixup strategy, the first clean subset and the first noisy subset after the labels are updated are combined to obtain the first combined sample set, and the second clean subset and the second noisy subset after the labels are updated are combined to obtain the second combined sample set. The second fusion network is trained using the first combined sample set, the loss is calculated and the second fusion network is updated in reverse. The first fusion network is trained using the second combined sample set, the loss is calculated and the first fusion network is updated in reverse. Determine whether the number of updates of the first fusion network and / or the second fusion network is less than or equal to the preset number. If so, return to the step of inputting the original input data into the first fusion network and the second fusion network respectively. Otherwise, output the first fusion network and the second fusion network as the final SAR target recognition model.

[0008] In one embodiment, ASC data of a SAR target is obtained by inputting hybrid SAR data into an ASC parameter model.

[0009] In one embodiment, the steps of dividing the hybrid SAR data into a first clean subset and a first noisy subset based on a first fusion feature using a multi-class Gaussian mixture model and dividing the hybrid SAR data into a second clean subset and a second noisy subset based on a second fusion feature are the same, both including: Calculate the loss for each sample of each category in the hybrid SAR data based on the fusion features of each sample; Fit a binary Gaussian mixture model using the loss distribution for each category; Data is divided into each class based on a set probability threshold. Samples with a probability of being correctly labeled that is greater than or equal to the set probability threshold are assigned to the clean subset, while those with a probability of being incorrectly labeled are assigned to the noisy subset.

[0010] In one embodiment, a semi-supervised learning strategy of joint distribution alignment is employed to update the sample labels of the first clean subset and the first noisy subset using a second fusion network, including: For each sample in the first clean subset, the original sample and multiple samples after sample augmentation are input into the second fusion network to obtain multiple prediction vectors. The multiple prediction vectors are combined with the original label of the original sample to obtain a refined label. The refined label is then temperature-sharpened and replaced with the original clean label. For each sample in the first noise subset, the original sample and multiple samples after sample enhancement are simultaneously input into the first fusion network and the second fusion network to obtain multiple prediction vectors. The average of these vectors is then used to obtain the guessed label. The joint distribution of the refined label and the guessed label is matched with the prior predicted edge class distribution to calibrate the guessed label. The calibrated guessed label is then temperature-sharpened and replaced with the original noise label.

[0011] In one embodiment, the formula for calculating the loss of each sample in each category of the hybrid SAR data based on the fusion features of each sample is as follows: In the formula, y n This represents the original label of the nth sample. Let represent the fusion feature of the nth sample, C represent the total number of categories, and ℓ(·) represent the classification function of the fusion network corresponding to the sample.

[0012] In one embodiment, the formula for combining multiple predicted vectors with the original labels of the original samples to obtain refined labels is as follows: In the formula, This represents the refined label of the nth sample. Let M represent the probability that the nth sample is correctly labeled, and M represent the number of samples augmented. This represents the fusion feature of the m-th enhanced sample out of n samples; The formula for guessing the label is: In the formula, q n This represents the guessed label for the nth sample.

[0013] In one embodiment, the expression for calibrating the guessed label is: In the formula, This represents the calibration guess label for the nth sample. This represents the distribution of clean labels relative to the overall data. This represents the cumulative average prediction of clean labels relative to relabeled samples. This represents the marginal class distribution of the prior predictions for the practice labels and the guessed labels.

[0014] In one embodiment, the loss during training of the fusion network using a combined sample set consists of cross-loss and mean squared error loss.

[0015] In one embodiment, the image feature extraction network is a convolutional neural network, and the ASC feature extraction network is a dynamic graph convolutional neural network.

[0016] Secondly, the present invention also provides a SAR target recognition model construction device for noise tags, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-mentioned SAR target recognition model construction method for noise tags.

[0017] The beneficial effects of this invention are: (1) The data input to the fusion network of this invention is a combination of ASC data and image data. The attribute scattering center is treated as graph structure data. The physical features extracted by the dynamic graph convolutional neural network effectively enrich the representation of deep image features, strengthen the fusion network, i.e. the SAR target recognition model, for the representation of SAR targets, and make the model classification basis have more reference information. (2) This invention uses a multi-class Gaussian mixture model to screen and classify mixed SAR data, avoiding the problem in traditional label noise learning algorithms where the distribution modeled uniformly for all categories of data is inconsistent with the actual distribution of some categories of data. At the same time, distribution modeling is performed on individual categories, avoiding the influence of misaligned superposition of other category distributions, making the boundary between the modeled clean label data distribution and the noisy label data distribution clearer, with stronger separability and higher screening accuracy; (3) The present invention adopts a semi-supervised learning strategy of joint distribution alignment to replace the labels of the clean subset and the noisy subset divided by the current fusion network using another fusion network. In the process of replacement, label distribution correction is performed. The distribution offset phenomenon caused by excessive accumulation of bias in the label guessing results is reduced by the joint distribution alignment strategy of clean labels and guessed labels, thereby improving the training accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method for constructing a noise-label-oriented SAR target recognition model provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a dynamic graph convolutional neural network provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In one embodiment, such as Figure 1 As shown in the flowchart, the method for constructing a SAR target recognition model for noise tags in this embodiment includes the following steps: S101. Determine the original input data based on the hybrid SAR data. The original input data includes the SAR target's ASC data and image data.

[0021] Hybrid SAR data consists of multiple samples, each of which includes data and a corresponding target class label.

[0022] Specifically, ASC data of SAR targets is obtained by inputting hybrid SAR data into the ASC parameter model.

[0023] Compared to image features, ASC features can more deeply reflect the geometric structure and material properties of a target, and further consider the variation of scattering response with frequency and angle, making the final model more consistent with actual physical conditions. Learning the feature representations in the ASC dataset can stably represent the target under complex operational conditions.

[0024] S102. Input the original input data into the first fusion network and the second fusion network respectively. Both the first fusion network and the second fusion network are composed of an image feature extraction network and an ASC feature extraction network.

[0025] In this embodiment, the image feature extraction network is a convolutional neural network (CNN), and the ASC feature extraction network is a dynamic graph convolutional neural network (DGCNN). The image feature extraction network uses ResNet18, a classic CNN specifically designed as the backbone network for image feature extraction. For ASC data, it is treated as point data for feature extraction. The parameter set contained in each scattering center point is used as the feature vector for that point. The feature extraction network used is a dynamic graph convolutional neural network (DGCNN), with the network structure as shown below. Figure 2 As shown, DGCNN can capture topological information based on dynamically updated local neighborhood graphs. Furthermore, CNN and DGCNN can extract features from different perspectives, and their combined use can facilitate the exploration of complementary information. CNN and DGCNN are fused through feature concatenation.

[0026] The dynamic graph convolutional network extracts features from the scattering center by combining local and global features. The resulting feature vector is then concatenated with the convolutional neural network as the extracted image features, preserving data from two different observation angles to the greatest extent possible.

[0027] In this embodiment, the attribute scattering center is treated as graph structure data. The physical features extracted by the dynamic graph convolutional neural network effectively enrich the representation of deep image features and effectively solve the problem of insufficient SAR image information. This overcomes the limitations of existing image-based methods when applied to real-world SAR data.

[0028] S103. The ASC features and image features obtained from the first fusion network and the second fusion network are respectively spliced ​​to obtain the first fusion feature and the second fusion feature.

[0029] It should be noted that both the first and second fusion features are composed of fusion features from multiple samples. The first fusion feature is obtained by concatenating the ASC features of the first fusion network with the image features, and the second fusion feature is obtained by concatenating the ASC features of the second fusion network with the image features.

[0030] S104. Using a multi-class Gaussian mixture model, the mixed SAR data is divided into a first clean subset and a first noisy subset based on the first fusion feature, and the mixed SAR data is divided into a second clean subset and a second noisy subset based on the second fusion feature.

[0031] After integrating the scattering and deep features of the target, each branch should divide the training samples into a clean subset and a noisy subset for semi-supervised learning of the other branch. Since CNNs tend to learn clean samples faster than noisy samples, existing methods attempt to model the loss distribution of each sample using a single mixture model for data partitioning. However, the loss distribution of each sample exhibits inconsistencies across different classes, making data partitioning based on a single-class independent mixture model imprecise.

[0032] This embodiment improves the selection accuracy of clean and noisy label subsets by employing multiple independent binary Gaussian mixture models for the loss distribution of each category in the dataset. Specifically, using a multi-class Gaussian mixture model, the steps of dividing the mixed SAR data into a first clean subset and a first noisy subset based on the first fusion feature and dividing the mixed SAR data into a second clean subset and a second noisy subset based on the second fusion feature are the same, both including: S1041. Calculate the loss of each sample in each category in the hybrid SAR data based on the fusion features of each sample.

[0033] Specifically, the formula for calculating the loss of each sample in each category of the hybrid SAR data based on the fusion features of each sample is as follows: In the formula, y n This represents the original label of the nth sample. Let represent the fusion feature of the nth sample, C represent the total number of categories, and ℓ(·) represent the classification function of the fusion network corresponding to the sample.

[0034] S1042. Fit a binary Gaussian mixture model using the loss distribution for each class. For each sample in the corresponding class, the probability π of it being correctly labeled is... n It can be derived from a Gaussian distribution with a small mean.

[0035] S1043. Based on the set probability threshold, divide the data into each class and classify the samples whose probability of being correctly labeled is greater than or equal to the set probability threshold into the clean subset, and otherwise classify them into the noisy subset.

[0036] It should be noted that the probability threshold δ is selected based on the specific circumstances, generally set to 0.5 or 0.7. Samples assigned to the clean subset retain their original class labels.

[0037] In the formula, x n D represents the SAR data of the nth sample. clean D represents a clean subset. noisy This represents a subset of noise.

[0038] S105. A semi-supervised learning strategy with joint distribution alignment is adopted. The sample labels of the first clean subset and the first noisy subset are updated using the second fusion network, and the sample labels of the second clean subset and the second noisy subset are updated using the first fusion network.

[0039] Semi-supervised learning utilizes unlabeled samples in noisy subsets. To ensure a comprehensive exploration of the training data, guessed / generated pseudo-labels are often assigned to unlabeled samples. However, due to confirmation bias, label guesses can easily accumulate over-accumulate in certain categories, causing a distribution shift in those categories. Therefore, semi-supervised learning can be severely affected by unreliable relabeled samples. To prevent confirmation bias, in this embodiment's joint distribution alignment semi-supervised learning strategy, the first and second fusion networks exchange their respective partitioning results. Clean and noisy samples partitioned by the first fusion network are used to train the second fusion network, and vice versa. This mechanism ensures the differentiation between the two networks and avoids overfitting a single network to incorrect samples.

[0040] It should be noted that the process of updating the sample labels of the first clean subset and the first noisy subset using the second fusion network and updating the sample labels of the second clean subset and the second noisy subset using the first fusion network are basically similar. This embodiment only elaborates on the process of updating the sample labels of the first clean subset and the first noisy subset using the second fusion network.

[0041] Specifically, a semi-supervised learning strategy of joint distribution alignment is adopted, and the sample labels of the first clean subset and the first noisy subset are updated using a second fusion network, including: For each sample in the first clean subset, the original sample and multiple samples after sample augmentation are input into the second fusion network to obtain multiple prediction vectors. The multiple prediction vectors are combined with the original label of the original sample to obtain a refined label. The refined label is then temperature-sharpened and replaced with the original clean label.

[0042] After performing the above process on all samples in the first clean subset, the first clean subset is updated.

[0043] Sample enhancement includes rotation and flipping.

[0044] The formula for combining multiple predicted vectors with the original labels of the original samples to obtain refined labels is as follows: In the formula, This represents the refined label of the nth sample. Let M represent the probability that the nth sample is correctly labeled, and M represent the number of samples augmented. This represents the fusion feature of the m-th enhanced sample of the n-th sample.

[0045] For each sample in the first noise subset, the original sample and multiple samples after sample enhancement are simultaneously input into the first fusion network and the second fusion network to obtain multiple prediction vectors. The average of these vectors is then used to obtain the guessed label. The joint distribution of the refined label and the guessed label is matched with the prior predicted edge class distribution to calibrate the guessed label. The calibrated guessed label is then temperature-sharpened and replaced with the original noise label.

[0046] After performing the above process on all samples in the first noise subset, the first noise subset is updated.

[0047] In this embodiment, a joint distribution alignment strategy is proposed to correct the label distribution during the relabeling process of noisy labels in semi-supervised learning, thereby improving the error accumulation phenomenon that occurs under high noise rate conditions. The expected sample category distribution in the noisy labels is obtained using the category distribution of clean labeled samples. The distribution adjustment coefficient for each category is calculated by combining the prior distribution of the dataset, and the confidence level of each category in the noisy labeled samples is adjusted to ensure that the actual sample distribution better matches the expected distribution, improving the credibility of pseudo-labels and thus improving the accuracy of semi-supervised training.

[0048] The formula for guessing the label is: In the formula, q n This represents the guessed label for the nth sample.

[0049] The expression for calibrating the guessed label is: In the formula, This represents the calibration guess label for the nth sample. This represents the distribution of clean labels relative to the overall data. This represents the cumulative average prediction of clean labels relative to relabeled samples. This represents the marginal class distribution of the prior predictions for the practice labels and the guessed labels.

[0050] The temperature sharpening function is applied to the refined labels and the calibrated guessed labels to minimize the distribution entropy. The formula is as follows: In the formula, p is the output vector of the refined label or the guessed label, K is the length of the output vector, and T is the temperature coefficient. The smaller T is, the closer the output is to the unique heat distribution.

[0051] S106. Based on the Mixup strategy, the first clean subset and the first noisy subset after the label update are combined to obtain the first combined sample set, and the second clean subset and the second noisy subset after the label update are combined to obtain the second combined sample set.

[0052] S107. The second fusion network is trained using the first combined sample set, the loss is calculated and the second fusion network is updated in reverse. The first fusion network is trained using the second combined sample set, the loss is calculated and the first fusion network is updated in reverse.

[0053] In this embodiment, both the first and second fusion networks are trained using a stochastic gradient descent (SGD) optimizer with a batch size of 16 and an initial learning rate of 0.02. The input data is pruned to 64×64. The hyperparameters δ and T are set to 0.6 and 0.5, respectively.

[0054] The loss during training of the fusion network using a combined sample set consists of cross-entropy loss and mean squared error loss. For clean labeled data, cross-entropy loss is used to measure the difference between the model's prediction and the true label; for noisy labeled data, mean squared error loss is used to measure the difference between the model's prediction and the new label. Mean squared error loss is relatively mild and can prevent the model from overfitting to noisy labels.

[0055] In this embodiment, the loss is calculated as follows: In the formula, L represents the total loss, L CE L represents the cross-entropy loss. MES λ represents the mean squared error loss. u It is a hyperparameter that controls the weight of the mean squared error loss. Indicates a sample, Indicates sample label, This indicates the fusion feature.

[0056] S108. Determine whether the number of updates of the first fusion network and / or the second fusion network is less than or equal to the preset number. If yes, return to the step of inputting the original input data into the first fusion network and the second fusion network respectively. Otherwise, output the first fusion network and the second fusion network as the final SAR target recognition model.

[0057] The preset number of iterations can be, but is not limited to, 300. In the step of returning to the first and second fusion networks respectively, both the first and second fusion networks are fusion networks with updated parameters during the current iteration.

[0058] It should be noted that the two fusion networks obtained are the final SAR target recognition models, and either one can be used, or the one with better performance can be selected.

[0059] In addition, during the entire SAR target recognition model construction process, five rounds of pre-training can be conducted first. In each subsequent round of training, all data are evaluated and tested to obtain the loss value. Then, a Gaussian mixture model is used to filter the data. Noise-labeled data is treated as unlabeled data, and a Mini-batch strategy is used to perform semi-supervised training on the divided data. This process is repeated iteratively.

[0060] The present invention also provides a SAR target recognition model construction device for noise tags, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the SAR target recognition model construction method for noise tags in any of the above embodiments.

[0061] In one specific embodiment, in order to evaluate the effectiveness of the proposed noise-label-oriented SAR target recognition model construction method, extensive experiments were conducted on the MSTAR dataset. The method of the present invention was compared with the current state-of-the-art SDF-Net algorithm that fuses scattering and deep features in a weighted manner, the classic noise label learning algorithm with added ASC feature extraction network, the noise label learning algorithm benchmark DivideMix, and the recently proposed noise label learning algorithm UNICON.

[0062] The MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset was compiled by DARPA using high-resolution focusing synthetic aperture radar (SAR) technology to collect SAR images of various Soviet-made military vehicles. In addition, a series of field tests were conducted, covering various scenarios including partially obscured targets, camouflaged targets, and different deployment configurations, creating a systematic and comprehensive database of actual measurement data. Depending on the observation configuration, the operating conditions are divided into Standard Operating Condition (SOC) and various Extended Operating Conditions (EOC).

[0063] Data description for SOC: The training and testing data for SOC were acquired under similar imaging conditions with approximate tilt angles (17° for training data and 15° for testing data). See Table 1 for specific dataset details.

[0064] Table 1. Basic Information about the SOC Dataset Data description for EOC-1 (four depression angle variants): Three types of SAR targets were observed at different depression angles: 17° is the training data, and 15°, 30°, and 45° are the test data. The data distribution is shown in Table 2. Since SAR imaging is very sensitive to changes in depression angle, maintaining robust performance under these conditions is crucial.

[0065] Table 2 Basic Information about the EOC-1 Dataset The performance of the final SAR target recognition model is evaluated by the classification accuracy of symmetric and asymmetric noise under different noise rates. Symmetric noise is generated by uniformly changing a certain percentage of the training data's true labels to other class labels. The percentage of label change refers to the noise rate. Conversely, for asymmetric noise generation, the label change for each class is restricted to specific classes based on inter-class similarity.

[0066] The ratios of symmetric and asymmetric noise were varied from 20% to 80% and 20% to 50%, respectively. The results under SOC are shown in Table 3, demonstrating that the method of this invention consistently achieves superior recognition performance compared to the comparative methods. Compared to other methods, the performance of this invention remains stable with increasing noise ratio. Furthermore, the superior noise robustness becomes even more pronounced for more challenging asymmetric noise. The above sufficiently demonstrates the effectiveness of the solution of this invention in the presence of noisy labels.

[0067] Table 3. Recognition accuracy of the present invention and comparative methods at different noise rates. This invention can be viewed as an extension of DivideMix in terms of model architecture design. DivideMix is ​​an advanced image-based collaborative learning framework containing two network branches. The improvements involved involve three main parts: Multimodal Feature Extraction (MMFE), Multi-class Gaussian Mixture Model Sample Selection (CWSS), and Joint Distribution Alignment (JDA). Symmetric and asymmetric noise ablation experiments were conducted on SOC with a 40% noise rate to demonstrate the contribution of each module to noise-robust learning. Specific results are shown in Table 4. First, it was found that only by providing richer representations through MMFE during the learning process can a certain degree of performance improvement be achieved. This highlights the necessity of combining scattering and depth features in noisy label learning. However, there is still considerable room for performance improvement, especially when dealing with symmetric noise. Then, utilizing CWSS and JDA alone can further improve recognition performance based on the fused features extracted by MMFE. Furthermore, the method based solely on amplitude images also outperforms the baseline framework, demonstrating the versatility of CWSS and JDA for image-based noise-robust learning. Finally, combining all these parts together achieves optimal recognition performance, namely the method of this invention.

[0068] Table 4 Ablation Experiment Results The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for constructing a SAR target recognition model oriented towards noise tags, characterized in that, The method comprises the following steps: determining original input data based on the mixed SAR data, the original input data comprising ASC data and image data of a SAR target; inputting the original input data into a first fusion network and a second fusion network respectively, the first fusion network and the second fusion network each being composed of an image feature extraction network and an ASC feature extraction network; splicing the ASC features and the image features obtained by the first fusion network and the second fusion network respectively to obtain first fusion features and second fusion features; dividing the mixed SAR data into a first clean subset and a first noise subset based on the first fusion features and into a second clean subset and a second noise subset based on the second fusion features by using a multi-class Gaussian mixture model; updating the sample labels of the first clean subset and the first noise subset by using the second fusion network and updating the sample labels of the second clean subset and the second noise subset by using the first fusion network by adopting a joint distribution alignment semi-supervised learning strategy; combining the first clean subset and the first noise subset after the labels are updated to obtain a first combined sample set and combining the second clean subset and the second noise subset after the labels are updated to obtain a second combined sample set based on a Mixup strategy; training the second fusion network by using the first combined sample set, calculating a loss and updating the second fusion network reversely, and training the first fusion network by using the second combined sample set, calculating a loss and updating the first fusion network reversely; determining whether the number of updates of the first fusion network and / or the second fusion network is less than or equal to a preset number, if yes, returning to the step of inputting the original input data into the first fusion network and the second fusion network, and if not, outputting the first fusion network and the second fusion network as a final SAR target recognition model.

2. The noise-targeted label facing SAR target recognition model construction method according to claim 1, characterized in that, The ASC data of the SAR target is obtained by inputting the mixed SAR data into an ASC parameter model.

3. The method of claim 1, wherein the method further comprises: The steps of dividing the mixed SAR data into a first clean subset and a first noise subset based on the first fusion features and into a second clean subset and a second noise subset based on the second fusion features by using a multi-class Gaussian mixture model are the same and each comprises: calculating a sample loss of each class in the mixed SAR data based on the fusion features of each sample; fitting a binary Gaussian mixture model using the loss distribution of each class; dividing the data of each class according to a set probability threshold, dividing samples with a probability greater than or equal to the set probability threshold to the clean subset and otherwise to the noise subset.

4. The method of claim 1, wherein the method further comprises: updating the sample labels of the first clean subset and the first noise subset by using the second fusion network by adopting a joint distribution alignment semi-supervised learning strategy comprises: for each sample in the first clean subset, inputting the original sample and a plurality of samples after sample enhancement into the second fusion network to obtain a plurality of prediction vectors, combining the plurality of prediction vectors with the original label of the original sample to obtain a refined label, and replacing the original clean label with the refined label after temperature sharpening. For each sample in the first noise subset, the original sample and the plurality of samples after sample enhancement are simultaneously input into the first fusion network and the second fusion network to obtain a plurality of prediction vectors, and an average of the plurality of prediction vectors is taken to obtain a guess label; a joint distribution of the refined label and the guess label is matched with an edge class distribution of a prior prediction to calibrate the guess label; and the calibrated guess label is subjected to temperature sharpening to replace the original noise label.

5. The method of claim 2, wherein the method further comprises: The calculation formula of each sample loss of each class in the mixed SAR data based on the fusion features of each sample is as follows: where y n denotes the original label of the nth sample, denotes the fusion feature of the nth sample, C denotes the total number of classes, and ℓ(·) denotes the classification function of the sample corresponding to the fusion network.

6. The noise-targeted label facing SAR target recognition model construction method according to claim 5, characterized in that, The calculation formula of the refined label obtained by combining the plurality of prediction vectors with the original label of the original sample is as follows: In the formula, denotes the refined label of the nth sample, denotes the probability that the nth sample is correctly labeled, and M denotes the number of sample augmentations, denotes the fusion feature of the mth augmented sample of the n samples; The calculation formula of the guess label is as follows: In the formula, q n denotes the guessed label of the nth sample.

7. The method of claim 3, wherein the method further comprises: The expression for calibrating the guess label is as follows: wherein, represents the calibrated guess label for the nth sample, represents the distribution of clean labels with respect to the overall data, represents the cumulative average prediction of clean labels with respect to the re-labeled samples, represents the marginal class distribution of the ground truth labels and the prior prediction of the guess labels.

8. The method of claim 1, wherein the method further comprises: The loss of the fusion network when the combined sample set is used for training is composed of a cross loss and a mean square error loss.

9. The noise-targeted label facing SAR target recognition model construction method according to claim 8, characterized in that, The image feature extraction network is a convolutional neural network, and the ASC feature extraction network is a dynamic graph convolutional neural network.

10. A noise label oriented SAR target recognition model construction device, comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor implements the steps of the method of any one of claims 1 to 9 when executing the computer program.

Images