A Progressive Domain Adaptation Recognition Method for Zero-Shot SAR Target Recognition

Abstract

The present invention discloses a progressive domain adaptation recognition method for zero-shot SAR target recognition, including: importing simulated data samples and measured data samples into a feature alignment network to extract first feature information and second feature information; calculating a distribution difference based on the first feature information and the second feature information; obtaining a trained feature alignment network based on a first loss function; using the trained feature alignment network to predict an unlabeled measured data set and assigning pseudo-labels to the unlabeled measured data set to obtain a first measured data set with pseudo-labels; performing pseudo-label denoising processing on the first measured data set to obtain a measured data set with pseudo-label denoising; importing the measured data set with pseudo-label denoising into the above-mentioned trained feature alignment network to adjust network parameters; inputting the data to be classified into the adjusted feature alignment network to obtain a final classification and recognition result. The present invention effectively removes pseudo-label data containing noise, thereby making the recognition result more stable.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and particularly relates to a progressive domain adaptation recognition method for zero-shot SAR target recognition. Background Art

[0002] Synthetic Aperture Radar (SAR) data has been widely used in many fields such as military reconnaissance, disaster warning, and environmental monitoring. The SAR Automatic Target Recognition (ATR) technology aims to identify the target types in SAR images, and this technology has been widely studied. For SAR-ATR, in order to ensure the effective recognition performance of the classifier, a large amount of labeled data is usually required to train the classifier. Therefore, the lack of training data will affect the recognition performance. Generating SAR images through simulation technology is an effective way to obtain data. However, due to various reasons, there will be differences between the distributions of simulated images and measured images, resulting in limitations in the application of simulated images. Therefore, the key to using simulated images lies in overcoming the differences between simulated images and measured images.

[0003] In recent years, the SAMPLE dataset provided by the Air Force Research Laboratory (AFRL) in the United States, which contains simulated SAR target data, has been used for the research on zero-shot SAR target recognition problems.

[0004] At present, some scholars have studied the application of simulated data in recognition, which can be roughly divided into the following categories:

[0005] (1) Pre-training.

[0006] In the paper "Improving SAR Automatic Target Recognition Models With Transfer Learning From Simulated Data" (IEEE Geoscience and Remote Sensing Letters) published by Malmgren-Hansen et al. in 2017, the classifier was first pre-trained with a large amount of simulated SAR data, and then the pre-trained network was fine-tuned with a small amount of measured SAR data.

[0007] (2) Data transformation or data augmentation.

[0008] In the paper "SAR Target Classification with CycleGAN Transferred Simulated Samples" (IEEE International Geoscience & Remote Sensing Symposium) published by Liu et al. in 2018, it was proposed to use CycleGAN (Cycle Generative Adversarial Network) to transform simulated data to make it more similar to real data, and then use the transformed simulated data and measured data simultaneously to train the network.

[0009] (3) Feature design.

[0010] In the paper "A hierarchical receptive network oriented to target recognition in SAR images" (Pattern Recogn) published by Dong et al. in 2022, a new hierarchical receptive neural network was proposed. First, a signal-guided receptive module was established through a series of fine convolutional filters and then used to encode empirical features and knowledge. Finally, the further refined features were used for model training.

[0011] (4) Domain adaptation method.

[0012] In the paper "SAR target recognition based on task-driven domain adaptation using simulated data" (IEEE Geosci. Remote. Sens. Lett) published by He et al. in 2021, a task-driven domain adaptation (TDDA) transfer learning method was proposed, which can reduce the decline in recognition ability caused by the difference in pitch angle between training and test data. This method reduces the domain distribution difference between training and test data by introducing relevant information on SAR imaging conditions during training.

[0013] (5) Other methods.

[0014] For example, in neural architecture search, or exploring the effectiveness of existing methods and integrating some effective methods. For example, in the paper "Bridging a gap in SAR-ATR: training on fully synthetic and testing on measured data" published by Inkawhich et al. in 2021, by using a deep learning-based ATR model, methods of data augmentation, model construction, loss function selection and combined use are proposed to enhance the feature information learned from simulation data.

[0015] The above methods all use more or less measured data with labels for training during the training process. Currently, it is still not easy to obtain enough measured SAR images for training. Even for some complex tasks, it is difficult to obtain even a small amount of measured data. Summary of the Invention

[0016] To solve the above problems existing in the prior art, the present invention provides a progressive domain adaptation recognition method for zero-shot SAR target recognition. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0017] A progressive domain adaptation recognition method for zero-shot SAR target recognition, the progressive domain adaptation recognition method includes:

[0018] Step 1, obtain labeled simulation data samples and unlabeled measured data samples;

[0019] Step 2, import the labeled simulation data samples and the unlabeled measured data samples into a feature alignment network respectively, and extract first feature information and second feature information correspondingly;

[0020] Step 3, calculate the distribution difference according to the first feature information and the second feature information, and use the distribution difference as the domain adaptation loss;

[0021] Step 4, obtain a first loss function according to the domain adaptation loss and the first classification loss corresponding to the labeled simulation data samples, and obtain a trained feature alignment network based on the first loss function;

[0022] Step 5, use the trained feature alignment network to predict the unlabeled measured data samples, and assign pseudo-labels to the unlabeled measured data samples to obtain a first measured data set with pseudo-labels;

[0023] Step 6, perform pseudo-label denoising processing on the first measured data set to obtain a measured data set with pseudo-label denoising;

[0024] Step 7: Import the measured data set with pseudo-labels denoised into the trained feature alignment network to adjust the network parameters and obtain an adjusted feature alignment network;

[0025] Step 8: Input the data to be classified into the adjusted feature alignment network to obtain the final classification and recognition result.

[0026] In an embodiment of the present invention, the Step 1 includes:

[0027] Step 1.1: Add speckle noise to the first simulation image to obtain a second simulation image;

[0028] Step 1.2: Crop the first simulation image, the second simulation image, and the first measured image to the same size to obtain a third simulation image, a fourth simulation image, and a second measured image respectively;

[0029] Step 1.3: Normalize the third simulation image, the fourth simulation image, and the second measured image to the range of [0, 255] to obtain a fifth simulation image, a sixth simulation image, and a third measured image. All the labeled fifth simulation images and labeled sixth simulation images form the simulation data sample, and all the unlabeled third measured images form the measured data sample.

[0030] In an embodiment of the present invention, the feature alignment network includes a ten-layer network architecture. The first layer is a convolutional layer L1, the second layer is a maxPooling layer L2, the third layer is a convolutional layer L3, the fourth layer is a maxPooling layer L4, the fifth layer is a convolutional layer L5, the sixth layer is a maxPooling layer L6, the seventh layer is a convolutional layer L7, the eighth layer is an avgPooling layer L8, the ninth layer is a convolutional layer L9, and the tenth layer is a softmax classifier layer L10;

[0031] Among them, the convolutional kernels of the convolutional layer L1, the convolutional layer L3, the convolutional layer L5, and the convolutional layer L7 are 5×5, and a dropout operation is added after the convolutional layer L7. The convolutional kernel of the convolutional layer L9 is 4×4, and finally the recognition result is output through the softmax classifier layer L10.

[0032] In an embodiment of the present invention, both the first feature information and the second feature information are the feature information output by the convolutional layer L7 of the feature alignment network.

[0033] In an embodiment of the present invention, the domain adaptation loss is:

[0034]

[0035] Where D(XL , X U ) is the domain adaptation loss, f(x l ) is the first feature information, f(x u ) is the second feature information, and E is for calculating the mean;

[0036] The first loss function is:

[0037] L1 = L C (X L ) + λD(X L , X U )

[0038] where L1 is the first loss function, L C (X L ) is the first classification loss, and λ is a hyperparameter.

[0039] In an embodiment of the present invention, step 6 includes:

[0040] Step 6.1: Determine the relationship between the probability of the predicted label output after the measured data set with pseudo - labels is input into the feature alignment network and the threshold. If the probability of the predicted label is less than or equal to the threshold, remove the measured data with pseudo - labels; if the probability of the predicted label is greater than the threshold, retain the measured data with pseudo - labels. All the retained measured data with pseudo - labels form the second measured data set;

[0041] Step 6.2: Based on the similarity between every two measured data in the second measured data set, establish a similarity matrix. For each measured data in the second measured data set, find the pseudo - label of the measured data with the highest similarity to the current measured data in the similarity matrix, and denote it as y tmax , if the pseudo - label y t of the current measured data is the same as y tmax , add the current measured data to the measured data set with pseudo - label denoising; otherwise, do not add it.

[0042] In an embodiment of the present invention, the similarity of the measured data includes cosine similarity, normalized mutual information similarity, or structural similarity.

[0043] In an embodiment of the present invention, step 7 includes:

[0044] Step 7.1: Calculate the second classification loss according to the measured data set with pseudo - label denoising;

[0045] Step 7.2: Obtain the total loss function according to the second classification loss and the first loss function, and adjust the network parameters of the trained feature alignment network based on the total loss function to obtain the adjusted feature alignment network.

[0046] In one embodiment of the present invention, the second classification loss is:

[0047]

[0048] where is the pseudo-label of the preliminary prediction, is the new prediction label after adjustment of the measured data set for pseudo-label denoising in the feature alignment network;

[0049] The total loss function is:

[0050] L2 = L C (X L ) + λD(X L , X U ) + L C (X deno )

[0051] where L2 is the total loss function.

[0052] Advantages of the present invention:

[0053] The present invention only uses labeled simulation data for identifying measured data, effectively reduces the feature distribution difference between simulation data and measured data through the feature alignment network and the self-training method with pseudo-label denoising, enables the simulation data to be used for model training, reduces the difficulty of obtaining labeled training data, and saves the time and economic cost of obtaining measured data.

[0054] The progressive domain adaptation target recognition method for zero-shot SAR target recognition proposed by the present invention can obtain relatively stable preliminary recognition results through the feature alignment network, and then effectively removes the noisy pseudo-label data through the self-training method with pseudo-label denoising, thereby making the recognition results more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a schematic flowchart of a progressive domain adaptation recognition method for zero-shot SAR target recognition provided by an embodiment of the present invention;

[0056] Figure 2 is a network architecture diagram of a step-by-step domain adaptation provided by the present invention;

[0057] Figure 3 is a structural diagram of a feature alignment network provided by an embodiment of the present invention;

[0058] Figure 4 is a process diagram of a self-training method provided by an embodiment of the present invention;

[0059] Figure 5Explanation diagram of a pseudo-label denoising method provided by an embodiment of the present invention;

[0060] Figure 6 Sample diagram of a SAMPLE dataset for experiments provided by an embodiment of the present invention. Detailed implementation manners

[0061] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0062] Embodiment 1

[0063] Please refer to Figure 1 and Figure 2 , Figure 1 which is a schematic flowchart of a progressive domain adaptation recognition method for zero-shot SAR target recognition provided by an embodiment of the present invention; Figure 2 which is a network architecture diagram of a step-by-step domain adaptation provided by the present invention. The present invention proposes a progressive domain adaptation recognition method for zero-shot SAR target recognition. The progressive domain adaptation recognition method includes steps 1-step 8, where:

[0064] Step 1: Obtain labeled simulation data samples and unlabeled measured data samples, where the labels are used to identify category information.

[0065] Step 1.1: Add speckle noise to the first simulation image to obtain a second simulation image;

[0066] Step 1.2: Crop the first simulation image, the second simulation image, and the first measured image to a unified size to obtain a third simulation image, a fourth simulation image, and a second measured image respectively. For example, crop to a size of 64×64 to exclude the influence of the background;

[0067] Step 1.3: Normalize the third simulation image, the fourth simulation image, and the second measured image to the range of [0, 255] to obtain a fifth simulation image, a sixth simulation image, and a third measured image. All the labeled fifth simulation images and labeled sixth simulation images form the simulation data samples, and all the unlabeled third measured images form the measured data samples. The simulation data samples are denoted as X L , and the measured data samples are denoted as X U .

[0068] Step 2: Import the labeled simulation data samples and the unlabeled measured data samples into the feature alignment network respectively to extract the first feature information and the second feature information.

[0069] In this embodiment, for the structure of the feature alignment network Ω, please refer to Figure 3, the feature alignment network includes a ten-layer network architecture. The first layer is the convolutional layer L1, the second layer is the maxPooling layer L2, the third layer is the convolutional layer L3, the fourth layer is the maxPooling layer L4, the fifth layer is the convolutional layer L5, the sixth layer is the maxPooling layer L6, the seventh layer is the convolutional layer L7, the eighth layer is the avgPooling layer L8, the ninth layer is the convolutional layer L9, and the tenth layer is the softmax classifier layer L10;

[0070] Among them, the convolutional kernels of the convolutional layer L1, the convolutional layer L3, the convolutional layer L5, and the convolutional layer L7 are 5×5. A dropout operation is added after the convolutional layer L7. The convolutional kernel of the convolutional layer L9 is 4×4. Finally, the classification result is output through the softmax classifier layer L10.

[0071] Furthermore, both the first feature information and the second feature information are the feature information output by the convolutional layer L7 of the feature alignment network, that is, the labeled simulation data sample X L and the unlabeled measured data sample X U both pass through the feature alignment network, and the corresponding feature information of the two data samples after passing through the convolutional layer L7 is extracted respectively.

[0072] Step 3: Calculate the distribution difference based on the first feature information and the second feature information, and use the distribution difference as the domain adaptation loss.

[0073] Specifically, calculate the distribution difference using the first feature information and the second feature information, and denote it as the domain adaptation loss D(X L , X U ), and the calculation formula is as follows:

[0074]

[0075] Among them, X L represents the labeled simulation data sample, X U represents the unlabeled measured data sample, f(x l ) is the first feature information, f(x u ) is the second feature information, and E is to calculate the mean.

[0076] Step 4: Obtain the first loss function based on the domain adaptation loss and the first classification loss corresponding to the labeled simulation data sample, and obtain the trained feature alignment network based on the first loss function, that is, obtain the trained feature alignment network by minimizing the loss function.

[0077] Specifically, the labeled simulation data sample obtains the corresponding classification loss L C (X L), and then form a new loss function with the domain adaptation loss D(X L , X U ) between the first feature information and the second feature information obtained in step 3, that is, the first loss function. The calculation formula of the first loss function is as follows:

[0078] L1 = L C (X L ) + λD(X L , X U )

[0079] where L1 is the first loss function, L C (X L ) is the first classification loss, and λ is a hyperparameter that determines the proportion of the domain adaptation loss in the loss L1.

[0080] Step 5: Use the trained feature alignment network to predict the unlabeled measured data samples, and assign pseudo-labels to the unlabeled measured data set to obtain the first measured data set with pseudo-labels.

[0081] Specifically, after the unlabeled measured data samples pass through the trained feature alignment network, corresponding predicted labels will be obtained, and these predicted labels will be used as pseudo-labels. Therefore, the measured data set with pseudo-labels is used as the first measured data set with pseudo-labels.

[0082] Step 6: Perform pseudo-label denoising processing on the first measured data set to obtain the measured data set with pseudo-labels denoised.

[0083] Since there are inevitably noises in the pseudo-labels, that is, the predicted labels may be incorrect. Therefore, in order to make the pseudo-labels closer to the true labels, in this embodiment, pseudo-label denoising processing is performed on the first measured data set, which can be specifically carried out through step 6.1 and step 6.2.

[0084] Step 6.1: Judge the relationship between the probability of the predicted label output after the measured data set with pseudo-labels is input into the feature alignment network (that is, the result output by the softmax classifier layer L10 after the measured data set is input into the feature alignment network) and the threshold. If the probability of the predicted label is less than or equal to the threshold, then remove the measured data with the pseudo-labels in the measured data set whose probability of the predicted label is less than or equal to the threshold. If the probability of the predicted label is greater than the threshold, then retain the measured data with the pseudo-labels. All the retained measured data with pseudo-labels form the second measured data set.

[0085] Specifically, set a threshold γ, that is, for the measured data with pseudo-labels If the probability P pred > γ of its predicted label, then retain the measured data with the pseudo-labels for subsequent network fine-tuning, otherwise do not use this data for network fine-tuning.

[0086] Optionally, γ is taken as 0.9.

[0087] Step 6.2: Establish a similarity matrix based on the similarity between every two measured data in the second measured dataset. For each measured data in the second measured dataset, find the pseudo-label of the measured data with the highest similarity to the current measured data (this measured data cannot be the current measured data itself) in the similarity matrix, and denote it as y tmax , if the pseudo-label y t of the current measured data is the same as y tmax , then add the current measured data to the measured dataset with pseudo-label denoising, otherwise, do not add it.

[0088] Specifically, the measured data with pseudo-labels selected by the threshold γ is not completely reliable. In this embodiment, in addition to setting the threshold γ, the image information is also combined to select the pseudo-labels. The present invention uses a method that combines calculating image similarity and the KNN algorithm to remove some pseudo-labeled data containing noise.

[0089] For the denoising process, refer to Figure 4 , Figure 4 The left figure in shows that there are some samples with incorrect predictions in the initial second measured dataset, and the similarity between the samples within the same circle is relatively high as represented by the similarity metric. From Figure 4 , it can be seen that between the two samples with the highest similarity, the predicted pseudo-labels should be consistent. If they are consistent, the pseudo-label samples are retained, otherwise they are excluded. Therefore, according to the comparison of the similarity and its predicted pseudo-labels, the incorrect predicted samples can be excluded, as shown in the right figure of Figure 4 . The specific method for pseudo-label denoising is as follows:

[0090] Establish a similarity matrix S through all the measured data in the second measured dataset. The similarity matrix S is:

[0091]

[0092] In the formula, n represents the number of measured data in the second measured dataset, s tt = 0, 1 ≤ t ≤ n.

[0093] For the measured data x t , the pseudo-label is y t . For the measured data x t , find the pseudo-label of the measured data with the highest similarity in the similarity matrix S, which is denoted as y tmax according to S. Then judge whether the pseudo-labels y tmax and y t are consistent. If they are consistent, then the measured data x tAdd the measured dataset with denoising. Otherwise, do not add it. Perform the above processing on all measured pseudo-label data to obtain the final measured dataset with denoising.

[0094] In this embodiment, three different methods, cosine similarity (COSS), normalized mutual information similarity (NMIS), and structural similarity (SSIM), are used to calculate the similarity relationship between two images x1 and x2, which are briefly introduced as follows:

[0095] 1) COSS:

[0096]

[0097] where a ij and bij represent the pixel values of images x1 and x2 at coordinates (i, j) respectively, n is the number of rows of the image, and m is the number of columns of the image. Cosine similarity measures whether two images point in the same direction by calculating the cosine value of the angle between the column vectors of the two images.

[0098] 2) NMIS:

[0099]

[0100] where H(·) and H(·,·) represent information and joint information entropy respectively. Normalized mutual information similarity measures the similarity between two images x1 and x2 from the perspective of entropy.

[0101] 3) SSIM:

[0102]

[0103] where μ1 or μ2, δ1 or δ2, and δ 12 represent the mean, variance, and covariance of the two images respectively. Structural similarity measures the correlation between two images from three parts: image brightness, contrast, and structure. C1, C2, and C3 are constants set according to relevant literature. C1 is a constant to prevent division by zero in the image brightness comparison part, C2 is a constant to prevent division by zero in the image contrast comparison part, and C3 is a constant to prevent division by zero in the image structure comparison part. Usually, C1 = 0.0001, C2 = 0.0009, and C3 = 0.5 * C2.

[0104] Step 7: Import the measured dataset with pseudo-label denoising into the trained feature alignment network to adjust the network parameters and obtain the adjusted feature alignment network.

[0105] That is to say, in this embodiment, the measured dataset with pseudo-label denoising is input into the trained feature alignment network to fine-tune the network parameters.

[0106] Step 7.1. Calculate the second classification loss L based on the measured dataset with pseudo-label denoising C (X deno ) is expressed as:

[0107]

[0108] where X deno is the measured dataset after pseudo-label denoising, is the preliminarily predicted pseudo-label, is the new predicted label after adjustment of the measured dataset after pseudo-label denoising through the feature alignment network, and will be continuously updated as the network is fine-tuned.

[0109] Step 7.2. Obtain the total loss function based on the second classification loss and the first loss function, and adjust the network parameters of the trained feature alignment network based on the total loss function to obtain the adjusted feature alignment network, that is, fine-tune the network parameters of the feature alignment network until the total loss function is minimized.

[0110] Please refer to Figure 5 , during the fine-tuning process, the total loss function L2 can be expressed as:

[0111] L2 = L C (X L ) + λD(X L , X U ) + L C (X deno )

[0112] where L2 is the total loss function.

[0113] Step 8. Input the data to be classified into the adjusted feature alignment network to obtain the final classification and recognition result.

[0114] The purpose of the present invention is to propose a progressive domain adaptation recognition framework for zero-shot SAR target recognition to address the distribution difference problem between simulated images and measured images, including a feature alignment network and a self-training method with pseudo-label denoising, so as to reduce the feature distribution difference between simulated data and measured data and achieve good recognition performance.

[0115] The present invention proposes a progressive domain adaptation target recognition framework for zero-shot SAR target recognition, which reduces the feature distance between training data and test data by training a feature alignment network, predicts pseudo-label samples for unlabeled samples, and then further improves the accuracy of the recognition result by fine-tuning the feature alignment classification network through a self-training method with pseudo-label denoising.

[0116] The present invention only uses labeled simulation data for identifying measured data, effectively reducing the difference in feature distributions between simulation data and measured data through a feature alignment network and a self-training method with pseudo-label denoising, enabling the simulation data to be used for model training, reducing the difficulty of obtaining labeled training data, and saving the time and economic costs of obtaining measured data.

[0117] The progressive domain adaptation target recognition method for zero-shot SAR target recognition proposed by the present invention can obtain relatively stable preliminary recognition results through a feature alignment network, and then effectively removes the noisy pseudo-label data through a self-training method with pseudo-label denoising, making the recognition results more stable.

[0118] The effects of the present invention can be further illustrated by the following experimental data:

[0119] I. Experimental Data

[0120] The dataset used in this experiment is the SAMPLE dataset, which contains a total of 10 types of data, and each type of data contains measured data and simulation data. The original image size is 128×128, and the resolution is 0.3m×0.3m. The pitch angles of the simulation data and the measured data are between 14° and 17°, and the simulation samples correspond one-to-one with the measured samples. Table 1 introduces the data types and the number of data in the original SAMPLE dataset. This experimental scenario setting is called "Training Scenario 1". See Figure 6 , Figure 6 shows examples of a pair of measured data and simulation data for each type. It can be clearly seen that the simulation images lack background clutter. In addition, the strong scattering centers of the simulation data and the measured data are also different.

[0121] Table 1 Training Scenario 1: Data Types and Quantities of Samples in the Original SAMPLE Dataset

[0122]

[0123] In the experiment of the present invention, the simulation data with a pitch angle range of 14° to 16° is also used as training data, and the measured data with a pitch angle range of 17° is used as test data. This experimental setting is called "Training Scenario 2". Compared with "Training Scenario 1", the pitch angles of the simulation data and the measured data are different. Table 2 introduces the specific data types and quantities of the samples used for training and testing in "Training Scenario 2".

[0124] Table 2 Data Types and Quantities of Samples Used for Training and Testing in Training Scenario 2

[0125]

[0126]

[0127] In this experiment, simulation data was used for training and measured data was used for testing. During the experiment, when preprocessing the images, the data was cropped to a size of 64×64 to exclude background influence, and then the data was normalized to the range of [0, 255] through the min-max scaling method.

[0128] For the network hyperparameters, this experiment used a parameter combination based on experimental exploration, with a batch size value of 256, a dropout value of 0.2, a λ value of 10, the total number of epochs set to 500, and the Adam optimizer was used with a learning rate set to 10 -4 。

[0129] For the given experimental results, each group of experiments was conducted 20 times, and the lowest and highest recognition accuracies were given, denoted by "Min" and "Max" respectively. Ave represents the average recognition accuracy of the 20 experiments, and the standard deviation of the experimental results was given.

[0130] II. Evaluation Criteria

[0131] The following criteria were used to evaluate the experimental results:

[0132] Recognition accuracy Accuracy, minimum recognition accuracy Min, maximum recognition accuracy Max, average recognition accuracy Ave.

[0133] For the above evaluation criteria, the following formulas were used for calculation:

[0134] Recognition accuracy Accuracy:

[0135]

[0136] Minimum recognition accuracy Min:

[0137] Min = min(Accuracy1, Accuracy2,..., Accuracy N )

[0138] In the formula, n represents the number of times the same experiment is repeated under the same parameter settings, Accuracy n represents the recognition accuracy of the nth time of this experiment, and min(·) represents taking the minimum value among the given parameters.

[0139] Maximum recognition accuracy Max:

[0140] Max = max(Accuracy1, Accuracy2,..., Accuracy N )

[0141] where n represents the number of times the same experiment is repeated under the same parameter settings, and Accuracy n represents the recognition accuracy of the nth time of this experiment, and max(·) represents taking the maximum value among the given parameters.

[0142] Average recognition accuracy Ave:

[0143]

[0144] where n represents the number of times the same experiment is repeated under the same parameter settings, and Accuracy n represents the recognition accuracy of the nth time of this experiment.

[0145] III. Experimental Content

[0146] Experiment 1: Conduct experiments using the progressive domain adaptation framework proposed in the present invention. Since this framework is a two-stage training process, feature alignment and self-training processes need to be carried out separately to verify the effectiveness of each stage. Pseudo-label denoising is not performed in this experiment. Tables 3 and 4 give the experimental results. "Baseline" is to add speckle noise to the image enhancement and augmentation of the training data.

[0147] Table 3 Training Scenario 1: Recognition Results of the Progressive Domain Adaptation Framework (%)

[0148]

[0149] Table 4 Training Scenario 2: Recognition Results of the Progressive Domain Adaptation Framework (%)

[0150]

[0151] It can be seen from the above results that the progressive domain adaptation framework proposed in the present invention is effective in both training scenario settings, and the recognition results show a gradually improving trend. In Experiment 2, pseudo-label denoising is considered in the self-training process to obtain better recognition results.

[0152] Experiment 2: In this experiment, pseudo-label dataset denoising is carried out in the self-training process. Three similarity metrics, COSS, NMIS, and SSIM, are used to calculate the similarity between different images for pseudo-label denoising. Tables 5 and 6 give the experimental results of the progressive domain adaptation framework with pseudo-label denoising.

[0153] Table 5 Training Scenario 1: Recognition Results of the Progressive Domain Adaptation Framework with Different Similarity Pseudo-Label Denoising (%)

[0154]

[0155] Table 6 Training scenario 2: Recognition results (%) of the progressive domain adaptation framework with different similarity pseudo-label denoising

[0156]

[0157] By using different similarity metrics for pseudo-label denoising, the average recognition performance can be improved by 1 to 2 percentage points compared with that before denoising, and the standard deviation of the results of multiple experiments becomes smaller. In training scenarios 1 and 2, SSIM performs better and has a higher average recognition rate.

[0158] Experiment 3: Using the SAMPLE dataset, compare the framework of the present invention with other existing frameworks. Tables 7 and 8 respectively give the experimental results of two training scenarios under different framework methods.

[0159] Table 7 Training scenario 1: Comparison with other methods (%)

[0160]

[0161] Table 8 Training scenario 2: Comparison with other methods (%)

[0162]

[0163] In comparison, the method in the present invention achieves a relatively high accuracy rate in the recognition results under training scenario 1. Under training scenario 2, the method in the present invention also exceeds the results of other frameworks in the comparative experiment. There are also other studies that use the SAMPLE dataset to conduct experiments under different experimental settings. For example, in the experiment of the paper "Augmenting simulations for SAR ATR neural network training" published by Sellers et al. in 2020, in addition to the labeled simulation data, one or two measured samples of each class are used for training. The method proposed in this experiment can achieve a classification accuracy rate of 95.1%, while the experimental results in the present invention are still better than it. Therefore, the method of the present invention can be considered to have reached an advanced level.

[0164] In summary, under the condition of only using simulation data for training, the present invention designs a progressive domain adaptation framework by using the pseudo-label denoising method, and conducts experimental verification on the SAMPLE dataset. This framework effectively reduces the distribution difference between simulation data and measured data through the feature alignment network and the self-training method containing pseudo-label denoising, and obtains experimental results superior to other existing frameworks.

[0165] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0166] Although the present application has been described herein in connection with various embodiments, however, in the process of implementing the claimed present application, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases. A single processor or other unit can implement several functions recited in the claims. Certain measures are recited in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0167] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A progressive domain adaptation recognition method for zero-shot SAR target recognition, the progressive domain adaptation recognition method comprising: Step 1, obtaining labeled simulated data samples and unlabeled measured data samples; Step 2, respectively importing the labeled simulated data samples and the unlabeled measured data samples into a feature alignment network to extract first feature information and second feature information correspondingly; Step 3, calculating a distribution difference according to the first feature information and the second feature information, and taking the distribution difference as a domain adaptation loss; Step 4, obtaining a first loss function according to the domain adaptation loss and a first classification loss corresponding to the labeled simulated data samples, and obtaining a trained feature alignment network based on the first loss function; Step 5, using the trained feature alignment network to predict unlabeled measured data samples and assigning pseudo-labels to the unlabeled measured data samples to obtain a first measured data set with pseudo-labels; Step 6, performing pseudo-label denoising processing on the first measured data set to obtain a measured data set with pseudo-label denoising; Step 7, importing the measured data set with pseudo-label denoising into the trained feature alignment network to adjust network parameters and obtain an adjusted feature alignment network; Step 8, inputting the data to be classified into the adjusted feature alignment network to obtain a final classification and recognition result.

2. The progressive domain adaptation recognition method according to claim 1, wherein The Step 1 includes: Step 1.1, adding speckle noise to a first simulation image to obtain a second simulation image; Step 1.2, cropping the first simulation image, the second simulation image and a first measured image to a unified size to correspondingly obtain a third simulation image, a fourth simulation image and a second measured image; Step 1.3, normalizing the third simulation image, the fourth simulation image and the second measured image to the range of [0, 255] to obtain a fifth simulation image, a sixth simulation image and a third measured image. All the labeled fifth simulation images and labeled sixth simulation images form the simulated data samples, and all the unlabeled third measured images form the measured data samples.

3. The progressive domain adaptation recognition method according to claim 1, characterized in that The feature alignment network includes a ten-layer network architecture. The first layer is a convolutional layer L1, the second layer is a maxPooling layer L2, the third layer is a convolutional layer L3, the fourth layer is a maxPooling layer L4, the fifth layer is a convolutional layer L5, the sixth layer is a maxPooling layer L6, the seventh layer is a convolutional layer L7, the eighth layer is an avgPooling layer L8, the ninth layer is a convolutional layer L9, and the tenth layer is a softmax classifier layer L10; Wherein, the convolutional kernels of the convolutional layer L1, the convolutional layer L3, the convolutional layer L5, and the convolutional layer L7 are 5×5, and a dropout operation is added after the convolutional layer L7. The convolutional kernel of the convolutional layer L9 is 4×4, and finally the recognition result is output through the softmax classifier layer L10.

4. The progressive domain adaptation recognition method according to claim 3, characterized in that, Both the first feature information and the second feature information are the feature information output by the convolutional layer L7 of the feature alignment network.

5. The progressive domain adaptation recognition method according to claim 1, characterized in that The domain adaptation loss is as follows: Among them, is the domain adaptation loss, is the first feature information, is the second feature information, is to calculate the mean; The first loss function is: Among them, is the first loss function, is the first classification loss, λ is a hyperparameter.

6. The progressive domain adaptation recognition method according to claim 1, characterized in that Step 6 includes: Step 6.1: Determine the relationship between the probability of the predicted label output after the input features of the measured data set with pseudo-labels are input into the feature alignment network and the threshold. If the probability of the predicted label is less than or equal to the threshold, remove the measured data with pseudo-labels. If the probability of the predicted label is greater than the threshold, retain the measured data with pseudo-labels. All the retained measured data with pseudo-labels form the second measured data set; Step 6.2: Establish a similarity matrix based on the similarity between every two measured data in the second measured dataset. For each measured data in the second measured dataset, find the pseudo-label of the measured data with the highest similarity to the current measured data in the similarity matrix, and denote it as y tmax . If the pseudo-label of the current measured data y t is the same as y tmax , add the current measured data to the measured dataset with pseudo-label denoising; otherwise, do not add it.

7. The progressive domain adaptation recognition method according to claim 6, characterized in that The similarity of the measured data includes cosine similarity, normalized mutual information similarity, or structural similarity.

8. The progressive domain adaptation recognition method according to claim 5, characterized in that Step 7 includes: Step 7.1: Calculate the second classification loss according to the measured data set with denoised pseudo-labels; Step 7.2: Obtain the total loss function according to the second classification loss and the first loss function, and adjust the network parameters of the trained feature alignment network based on the total loss function to obtain the adjusted feature alignment network.

9. The progressive domain adaptation recognition method according to claim 8, characterized in that The second classification loss is: Among them, is the pseudo-label predicted initially, is the new predicted label after adjustment by the feature alignment network for the measured data set with pseudo-label denoising; The total loss function is: Among them, is the total loss function.

Images