A dual-prototype-based region feature reinforced radar image open set identification method

Abstract

The application discloses a kind of based on dual prototype's regional feature reinforcement radar image open set identification method, belong to radar signal processing field.The method first proposes the feature extractor containing two modules: target area feature enhancement module is used to strengthen regional feature information, attention perception channel discarding module is used to improve structural discriminability and promote feature diversity.These components collectively enhance the distinguishing degree and robustness of feature under open set condition.Subsequently introduce dual prototype loss function, wherein prototype represents known class feature, complementary point predicts unknown class distribution, to effectively separate known and unknown feature space.To avoid misclassification, two kinds of regularization constraint mutual exclusion point distribution are used;Finally, the application uses the maximum-second largest class probability difference and preset threshold to compare to confirm whether the radar target sample to be measured belongs to known class.Compared with some advanced open set identification methods, the application realizes more excellent open set identification performance.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, specifically relating to a method for open set recognition of radar images based on dual-prototype regional feature enhancement. Background Technology

[0002] Automatic target recognition (ADR) of radar targets is a research hotspot in the field of radar applications, playing a vital role in disaster early warning and geological exploration. In recent years, ADR has been developing towards deep learning, but most research focuses on closed-set recognition, meaning it doesn't consider the existence of unknown radar targets, which often doesn't reflect reality. Therefore, research on open-set recognition methods for radar images is of great significance.

[0003] Open set recognition specifically refers to a recognition model that not only classifies known categories but also identifies unseen categories. In deep learning, this can be interpreted as the existence of categories in the test set that are not present in the training set. E. Giusti et al. first applied the OpenMax classifier to open set recognition of radar image targets, but this method requires a large number of radar image samples of known classes, which is a challenge in practice. G. Chen et al. proposed the concept of complementary points, which accurately predicted the distribution of unknown classes in the feature space, but this method emphasizes increasing the distance between complementary points and their corresponding classes, ignoring the compactness of intra-class distribution. H. Yang et al. introduced a prototype network into open set recognition, but this method could not effectively predict the distinction of unknown classes. Summary of the Invention

[0004] To address the above issues, this invention provides a dual-prototype regional feature enhancement radar image open set recognition method, which can extract more discriminative features and aggregate known radar targets while accurately predicting the feature distribution of unknown radar targets, thereby achieving excellent radar image target open set recognition performance.

[0005] The technical solution adopted in this invention is: a method for open set recognition of radar images based on dual-prototype regional feature enhancement, the specific steps of which are as follows:

[0006] Step 1: Obtain training samples of radar images with known categories and test data of radar images with unknown categories. The global feature space of each sample is divided into feature space and open space.

[0007] Step 2: Use a feature extractor to extract features from the input radar image;

[0008] Step 3: Dual Prototype Learning; Obtain the prototype and complementary points of the radar target, with the complementary points serving as the non-first prototype. The centers of known and unknown radar image samples;

[0009] The distance from the feature extractor output of the radar image sample to its corresponding complementary point and prototype is calculated. The dual prototype loss function is used to make the radar image sample move closer to its corresponding prototype while moving away from its corresponding complementary point.

[0010] Furthermore, a loss function is established to constrain the range of the open space; at the same time, the loss function is constrained by complementary points to avoid the uneven distribution of complementary points causing unknown radar targets to be misclassified as known targets;

[0011] Step 4: Use the trained model to identify unknown radar targets; and set a threshold for the difference between the probability of the largest class and the probability of the second largest class of radar image samples to distinguish between unknown and known classes.

[0012] The feature extractor includes a target region feature enhancement module and an attention-aware channel discarding module, as detailed below:

[0013] Target region feature enhancement module: A set of filters is applied to the input radar image to generate a binary space mask that highlights the target region; the input radar image is passed through a convolutional layer to obtain a feature map, and the binary space mask is adjusted to match the spatial resolution of the feature map to obtain an excitation map; the excitation map and the feature map are added using residual addition to obtain the enhanced feature map;

[0014] The enhanced feature map output by the target region feature enhancement module is passed through a convolutional layer and then used as the input to the attention-aware channel discarding module.

[0015] Attention-aware channel drop-off module: First, the average response of the channel dimension is calculated by global average pooling. Then, the channel response vector in each sample is standardized to obtain the attention score of each channel. Then, a binary drop-off mask is generated based on the attention score, and the maximum drop-off ratio is set. The binary drop-off mask is adjusted to be consistent with the shape of the input of this module, and the output of the attention-aware channel drop-off module is obtained by element-wise multiplication.

[0016] Finally, after average pooling and random dropout functions, the output of the feature extractor is obtained.

[0017] Step 3 is as follows:

[0018] The prototype of the k-th type of radar target is represented as: Complementary points are represented as ,in and These represent the number of prototypes and complementary points, respectively.

[0019] Class k radar image samples The distance to its corresponding complementary point and prototype is expressed as:

[0020] ;

[0021] in, This represents the distance between two variables. This represents the output of the feature extractor, which has learnable parameters. , Represents the square of the Euclidean distance;

[0022] Therefore, the biprototype loss function The mathematical expression is:

[0023] ;

[0024] in, K is the number of known categories. It has a positive temperature coefficient. Indicates the sample index. It is the first Complementary points It is the first Class prototype;

[0025] Loss function of constrained open space Represented as:

[0026] ;

[0027] in It is a learnable parameter;

[0028] Complementary point constraint loss function The expression is as follows:

[0029] ;

[0030] in .

[0031] Therefore, the total loss function is expressed as:

[0032] ;

[0033] in and It's a hyperparameter.

[0034] The criteria for distinguishing between unknown and known classes are as follows:

[0035] ;

[0036] in It is a test radar image sample Classified as the first The probability of a class and These represent the probability of the largest class and the probability of the second largest class, respectively. It is a preset threshold. It is a predicted label.

[0037] The beneficial effects of the present invention are as follows: The dual-prototype regional feature enhancement radar image open set recognition method of the present invention accurately predicts the distribution of unknown radar targets in the feature space, effectively separates the known space and the unknown space, and achieves excellent radar image open set recognition effect. Attached Figure Description

[0038] Figure 1 This is a block diagram of the feature extractor proposed in this invention.

[0039] Figure 2 A schematic diagram of the dual-prototype learning proposed in this invention.

[0040] Figure 3 The SAR image used in the specific implementation method of this invention.

[0041] Figure 4 The confusion matrix obtained by the specific implementation method of this invention.

[0042] Figure 5 The feature distribution maps of known and unknown radar targets obtained by the specific implementation method of this invention.

[0043] Figure 6 The receiver operating curves (ROC) obtained by the specific implementation method of this invention. Detailed Implementation

[0044] This invention is primarily verified through simulation experiments. All steps and conclusions were verified correctly on the Windows 11 operating system platform using the PyTorch deep learning framework and the Stochastic Gradient Descent (SGD) optimizer. The invention is further elaborated below with reference to the accompanying drawings.

[0045] A method for open-set recognition of radar images based on dual-prototype region feature enhancement includes the following steps:

[0046] S1. Problem modeling for open set target recognition in radar images;

[0047] In radar image target open set identification tasks, given a set of... Training samples of radar images with known categories and labels. and a group of Radar image test data of unknown category ,in It is the i-th sample The tag, and These represent the total number of radar image samples of known class and unknown class, respectively.

[0048] Let the known radar image sample of the k-th class be represented as... The feature space and the corresponding open space are respectively represented as: and ,in yes A global feature space of dimension 1. The open space of the k-th class of radar image samples can be further divided into: and ,in It is the positive open space of radar image samples of class k, and is the known class of radar image samples that are not class k. The feature space in which it is located It is the negative open space of the k-th type of radar image sample, and it is the radar image sample of the potential unknown class. The feature space in which it resides.

[0049] For a multi-class radar image open set recognition problem, the overall objective is to optimize a multi-class classification function. To minimize risk Then the problem can be expressed as:

[0050]

[0051] in and These represent risk categorized by experience and risk in open spaces, respectively. It is a regularization coefficient. Solving the problem of open set recognition of multi-class radar images can be equated to jointly optimizing the empirical classification loss and the open space risk.

[0052] S2, Feature Extractor:

[0053] Figure 1 The framework of the proposed feature extractor is shown. The feature extractor adds a target region feature enhancement module and an attention-aware channel discarding module to the convolutional layer, as detailed below:

[0054] Target Region Feature Enhancement Module: This module first applies a set of Gabor filters to the input radar image. The filtered responses are then fused to generate a binary space mask that highlights the target region. This mask will be used to modulate intermediate feature maps during training. Given an input image... The mask can be obtained through the following operation:

[0055]

[0056] in express Gabor filter response in the direction, Indicates the direction angle. The set of azimuth angles is represented by X, where X represents the input radar image. It is an indicator function. It is a threshold.

[0057] The mask is then adjusted to match the feature map obtained by passing the input radar image through the first convolutional layer. The spatial resolution is matched, where B represents the batch size, C represents the number of channels, H represents the feature map height, and W represents the feature map width. These values ​​are then applied to the channel-averaged features to obtain the final excitation map E.

[0058]

[0059] This is an element-wise multiplication; to ensure dynamic adjustment of feature enhancement strength during training, while avoiding the model from over-focusing on target region features, a time-varying coefficient is introduced. Its definition is as follows:

[0060]

[0061] in t represents the total number of training epochs, and t represents the current training epoch. The enhanced feature map is finally obtained through residual addition. :

[0062]

[0063] Attention-aware channel drop-off module: Given an input feature map :

[0064]

[0065] in Indicates the first The output of each convolutional layer, Indicates the first SiLu activation functions Indicates the first Individual batch normalization, Indicates the first One convolution, It's an index. First, the average response across the channel dimensions is calculated using global average pooling:

[0066]

[0067] in This is the channel number index.

[0068] Then, the channel response vector within each sample is standardized:

[0069]

[0070] in and They represent the average response. The mean and standard deviation. Sigmoid function. The standardized scores are mapped to the (0,1) interval, resulting in attention scores. This reflects the relative importance of each channel.

[0071] Then, a binary dropout mask is generated based on the attention score. :

[0072]

[0073] in It is a global drop intensity coefficient. This represents a uniform distribution of values ​​from 0 to 1. Attention score. The lower the value, the more likely the channel is to be dropped.

[0074] To avoid dropping too many channels, a maximum drop ratio was set. For each sample, at least one channel is retained by preserving the channel with the highest attention score. One channel, namely:

[0075]

[0076] Finally, the binary dropout mask is adjusted to match the module's input feature map. The shapes are consistent, and the module output is obtained through element-wise multiplication. :

[0077]

[0078] The final output feature vector of the feature extractor can be represented as:

[0079]

[0080] in, For average pooling function, It is a random discard function.

[0081] S3, Dual Prototype Learning:

[0082] Prototypes and complements are a pair of complementary feature representations. Specifically, a prototype is a representation of the typical features of a category by calculating the category center or selecting the category core; conversely, a complement of a category does not contain information about its own category, but rather aggregates information from all other categories to represent the center or typical features of other categories. Essentially, a complement is a special form of prototype.

[0083] The prototype of the k-th type of radar target is represented as: Complementary points are represented as ,in and These represent the number of prototypes and complementary points, respectively. Then... Can be considered as non-first Known radar image samples and unknown radar image samples The center of . Theoretically, its characteristic space Data to complementary points The distance is much greater than the negative open space. Data to The distance, at the same time Data to Prototype The distance is much smaller than Data to The distance, that is:

[0084]

[0085] in This represents the distance between two variables. Data representing negative open space distance, Data representing negative open space The distance.

[0086] On the one hand, prototype It can effectively aggregate the k-th type of radar image data However, radar image data of unknown types is lacking. The information is insufficient to accurately predict the characteristic distribution of unknown radar targets, thus easily leading to misclassification of unknown radar targets as known radar targets. On the other hand, complementary points It contains information about unknown radar targets, which can accurately predict the feature distribution of unknown targets, but it lacks information about known radar targets and cannot effectively aggregate the k-th type of radar targets.

[0087] Therefore, to better distinguish between known and unknown radar targets, a feasible approach is to maximize the distance between any sample and its corresponding complementary point, while minimizing its distance to its corresponding prototype. This helps to effectively separate the known and unknown spaces. (k-th class radar image sample) The distance to its corresponding complementary point and prototype can be expressed as:

[0088]

[0089] in, This represents the output of the feature extractor described above, which has learnable parameters. , This represents the squared Euclidean distance. Based on the above distance, this invention proposes a dual-prototype loss function. The aim is to encourage radar image samples It approaches its corresponding prototype while moving away from its corresponding complementary point. The mathematical expression is:

[0090]

[0091] in , It is a positive temperature coefficient, and i represents the sample index. It is the i-th type of complementary point. It is the i-th prototype. By minimizing the biprototype loss function, the complementary points... Extend the range of the k-th type of radar image data, while simultaneously prototyping the radar. This brings it closer. This dual effect alleviates the problem of known radar targets excessively occupying the feature space, helping to reduce the risk of empirical classification. This achieves an effective separation between enclosed and open spaces.

[0092] However, maximizing complementary points The distance to radar targets of the corresponding category causes the open space to tend towards infinity, thus increasing the risk in open space. To avoid this problem, the extent of the open space needs to be constrained, and a loss function for constraining the open space needs to be defined. It can be represented as:

[0093]

[0094] in It is a learnable parameter. Minimize The size of the open space will be minimized as much as possible. Meanwhile, to avoid the misclassification of unknown radar targets as known targets due to uneven distribution of complementary points, this invention further proposes the following complementary point constraint loss function. :

[0095]

[0096] in .

[0097] Therefore, the total loss function can be expressed as:

[0098]

[0099] in and These are hyperparameters. A schematic diagram of dual-prototype learning is shown below. Figure 2 As shown, loss item By increasing the size of the k-th type radar image sample and its corresponding complementary point The distance between them is used to separate the feature distributions of known and unknown classes, while the loss term... Limiting this distance prevents increased risks in open space. On the other hand, by reducing the size of the k-th type radar image sample compared to its prototype... The spacing between classes increases intra-class compactness, thereby improving the ability to distinguish known classes. Furthermore, This ensures that complementary points are evenly distributed, avoiding any gap between complementary points. This avoids the problem of unknown classes being misclassified as known classes due to excessive distance. The three loss functions work together to reduce the risk of empirical classification. and open space risks .

[0100] S4. Identification of Unknown Radar Targets:

[0101] Typically, when a model performs known class recognition and unknown class detection, the standard criterion for determining whether a target belongs to the unknown class is: if the probability of the largest class is lower than a preset threshold, it is classified as an unknown class. However, this criterion overlooks a crucial piece of information—unknown class samples are generally close to their complementary points or prototypes in each class. This leads to the probability of the second largest class often being close to the probability of the largest class; that is, the difference between the maximum and second largest class probabilities of the unknown class is often smaller than that of the known class. Based on this finding, this invention proposes a new discrimination criterion:

[0102]

[0103] in It is a test radar image sample Classified as the first The probability of a class and These represent the probability of the largest class and the probability of the second largest class, respectively. It is a preset threshold. It is a predicted label.

[0104] This implementation case uses the MSTAR dataset. Figure 3SAR images were presented, with three types of targets selected as unknown targets. To fully verify the recognition performance of this invention, the following aspects were analyzed. The proposed method was compared with OpenMax, Generalized Convolutional Learning (GCPL), Complementary Point Learning (RPL), Maximum Log-Score (MLS), and Category-Aware Binary Classification (CBC) methods. First, to verify the superior overall radar image target open-set recognition performance of the proposed method, this implementation compared the overall performance of the proposed method with the other methods based on five evaluation metrics: accuracy, recall, F1 score, and open-set recognition rate. As shown in Table 1, the proposed method achieved the highest values ​​in accuracy, recall, F1 score, and open-set recognition rate. Overall, the proposed method's open-set recognition performance is superior to the other three methods.

[0105] Table 1. Comparison of the open set recognition performance of the specific implementation method of the present invention with that of other methods.

[0106] method Accuracy (%) Recall rate (%) F1 score (%) Accuracy (%) OpenMax 86.23 78.47 82.01 81.90 GCPL 89.80 87.57 88.26 87.59 RPL 88.98 87.62 87.97 86.56 MLS 86.18 75.00 79.62 85.44 CBC 84.73 78.07 80.87 83.18 The proposed method 92.72 93.20 92.85 92.08

[0107] To further visualize the effectiveness of the proposed method Figure 4 and Figure 5 The confusion matrix and feature distributions of known and unknown radar targets using the proposed method are shown respectively. Figure 4 The confusion matrix also indicates that the proposed method has overlap between unknown and known classes, but it can identify most known and unknown SAR targets. Figure 5 As can be seen, unknown radar targets are well separated from known radar targets, with only a small overlap. Meanwhile, known radar targets are also effectively grouped together. Finally, the ROC curves of the proposed method and five other methods in detecting unknown radar targets are shown in the figure. Figure 6 As shown, the area under the ROC curve of the proposed method is significantly larger than that of other comparative methods. The specific values ​​of the area under the curve (AUROC) are shown in Table 2.

[0108] Table 2. Comparison of the area under the receiver operating characteristic (ROC) curve obtained by the specific implementation method of the present invention with the area under the ROC curve obtained by other methods.

[0109] method Area under the operating characteristic curve (%) OpenMax 88.80 GCPL 93.00 RPL 92.54 MLS 93.23 CBC 90.98 The proposed method 97.11

[0110] The method proposed in this invention achieves an AUROC value of 97.11%, which is 3.88% higher than the second-highest method, indicating superior performance in detecting unknown radar targets. In summary, the method proposed in this invention outperforms existing open-set recognition methods in overall performance.

[0111] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for open-set recognition of radar images based on dual-prototype regional feature enhancement, characterized in that, Includes the following steps: Step 1: Obtain training samples of radar images with known categories and test data of radar images with unknown categories. The global feature space of each sample is divided into feature space and open space. Step 2: Use a feature extractor to extract features from the input radar image; Step 3: Dual Prototype Learning; Obtain the prototype and complementary points of the radar target, with the complementary points serving as the non-first prototype. The centers of known and unknown radar image samples; The distance from the feature extractor output of the radar image sample to its corresponding complementary point and prototype is calculated. The dual prototype loss function is used to make the radar image sample move closer to its corresponding prototype while moving away from its corresponding complementary point. Furthermore, a loss function is established to constrain the range of the open space; at the same time, the loss function is constrained by complementary points to avoid the uneven distribution of complementary points causing unknown radar targets to be misclassified as known targets; Step 4: Use the trained model to identify unknown radar targets; and set a threshold for the difference between the probability of the largest and second-largest class of radar image samples to distinguish between unknown and known classes; Step 3 is described in detail below: The prototype of the k-th type of radar target is represented as: Complementary points are represented as ,in and These represent the number of prototypes and complementary points, respectively. Type k radar image samples The distance to its corresponding complementary point and prototype is expressed as: ； in, This represents the distance between two variables. This represents the output of the feature extractor, which has learnable parameters. , Represents the square of the Euclidean distance; Therefore, the biprototype loss function The mathematical expression is: ； in, K is the number of known categories. It has a positive temperature coefficient. Indicates the sample index. It is the first Complementary points It is the first Class prototype; Loss function of constrained open space Represented as: ； in It is a learnable parameter; Complementary point constraint loss function The expression is as follows: ； in ; Therefore, the overall loss function is expressed as: ； in and It's a hyperparameter.

2. The method for open set recognition of radar images based on dual-prototype regional feature enhancement according to claim 1, characterized in that, The feature extractor includes a target region feature enhancement module and an attention-aware channel discarding module, as detailed below: Target region feature enhancement module: A set of filters is applied to the input radar image to generate a binary space mask that highlights the target region; the input radar image is passed through a convolutional layer to obtain a feature map, and the binary space mask is adjusted to match the spatial resolution of the feature map to obtain an excitation map; the excitation map and the feature map are added using residual addition to obtain the enhanced feature map; The enhanced feature map output by the target region feature enhancement module is passed through a convolutional layer and then used as the input to the attention-aware channel discarding module. Attention-aware channel drop-off module: First, the average response of the channel dimension is calculated by global average pooling. Then, the channel response vector in each sample is standardized to obtain the attention score of each channel. Then, a binary drop-off mask is generated based on the attention score, and the maximum drop-off ratio is set. The binary drop-off mask is adjusted to be consistent with the shape of the input of this module, and the output of the attention-aware channel drop-off module is obtained by element-wise multiplication. Finally, after average pooling and random dropout functions, the output of the feature extractor is obtained.

3. The method for open set recognition of radar images based on dual-prototype regional feature enhancement according to claim 1, characterized in that, The criteria for distinguishing between unknown and known classes are as follows: ； in It is a test radar image sample Classified as the first The probability of a class and These represent the probability of the largest class and the probability of the second largest class, respectively. It is a preset threshold. It is a predicted label.

4. The method for open set recognition of radar images based on dual-prototype regional feature enhancement according to claim 2, characterized in that, The target region feature enhancement module is as follows: The binary space mask is obtained through the following operation: ； in express Gabor filter response in the direction, Indicates the direction angle. The set representing the direction angles, It is an indicator function. It is a threshold; The binary space mask The feature map is adjusted to match the spatial resolution of the feature map obtained by passing the input radar image through the first convolutional layer, resulting in the excitation map E: ； in, B represents the feature map, C represents the batch size, H represents the number of channels, and W represents the feature map height. This is element-wise multiplication; Introduce a coefficient that varies with time. Its definition is as follows: ； in This is the total number of training rounds. This is the current training epoch, and the enhanced feature map is ultimately obtained through residual addition. : 。 5. The method for open set recognition of radar images based on dual-prototype regional feature enhancement according to claim 4, characterized in that, The enhanced feature map, after passing through a convolutional layer, is represented by the following formula: ； in, This is the input feature map for the attention-aware channel drop-out module. Indicates the first The output of each convolutional layer, Indicates the first SiLu activation functions Indicates the first Individual batch normalization, Indicates the first Each convolution.

6. The method for open set recognition of radar images based on dual-prototype regional feature enhancement according to claim 5, characterized in that, The attention perception channel discard module is as follows: First, the average response across the channel dimension is calculated using global average pooling: ； Where H represents the feature map height and W represents the feature map width. The channel number is used as the index; then, the channel response vector within each sample is standardized to generate an attention score. ; Then, a binary dropout mask is generated based on the attention score. : ； in It is a global drop intensity coefficient. It represents a uniform distribution of 0 to 1; Finally, the binary dropout mask is adjusted to match the module's input feature map. The shapes are consistent, and the module output is obtained through element-wise multiplication. : 。

Images