SAR image recognition system and method based on distance metric and fully convolutional network

Abstract

The application discloses a SAR image recognition system based on distance measurement and a full convolution network, a feature extraction network training module is used to obtain a trained feature extraction network; a model training module is used to obtain a trained synthetic aperture radar image recognition model; and a synthetic aperture radar image recognition module is used to input a synthetic aperture radar image to be recognized into the trained synthetic aperture radar image recognition model to obtain a synthetic aperture radar image recognition result. The application comprehensively utilizes sample and class center distance, variance of samples of the same kind, class center and center distance. Meanwhile, a feature extraction network composed of multiple small-size convolution layers and a 1x1 convolution layer classifier are designed, a two-stage training method is adopted, and finally the accuracy of SAR image recognition is improved.

Description

Technical Field

[0001] This invention relates to the field of SAR image recognition technology, specifically to a SAR image recognition system and method based on distance metrics and fully convolutional networks. Background Technology

[0002] Synthetic Aperture Radar (SAR) imagery is a high-resolution microwave remote sensing imaging technique that utilizes the synthetic aperture principle. It offers advantages such as all-weather, all-day operation and considerable penetration, making it valuable for applications in marine monitoring, environmental analysis, military reconnaissance, and geological exploration. However, the SAR imaging principle inherently results in strong speckle noise and geometric distortion in SAR images, posing challenges to their interpretation. Automatic Target Recognition (ATR), which integrates image detection, feature extraction, and image recognition processes, is one of the key technologies for achieving automatic SAR image interpretation.

[0003] Image recognition, as the final crucial step in ATR (Automatic Tracing), has always been a research hotspot in SAR imagery. Traditional recognition methods mainly consist of two steps: feature extraction and recognition. Commonly used features include geometric features, projection features, and scattering features. Then, machine learning algorithms are used for classification based on the extracted features. The main classification algorithms include K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and sparse representation classifiers. Traditional SAR image recognition methods have achieved good results and brought some progress to SAR ATR research. However, these methods rely on tedious, manually extracted features by experts, resulting in low efficiency and poor robustness. Therefore, feature extraction algorithms need to be redesigned for different SAR datasets and application scenarios.

[0004] In recent years, with the development of deep learning technology, especially convolutional neural networks, image recognition technology research has made groundbreaking progress. The proposal and successful application of AlexNet marked the beginning of the development of deep networks, followed by the well-known VGGNet and RestNet, which achieved breakthroughs in natural image recognition. With the achievements and influence of deep learning in optical images, researchers have attempted to apply deep learning to SAR image recognition. SAR image recognition methods based on deep convolutional networks are an "end-to-end" learning approach, unifying feature extraction and classification. The network automatically learns the required features according to the learning objective, eliminating the need for additional feature extraction algorithms, reducing manual work, and greatly improving the robustness of the algorithm. However, deep learning algorithms rely on large amounts of training data, while SAR image training data is more difficult to obtain than optical image training data, with a relatively smaller sample size. Directly applying ordinary image recognition algorithms to SAR image recognition results in low performance. How to design deep learning algorithm models and training methods to improve the recognition accuracy in SAR image recognition scenarios is an urgent problem to be solved.

[0005] Chinese invention patent "SAR Image Small Sample Recognition Method and Apparatus Based on Weighted Distance (Application No. 202210618273.7)" improves the recognition accuracy of SAR images by utilizing weighted sample-prototype distance and meta-learning methods. Chinese invention patent "SAR Target Recognition Method, Apparatus, Device and Medium Based on Prototype Comparative Learning (Application No. 202310919952.2)" performs two types of data augmentation on each sample, obtains sample features, uses clustering to obtain sample prototypes, compares the distance between samples and prototypes to calculate the loss, and considers prototype density. Both patents use distance metric algorithms, but only consider the distance from the sample to the prototype, neglecting the distance between prototypes themselves, thus failing to fully utilize various factors, and the training effect needs improvement. Summary of the Invention

[0006] The purpose of this invention is to provide a SAR image recognition system and method based on distance metrics and fully convolutional networks. This invention comprehensively utilizes the sample-to-class center distance, the variance of samples within the same class, and the distance between class centers. Furthermore, by designing a feature extraction network composed of multiple small-sized convolutional layers and a 1x1 convolutional classifier, and employing a two-stage training method, the accuracy of SAR image recognition is ultimately improved.

[0007] To achieve this objective, the SAR image recognition system based on distance metric and fully convolutional network designed in this invention is characterized by comprising a feature extraction network training module, a model training module, and a synthetic aperture radar image recognition module.

[0008] The feature extraction network training module is used to perform data augmentation on the synthetic aperture radar image training data, and input the data augmented synthetic aperture radar image training data into the feature extraction network to obtain feature vectors. The distance loss of the feature vectors is calculated, the distance loss is backpropagated, and the parameters of the feature extraction network are updated using an optimization algorithm to obtain the trained feature extraction network.

[0009] The model training module is used to add the classifier to the trained feature extraction network to form a synthetic aperture radar (SAR) image recognition model. It performs data augmentation on the SAR image training data, inputs the augmented SAR image training data into the SAR image recognition model, calculates the cross-entropy loss between the output data and the data labels of the SAR image recognition model, backpropagates the cross-entropy loss, and updates the SAR image recognition model parameters using an optimization algorithm to obtain the trained SAR image recognition model.

[0010] The synthetic aperture radar image recognition module is used to input the synthetic aperture radar image to be recognized into the trained synthetic aperture radar image recognition model to obtain the synthetic aperture radar image recognition result.

[0011] The beneficial effects of this invention are:

[0012] This invention constructs a convolutional network based on small-sized convolutional kernels, reducing the number of network parameters and making it more suitable for SAR image recognition scenarios where the number of samples is much smaller than that of optical images. A distance-based loss function is proposed, which calculates the final loss by combining the distance between samples and class centers, the variance of samples within the same class, and the distance between class centers, helping the model converge quickly towards intra-class aggregation and inter-class dispersion. A convolutional network classifier is constructed to reduce model parameters, prevent overfitting, and improve model robustness and performance. A two-stage training method is proposed: the first stage uses the distance-based loss function to train the feature extraction network, and the second stage uses cross-entropy for overall training. This training method helps the model distinguish between classes more clearly and has stronger classification capabilities. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the structure of the present invention;

[0014] Figure 2 This is the overall framework diagram;

[0015] Figure 3 Here is a diagram of the feature extraction network and classifier structure;

[0016] Figure 4 This diagram illustrates the impact of the ratio of inter-class distance to class variance on the discriminative power of the data.

[0017] Figure 5 The confusion matrix represents the experimental results on the MSTAR dataset.

[0018] Figure 6a The visualization of the features output by the model trained using the prototype network loss function after PCA dimensionality reduction;

[0019] Figure 6b Visualization of the features output by the model trained using the loss function of a supervised contrastive learning network after PCA dimensionality reduction.

[0020] Figure 6c Visualization of the features output by the model trained with the triplet loss function after PCA dimensionality reduction.

[0021] Figure 6d The visualization result of the features output by the model trained by the loss function of this invention after PCA dimensionality reduction.

[0022] Figure 7a The result is the dimensionality reduction of the features output by the model trained using distance metric after PCA.

[0023] Figure 7bThe result is the dimensionality reduction of the features output by the model trained without using distance metrics, after PCA. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0025] Example 1

[0026] like Figure 1 and 2 The SAR image recognition system shown is based on distance metric and fully convolutional network, and includes a feature extraction network training module, a model training module, and a synthetic aperture radar image recognition module.

[0027] The feature extraction network training module is used to perform data augmentation on the synthetic aperture radar (SAR) image training data (the training data consists of more than 2,000 SAR images). The augmented SAR image training data is then input into the feature extraction network to obtain feature vectors (the feature vectors are the output of the image after passing through the feature extraction network). The distance loss of the feature vectors is calculated, and the distance loss is backpropagated. The parameters of the feature extraction network are then updated using the AdamW (Adaptive Moment Estimation with Weight Decay) algorithm to obtain the trained feature extraction network. For details on the specific method of updating the feature extraction network parameters using the AdamW algorithm, please refer to the reference Loshchilov I, Hutter F. Decoupled Weight Decay Regularization[J]. 2017. DOI:10.48550 / arXiv.1711.05101.

[0028] The model training module adds a classifier to the trained feature extraction network to form a synthetic aperture radar (SAR) image recognition model (data passes through the feature extraction network and then the classifier). It augments the SAR image training data and inputs the augmented data into the SAR image recognition model. Each image in the SAR image recognition model's output data is assigned a vector indicating the probability of each sample. The module calculates the cross-entropy loss between the SAR image recognition model's output data and the data labels, indicating the accuracy of the current model's output. The cross-entropy loss is then backpropagated and the AdamW algorithm is used to update the SAR image recognition model parameters (all model parameters are updated, iterating in a direction that gradually reduces the loss until the optimal value is reached). This results in a trained SAR image recognition model. Since the model has already undergone feature extraction network training, the main purpose of this module is to fine-tune the feature extraction network parameters and optimize the classifier parameters to improve the model's classification accuracy.

[0029] The synthetic aperture radar image recognition module is used to input the synthetic aperture radar image to be recognized into the trained synthetic aperture radar image recognition model to obtain the synthetic aperture radar image recognition result.

[0030] In the above technical solution, the feature extractor is trained first instead of the entire model at once. Distance metric loss is used to increase the distance between categories, making the network more able to distinguish between image feature representations of different categories and improving its generalization ability. This makes the feature extraction network more capable of representing various types of images, thus improving the performance of subsequent image recognition models.

[0031] In the above technical solution, the data label is the specific category to which the sample belongs. It corresponds to the training data. For example, image 1, category 1, category 1 is the label of image 1, representing the true category of the image.

[0032] In the above technical solution, the feature vector is input into the classifier, such as... Figure 3 As shown, image recognition can be performed. The classifier outputs the probability that each sample (each image) belongs to each category (the category of the image). The classifier structure is a 1X1 convolutional network with dimensions from 512 to 10.

[0033] In the above technical solution, the specific method for data augmentation of the synthetic aperture radar image training data by the feature extraction network training module is as follows: the synthetic aperture radar image training data is centered and cropped to an 88x88 size (the original data has multiple sizes, but the model currently accepts images of a uniform size. To ensure that the feature size of the last layer of the feature extraction network is exactly 1x1, the input image size is 88x88 according to the model structure derivation). The cropped synthetic aperture radar image training data is then randomly flipped vertically with a 50% probability. The vertically flipped synthetic aperture radar image training data is then randomly flipped horizontally with a 50% probability. The horizontally flipped synthetic aperture radar image training data is then randomly rotated by 20 degrees to randomly deform the image, thereby increasing the diversity of the training data. The randomly rotated synthetic aperture radar image training data is then normalized to a mean of 0.485 and a standard deviation of 0.229.

[0034] In the above technical solution, the specific method for data augmentation of the synthetic aperture radar (SAR) image training data by the model training module is as follows: the SAR image training data is centered and cropped to an 88x88 size; the cropped SAR image training data is then randomly flipped vertically with a 50% probability; the vertically flipped SAR image training data is then randomly flipped horizontally with a 50% probability; the horizontally flipped SAR image training data is then randomly rotated by 20 degrees; and the randomly rotated SAR image training data is then normalized to a mean of 0.485 and a standard deviation of 0.229. Through data augmentation, the diversity of the data is increased, the generalization ability of the model is enhanced, and its representational ability is strengthened.

[0035] Data augmentation must be performed before each training session. Since data augmentation involves random factors, the data obtained after each augmentation of the same batch of data will differ, thereby increasing data diversity and improving the model's generalization ability.

[0036] In the above technical solution, the structure of the feature extraction network is shown below. Figure 3 The model comprises five convolutional blocks, each consisting of two 3x3 convolutional layers, a batch normalization layer, a ReLU activation layer, and a max-pooling layer. The pooling layer uses a 2x2 kernel with a stride of 2, thereby reducing the feature map size to half its original size. Except for the first convolutional block (the first of the five convolutional blocks is the first one in the sequence), the first convolutional layer of each of the other convolutional blocks doubles the number of channels. By stacking multiple small 3x3 convolutional modules, the model achieves the same receptive field as large convolutional modules while reducing the number of parameters. This reduces the risk of overfitting and improves the model's generalization ability in SAR image recognition scenarios with fewer samples.

[0037] In the above technical solution, the method for calculating the pre-training loss function of the feature extractor comprehensively considers three aspects: the distance from the sample to the class center, the distance from the class center to the center, and the class variance, thereby guiding the model to train in the direction of intra-class aggregation and inter-class dispersion. The distance loss of the feature vector is... distance The specific calculation method is as follows:

[0038] loss distance =α·loss s-c +β·loss var +(1-α-β)·loss c-c

[0039] Where, loss s-c The loss is the distance from the sample to the class center in the synthetic aperture radar image training data. var For the variance loss of similar samples, loss c-c For the inter-class center distance loss, α and β are two hyperparameters of the distance metric method. α is the proportion of sample-to-center distance loss, and β is the proportion of variance loss for samples within the same class. The above design comprehensively guides the model training towards intra-class aggregation and inter-class dispersion by integrating three aspects of loss, considering both the relationship between samples and class centers, as well as the relationship between class centers.

[0040] In the above technical solution, the variance loss of similar samples is... var The calculation method is as follows:

[0041]

[0042] var j The variance of the j-th type of sample is calculated as follows:

[0043]

[0044] Where X(j)≡{x i :i∈I(j)} is the set of samples of the j-th class, x i Let I(j) be the sample numbered i, which can generally be considered as the i-th sample. I(j) is the set of sample numbers for the j-th class. var(.) is the function for calculating variance. m is a variable in the formula that varies from 0 to M-1, where M is the dimension of the feature vector. f θ (X(j)) [m] Let m be the m-th dimension of the feature vector of the j-th class of samples.

[0045] In the above technical solution, the distance loss from the sample to the class center in the synthetic aperture radar image training data is... s-cThe distance loss from samples to their class centers is significant. When a sample's feature vector is far from its class center, a large loss occurs. The model optimizes in the direction that minimizes this loss. Therefore, this portion of the loss motivates the model to optimize and bring each sample closer to its class center. It is calculated by averaging the distance loss from each sample to the center, using the following formula:

[0046]

[0047] Wherein, loss(x) i ) represents the distance loss from the i-th sample to the class center, and N is the number of samples in this batch;

[0048] loss(x i The calculation formula is:

[0049]

[0050] Among them, y i ∈{0...L-1} is a sample x i Category, Dis(x) i ,c l Let be the distance from the i-th sample to the center of the l-th class. In this embodiment, Euclidean distance is used, where L is the number of classes and e is the natural constant. For the i-th sample x i To its corresponding tag category y i center distance,

[0051] Dis(x i ,c l The calculation formula is:

[0052]

[0053] Where M is the vector dimension, f θ (x i ) [m] For sample x i The m-th dimension, c, of the feature vector output by the feature extraction network is... l Let c be the center of class l, and let c be the average feature vector of class l in the current training batch. l[m] For c l The m-th dimension, c l The calculation formula is as follows:

[0054]

[0055] Where I≡{1...N} is the set of input sample numbers in this batch, I(l)≡{p∈I:y p=l} is the set of indices for the l-th class of samples, |I(l)| is the number of elements in the set, and c l Each batch will be updated.

[0056] In the above technical solution, loss c-c For the inter-class center distance loss, the larger the inter-class distance of the feature vectors after training, the better the classification performance of the model. However, the absolute distance cannot reflect the degree of distinction between classes. This invention compares the distance between class centers with the variance of class samples; the larger the value, the greater the difference between classes. See the schematic diagram. Figure 4 , Figure 4 For the loss function c-c Design intent, Figure 4 a and Figure 4 In case b, the class center distance is the same, but Figure 4 Classification performance is good, and the inter-class center distance loss is... c-c The calculation method is as follows:

[0057]

[0058] Among them, c j For the j-th class of sample centers, var i Let var be the variance of the i-th class of samples. j Let Dis(c) be the variance of the j-th class of samples. i ,c j Let be the Euclidean distance between the centers of classes i and j, and L be the number of classes. The more compact the samples of the same class are and the greater the distance between classes, the smaller the loss; conversely, the greater the distance between classes, the larger the loss. The model training process optimizes towards minimizing the loss. The loss is calculated by comparing the distance between class centers with the variance of the class samples, rather than simply comparing the distance between class centers. This results in a wider class boundary after training and a stronger generalization ability of the model.

[0059] Example 2

[0060] A SAR image recognition method based on distance metric and fully convolutional network, characterized by comprising the following steps:

[0061] Step 1.1: Perform data augmentation on the synthetic aperture radar image training data;

[0062] Step 1.2: Input the augmented synthetic aperture radar image training data into the feature extraction network to obtain the feature vector;

[0063] Step 1.3: Calculate the distance loss of the feature vector;

[0064] Step 1.4: Backpropagate the distance loss and update the feature extraction network parameters using the AdamW algorithm;

[0065] Step 1.5: Repeat steps 1.1 to 1.4 until all synthetic aperture radar image training data have been traversed and 200 iterations have been completed (each iteration completes the traversal of all training data, and multiple rounds of iterative training continuously optimize the parameters of the feature extraction network, eventually converging to a stable value), to obtain the trained feature extraction network.

[0066] Step 2.1: Add the classifier to the trained feature extraction network to form a synthetic aperture radar image recognition model;

[0067] Step 2.2: Perform data augmentation on the synthetic aperture radar image training data;

[0068] Step 2.3: Input the enhanced synthetic aperture radar image training data from Step 2.2 into the synthetic aperture radar image recognition model;

[0069] Step 2.4: Calculate the cross-entropy loss between the output data and data labels of the synthetic aperture radar image recognition model;

[0070] Step 2.5: Backpropagate the cross-entropy loss and update the parameters of the synthetic aperture radar image recognition model using the AdamW algorithm;

[0071] Step 2.6: Repeat steps 2.2 to 2.5 until all synthetic aperture radar image training data have been traversed and 150 iterations have been completed to obtain the trained synthetic aperture radar image recognition model.

[0072] Step 3.1: Input the synthetic aperture radar image to be identified into the trained synthetic aperture radar image recognition model to obtain the synthetic aperture radar image recognition result.

[0073] Step 2.4 uses the following cross-entropy formula to calculate the cross-entropy loss. cross_entropy :

[0074]

[0075] Where N is the number of samples in this batch, g(x) i ) is the sample x i The output of the classifier has a dimension of L (the dimension of the classifier is the number of classes), g(x) i ) [l] It is g(x) i The l-th dimension of y i ∈{0...L-1} is a sample x i Category label, g(x) i )[y i ] is g(x i Take the y-th... i Dimension.

[0076] Example 3

[0077] A computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method.

[0078] Example 4

[0079] This embodiment illustrates the technical effects of the present invention through the following experiments:

[0080] 1. Experimental conditions

[0081] The following experiments all use the MSTAR dataset to verify the effectiveness of this invention. This invention selects 10 classes of ground target data from the Standard Operating Environment (SOC) of the MSTAR dataset for experiments. The experimental data are shown in Table 1 below. The experiments were conducted on the Ubuntu 20.04 LTS operating system, using the PyTorch platform with an RTX 3090 graphics card. The experimental parameters were configured as follows: the batch size for both the feature extraction network pre-training and the SAR image recognition model fine-tuning training was 100; the learning rate of the Adamw optimizer was 0.001, the weight decay was 0.004, and the hyperparameters α and β of the distance metric loss function were set to 0.8 and 0.1, respectively.

[0082] Table 1 MSTAR SOC Training and Test Sets

[0083]

[0084] 2. Experiment Content

[0085] Experiment 1

[0086] The SAR image recognition model, consisting of a feature extraction network and a classifier, was trained using the training set in Table 1. The performance of the trained model was then tested using the test set. The confusion matrix results from the test set are shown below. Figure 5 , Figure 5 The results were used to verify the effectiveness of the invention on a public dataset. It can be seen that the recognition accuracy for most categories is very high, reaching 100%, and the average recognition accuracy is 99.67%, which proves the effectiveness of this model.

[0087] Experiment 2

[0088] To further demonstrate the effectiveness of the algorithm proposed in this invention, the proposed model and references are provided in the following documents: “S. Chen, H. Wang, F. Xu, and YQ Jin, “Target classification using the deep convolutional networks for SAR images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 8, pp. 4806–4817, 63” (hereinafter referred to as AConvnet); “J. Guan, J. Liu, P. Feng, and W. Wang, “Multiscale deep neural network with two-stage loss for SAR target recognition with small training set,” IEEE Geoscience and Remote Sensing Letters, pp. 1–5, 2021” (hereinafter referred to as MultiscaleNet); and “D. Cheng, Y. Gong, S. Zhou, J. Wang, and N. Zheng, “Person re-identification by multi-channel parts-based CNN with improved triplet loss function,” Proceedings of the IEEE conference on computer vision and pattern recognition.” The model was compared with the classic image recognition convolutional networks VGG16, ResNet-50, and ResNet-18. The comparative experimental results are shown in Table 2. The results of AConvnet and MultiscaleNet algorithms were taken from the original data in the literature. The results of VGG16, ResNet-50, and ResNet-18 were obtained by training with the same parameters as the model in this invention. The results of Improved Triplet Loss were obtained by replacing the distance metric method in this invention with the distance metric method in the Improved Triplet Loss paper. All other conditions were the same as in this invention. As can be seen from the table, VGG16 performed the worst, as it used multiple fully connected layers and had a large number of parameters. ResNet-50 was the second worst, also due to its large number of layers and parameters.Acquiring training data for SAR images is more difficult than for optical images, and the number of samples is relatively small. Models with a large number of parameters are prone to overfitting and performance degradation. Compared with the former two, ResNet-18, AConvnet, MultiscaleNet, Improved Triplet Loss, and the model in this paper have fewer parameters, so their performance is better, reaching over 99% in SOC. Among them, the algorithms MultiscaleNet, Improved Triplet Loss, and the method of this invention all further use distance metric algorithms, thus further improving accuracy.

[0089] Table 2. Experimental results comparing the present invention with algorithms in other literature.

[0090]

[0091]

[0092] Experiment 3

[0093] To demonstrate the superiority of the feature extraction network in this invention, this experiment replaced the feature extraction network with EfficientNetV2 and MobileNetV3, and conducted experiments under the same conditions. The comparison results are shown in Table 3. It can be seen that the feature extraction network of this invention has a significant advantage in recognition rate compared to the other two. This is because the feature extraction network of this invention uses small convolutions and has fewer parameters. In SAR image recognition scenarios with few samples, it avoids overfitting, resulting in better model training performance.

[0094] Table 3. Comparison of experimental results for different feature extraction networks.

[0095]

[0096] Experiment 4

[0097] To demonstrate the superiority of the distance metric loss function in this invention, the loss function trained in the feature network stage was replaced with the loss functions of the prototype network, the supervised contrastive learning network, and the triplet loss function. Experiments were conducted under the same conditions, and the comparison results are shown in Table 4. It can be seen that the algorithm in this paper outperforms the other three methods. This is because the prototype network only uses the distance between a sample and the center of a sample of the same class (the prototype), the supervised contrastive learning network only considers the distance between a sample and other samples of the same class, and the triplet loss function only considers the distance between samples of the same class and samples of different classes. All three only consider the relationship between samples and other samples. This invention, in addition to considering the relationship between samples, further considers the internal variance of samples of the same class and the distance between the centers of samples of different classes, while also considering both individual and overall losses of samples. This provides a more comprehensive guide to optimize the model in a direction that clusters samples of the same class and disperses samples of different classes.

[0098] Table 4. Comparison of experimental results with different loss functions

[0099]

[0100] Figure 6 shows a visualization of the dimensionality reduction results of the feature extraction networks trained with different algorithms on the test set after PCA. Each color represents a class, and high class clustering indicates good classification performance. Figure 6a The model trained using the prototype network loss function. Figure 6b For models trained using the loss function of supervised contrastive learning networks, Figure 6c The model is trained using the triplet loss function. Figure 6d The model trained by the algorithm of this invention shows that the algorithm of this invention distinguishes each category more clearly and has a stronger classification ability.

[0101] Experiment 5

[0102] To verify the advantages of the proposed model in terms of parameter quantity and performance, the parameter quantity of related algorithms was calculated. Table 5 compares the proposed algorithm with other algorithms in terms of parameter quantity, floating point operations (FLOPs), and accuracy. It can be seen that the proposed algorithm has the fewest parameters, with a computational load only higher than MobileNetV3_S, but the highest accuracy. While ensuring a high recognition rate, the overall algorithm complexity is low.

[0103] Table 5 Comparison of Algorithm Complexity

[0104] VGG16 134.3004M 2,442,049,536 96.37％ ResNet-50 23.5222M 685,958,400 98.27％ ResNet-18 11.1754M 292,672,256 99.01％ EfficientNetV2_S 20.1899M 484,495,360 97.90％ MobileNetV3_S 2.5426M 11,515,184 98.52％ This invention model 1.7757M 111,825,172 99.67％

[0105] Experiment 6

[0106] To verify the effectiveness of the three design methods proposed in this invention—the convolutional classifier, the small-sized convolutional kernel, and the distance metric loss function—this paper combines the three methods (denoted as A, B, and C) into several variants: Variant 1: using a linear layer as the final classifier; Variant 2: using a 7x7 convolutional kernel instead of three 3x3 convolutional kernels, and using a 5x5 convolutional kernel instead of two 3x3 convolutional kernels; Variant 3: not using the distance metric loss function, but directly using cross-entropy for one-stage training. The experimental results of the three variants are shown in Table 6. It can be seen that all three variants have varying degrees of accuracy loss. Variant 1 has the worst performance, followed by the large-sized convolutional kernel, then the model without using the distance metric, and the best performance is the model without any changes. This proves the effectiveness of the convolutional classifier, the distance metric loss function, and the small-sized convolutional kernel. Furthermore, the experimental results show that the accuracy of variant 1 and variant 2 decreased. This is because using a linear classifier and a large convolutional kernel both increased the number of parameters in the model to some extent, leading to overfitting. This invention simplifies the network to reduce the number of parameters, thereby avoiding overfitting to some extent.

[0107] Table 6 Ablation experiments for various variant models

[0108]

[0109] Figure 7 shows the visualization results of the features output by the model after dimensionality reduction using PCA. Figure 7a These are features generated by a model trained using a distance metric loss function. Figure 7b The figures show the features generated by the model trained without using a distance metric function. As can be seen from the figures, after training with a distance metric loss function, the categories are more clearly distinguished, and the classification ability is stronger. This is because the distance metric algorithm of this invention compares the feature distances between samples of the same class, samples of different classes, and class centers, guiding the model to converge in the direction of intra-class aggregation and inter-class dispersion.

[0110] The experimental results above demonstrate that the algorithm of this invention exhibits more clustered features among similar samples, clearer distinctions between classes, and stronger classification capabilities. The method proposed in this paper is superior to methods proposed in other papers and classic image recognition models, thus validating the benefits of this invention.

[0111] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A SAR image recognition system based on distance metric and fully convolutional network, characterized in that: It includes a feature extraction network training module, a model training module, and a synthetic aperture radar image recognition module; The feature extraction network training module is used to perform data augmentation on the synthetic aperture radar image training data, and input the data augmented synthetic aperture radar image training data into the feature extraction network to obtain feature vectors. The distance loss of the feature vectors is calculated, the distance loss is backpropagated, and the parameters of the feature extraction network are updated using an optimization algorithm to obtain the trained feature extraction network. The model training module is used to add the classifier to the trained feature extraction network to form a synthetic aperture radar (SAR) image recognition model. It performs data augmentation on the SAR image training data, inputs the augmented SAR image training data into the SAR image recognition model, calculates the cross-entropy loss between the output data and the data labels of the SAR image recognition model, backpropagates the cross-entropy loss, and updates the SAR image recognition model parameters using an optimization algorithm to obtain the trained SAR image recognition model. The synthetic aperture radar image recognition module is used to input the synthetic aperture radar image to be recognized into the trained synthetic aperture radar image recognition model to obtain the synthetic aperture radar image recognition result; Distance loss of the feature vector The specific calculation method is as follows: in, The distance loss from samples to class centers in the synthetic aperture radar image training data. For the variance loss of similar samples, For the distance loss between class centers, and These are two hyperparameters of the distance metric method. It represents the percentage of loss due to distance from the sample to the center. Percentage of variance loss for similar samples; The same sample variance loss The calculation method is as follows: The variance of the j-th type of sample is calculated as follows: in, It is the set of samples of class j. x i It is numbered i The sample, I(j) It is the first j Set of sample IDs for each class It is a function for calculating variance, where m is a variable ranging from 0 to M-1 in the formula, and M is the dimension of the feature vector. Let m be the m-th dimension of the feature vector of the j-th class of samples; The inter-class center distance loss The calculation method is as follows: in, For the j-th type of sample center, Let be the variance of the i-th type of sample. Let be the variance of the j-th type of sample. Let L be the Euclidean distance between the centers of classes i and j, and L be the number of classes.

2. The SAR image recognition system based on distance metric and fully convolutional network according to claim 1, characterized in that: The specific method for data augmentation of synthetic aperture radar image training data by the feature extraction network training module is as follows: the synthetic aperture radar image training data is centered and cropped to a preset size; the cropped synthetic aperture radar image training data is randomly vertically flipped; the vertically flipped synthetic aperture radar image training data is randomly horizontally flipped; the horizontally flipped synthetic aperture radar image training data is randomly rotated; and the randomly rotated synthetic aperture radar image training data is normalized.

3. The SAR image recognition system based on distance metric and fully convolutional network according to claim 1, characterized in that: The specific method for data augmentation of the synthetic aperture radar image training data by the model training module is as follows: the synthetic aperture radar image training data is cropped to a preset size by centering; the cropped synthetic aperture radar image training data is randomly vertically flipped; the vertically flipped synthetic aperture radar image training data is randomly horizontally flipped; the horizontally flipped synthetic aperture radar image training data is randomly rotated; and the randomly rotated synthetic aperture radar image training data is normalized.

4. The SAR image recognition system based on distance metric and fully convolutional network according to claim 1, characterized in that: The feature extraction network includes 5 convolutional blocks. Each convolutional block includes two 3x3 convolutional layers, one batch normalization layer, one ReLU activation layer, and one max-pooling layer. The pooling layer uses a 2x2 kernel size with a stride of 2. Except for the first convolutional block, the first convolutional layer of each of the remaining convolutional blocks doubles the number of channels.

5. The SAR image recognition system based on distance metric and fully convolutional network according to claim 1, characterized in that: Distance loss from sample to class center in synthetic aperture radar image training data The distance loss from each sample to its respective center is calculated by averaging the distance loss from each sample to the center. The formula is as follows: in, The distance loss is the distance from the i-th sample to the class center, and N is the number of samples in this batch; The calculation formula is: in, It is a sample Category For the i-th sample to the i-th Class center distance, where L is the number of classes and e is the natural constant. For the i-th sample To its corresponding tag category center distance, The calculation formula is: Where M is the vector dimension, For the sample The m-th dimension of the feature vector output by the feature extraction network For the first The center of the class is the first in the current training batch. The average value of the eigenvectors of the class. for The m-th dimension, The calculation formula is as follows: in, This is the set of sample IDs for the input samples in this batch. It is the first The set of sample IDs for each class. The number of elements in the set. Each batch will be updated.

6. A SAR image recognition method based on distance metric and fully convolutional network according to claim 1, characterized in that, It includes the following steps: Step 1.1: Perform data augmentation on the synthetic aperture radar image training data; Step 1.2: Input the augmented synthetic aperture radar image training data into the feature extraction network to obtain the feature vector; Step 1.3: Calculate the distance loss of the feature vector; Step 1.4: Backpropagate the distance loss and update the feature extraction network parameters using an adaptive learning rate algorithm; Step 1.5: Repeat steps 1.1 to 1.4 until all synthetic aperture radar image training data have been traversed and the preset number of iterations have been completed to obtain the trained feature extraction network; Step 2.1: Add the classifier to the trained feature extraction network to form a synthetic aperture radar image recognition model; Step 2.2: Perform data augmentation on the synthetic aperture radar image training data; Step 2.3: Input the enhanced synthetic aperture radar image training data from Step 2.2 into the synthetic aperture radar image recognition model; Step 2.4: Calculate the cross-entropy loss between the output data and data labels of the synthetic aperture radar image recognition model; Step 2.5: Backpropagate the cross-entropy loss and update the parameters of the synthetic aperture radar image recognition model using an adaptive learning rate algorithm; Step 2.6: Repeat steps 2.2 to 2.5 until all synthetic aperture radar image training data have been traversed and the preset number of iterations have been completed to obtain the trained synthetic aperture radar image recognition model; Step 3.1: Input the synthetic aperture radar image to be identified into the trained synthetic aperture radar image recognition model to obtain the synthetic aperture radar image recognition result.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 6.

Images