Radar HRRP target open set identification method based on hierarchical prototype loss function constraint

Abstract

The invention discloses a radar HRRP target open set identification method based on hierarchical prototype loss function constraint, and aims at improving the generalization performance of a model for known categories when a test set comprises targets which do not appear in a training stage in an open environment in RATR practical application, and improving the recognition efficiency of the model. And information about the relationship between the unknown category and the known category is provided while the unknown category is identified. Specifically, firstly, a category hierarchical structure is constructed based on prior knowledge to guide training and testing stages. Then, a hierarchical prototype loss (HPL) is proposed to constrain the consistency of the feature space distribution and the category hierarchical structure. And finally, performing prediction according to a hierarchical structure by using the trained prototype network so as to provide information of a relationship between an unknown category and a known category. Wide experiments on actually measured HRRP data verify the effectiveness of the provided method on open set identification.

Description

Technical Field

[0001] The invention discloses a radar HRRP target open set recognition method involving a hierarchical prototype loss function constraint, belonging to the technical field of radar target recognition. Background Art

[0002] High Resolution Range Profile (HRRP) represents the one-dimensional distribution of the target scattering center along the radar line of sight, and contains the geometric structure information of the target. Compared with two-dimensional Synthetic Aperture Radar (SAR) images and micro-Doppler data, HRRP has a shorter acquisition time, is easier to acquire and process, and is increasingly used in Radar Automatic Target Recognition (RATR).

[0003] In practical applications, the targets to be identified by the RATR system are often non-cooperative targets or even hostile targets. It is difficult to establish a complete recognition template library of target classes during the training phase. The test set contains targets that have never appeared in the training phase, that is, targets outside the library. However, traditional methods identify all test samples as targets in the library, resulting in classification errors. Therefore, it is urgent to study open set recognition (OSR) methods that can not only identify targets in the library, but also reject targets outside the library.

[0004] The OSR research based on SAR and HRRP initially adopted traditional machine learning classifiers such as support vector machines and random forests. By setting a threshold and comparing it with the maximum confidence output of the classifier, it was determined whether the test sample was an out-of-library target. Subsequently, deep learning was increasingly widely used in open set recognition. Deep learning methods extract information directly related to the label from the input data and compress or ignore other information, so they will also give a higher confidence to out-of-library samples. Therefore, the researchers designed a prototype loss function to replace the original cross entropy loss function, compressing the feature space where the in-library categories are located, thereby leaving more open space to distinguish out-of-library samples. Then, in the test phase, the output category is determined based on the distance between the sample and the known class prototype in the feature space.

[0005] The distance between categories in the feature space learned by the prototype loss function is an important basis for determining its recognition performance. However, the semantic relationship between the feature space and the categories learned by the existing prototype learning methods is inconsistent; and there is also the influence of the posture sensitivity of HRRP. Even if the feature space distance of samples of the same category is small, their actual waveforms may be significantly different. Therefore, the feature space learned by the prototype learning method is not only difficult to describe the semantic relationship between categories, but also unable to reveal the similarity between the original waveforms. As a result, not only does this method have a decreased generalization performance when identifying samples in the library, but it is also difficult to divide the appropriate decision boundary to distinguish between objects in and outside the library.

[0006] In view of the above situation, the present invention proposes a radar HRRP target open set recognition method based on hierarchical classification based on prototype learning. This method uses hierarchical multi-granularity labels to describe objects with prior knowledge, and organizes the categories to be inferred into a category hierarchy to guide the training and testing stages. In the training stage, a hierarchical prototype loss (HPL) is proposed to constrain the feature space extracted by the prototype network so that the feature distribution of the sample is consistent with the hierarchical category structure; in the testing stage, the trained prototype network is used to make predictions according to the hierarchical structure, so as to obtain the relationship information between unknown categories and known categories. Summary of the invention

[0007] The purpose of the present invention is to provide a radar HRRP target open set recognition method constrained by a hierarchical prototype loss function. Through the present invention, HRRP can be used to enhance the open set recognition capability, improve the generalization performance of the model for known categories, and provide information on the relationship between unknown categories and known categories while identifying unknown categories, providing a more effective solution for the management of unknown categories in RATR.

[0008] The present invention introduces the idea of ​​hierarchical classification, defines the semantic categories of targets at four granularities: attribute, type, model, and posture based on prior knowledge, and organizes targets with consistent semantic categories at each granularity into a tree node, which is sequentially connected to each node containing a target, thereby organizing the category tree hierarchy.

[0009] Furthermore, a hierarchical prototype loss function is designed based on the category hierarchy to constrain the distribution of the feature space to be consistent with the semantic relationship between categories. That is, categories with similar category semantic relationships are also distributed in feature spaces with similar distances, thereby improving the generalization performance of the model for targets in the library.

[0010] Furthermore, a hierarchical classifier is designed to classify each non-leaf node in the category tree hierarchy in turn. Each classification node needs to distinguish whether the target is the category of its connected child node or the category outside the library. A complex classification task is broken down into each classification node. This approach not only simplifies the decision boundary, but also reduces the uncertainty space between samples in the library and samples outside the library. In addition, the semantic relationship between unknown categories and known categories can be obtained based on the results of each classification node.

[0011] The technical solution of the present invention is as follows:

[0012] A radar HRRP target open set recognition method constrained by a hierarchical prototype loss function comprises the following steps:

[0013] Step 1: Obtain HRRP data of radar targets: Use high-resolution radar to obtain HRRP echo data of the target, perform preprocessing such as data slicing and segmentation, amplitude normalization, and form a data set by arranging and adding labels. The data set is divided into training set and test set to facilitate subsequent model training and radar target recognition performance testing.

[0014] Step 2: Construct a tree hierarchy: Using the prior knowledge in RATR, all input HRRPs are labeled at multiple levels to guide the construction of the hierarchical tree T.

[0015] Step 3: Construct a deep learning network based on hierarchical prototype learning: the deep learning network mainly includes two modules: feature extraction module and hierarchical prototype learning module; extract the deep features of the original HRRP through the backbone network, and use the hierarchical prototype learning method to learn the class prototype for each category node, and optimize the feature space distribution in combination with the hierarchical prototype loss;

[0016] Step 4: Network training and parameter updating: The feature extractor and prototype are jointly trained based on the hierarchical prototype learning strategy and data. For each class node in the hierarchy, the corresponding prototype is calculated and updated according to the feature extraction results.

[0017] Step 5, network testing: The test sample is classified through the trained hierarchical classifier model, step by step, by the local classifiers associated with each non-leaf node in the hierarchical structure. Each local classifier determines whether the sample belongs to its sub-node category or an unknown category according to the set rejection threshold and the distance to the sub-class prototype, until it reaches the type layer or is identified as an unknown category, and finally obtains the prediction result.

[0018] Furthermore, in step 2, the prior knowledge of the recognition process includes three parts: target discrimination, i.e., identifying the target group; target classification, i.e., overall classification of the targets; and target recognition, i.e., multi-level labeling of all input HRRPs to construct a tree hierarchy including attributes, types, and models.

[0019] Furthermore, in step 2, a gesture layer is introduced below the recognition layer to capture the gesture information in the learned feature space and enhance the consistency between the distribution of sample points in the feature space and the original waveform.

[0020] Furthermore, in step 3, the deep learning network mainly includes two modules: a feature extraction module and a hierarchical prototype learning module; the deep features of the original HRRP are extracted through the backbone network, and the hierarchical prototype learning method is used to learn the class prototype for each category node, and the feature space distribution is optimized in combination with the hierarchical prototype loss.

[0021] Furthermore, in step 3, the hierarchical prototype loss is used to ensure that the feature distribution of the target in the feature space is consistent with the hierarchical category structure, so as to learn a better feature space and enhance the generalization ability of the model.

[0022] Furthermore, in step 3, the category of the leaf node is represented by c i Indicates that the nodes along the real path are represented by denotes, j denotes the layer from 0 (root) to L (leaf), and the original HRRP input u is denoted as c i , and its true value path within the classification standard T is recorded as P i , based on the prototype loss, the hierarchical prototype loss L defined HPL as follows:

[0023]

[0024] Among them, ω PL (j) and ω CE (j) is the weight factor used to adjust the importance of each level of loss, l CE and l PL They correspond to cross entropy loss and prototype loss, respectively, as shown below:

[0025]

[0026] if There are N child nodes, then the child nodes are recorded as They share a common parent node; is the feature representation f(u;θ) and the prototype The Euclidean distance between them; λ is a hyperparameter, X k Representative Category deep features, i.e., features of sibling nodes that share the same parent node in the hierarchy.

[0027] Furthermore, in step 4, the node The prototype of each class represented is represented as:

[0028]

[0029] Among them, f(u;θ) is the feature extractor.

[0030] Furthermore, in step 5, the local classifier F i j can be expressed as:

[0031]

[0032] Among them, η(j,τ) represents the rejection threshold of the jth layer, g k (u) indicates category The discriminant function of is:

[0033]

[0034] Furthermore, in step 5, when a test sample passes through node When , the associated local classifier Determine the next node for the test sample to pass; this process continues until one level above the leaf nodes - the type level.

[0035] Beneficial effects:

[0036] (1) The present invention proposes a radar HRRP target open set recognition method constrained by a hierarchical prototype loss function. The method introduces the idea of ​​hierarchical classification and defines the semantic categories of targets at four granularities: attribute, type, model, and posture based on prior knowledge. Targets with consistent semantic categories at each granularity are organized in a tree node, which is sequentially connected to each node containing a target, thereby organizing the category tree hierarchy and guiding subsequent deep learning training to capture posture-related features.

[0037] (2) The present invention introduces hierarchical prototype learning and designs a hierarchical prototype loss function based on the category hierarchy, which constrains the distribution of the feature space to be consistent with the semantic relationship between categories. That is, categories with similar category semantic relationships are also similarly distributed in the distance space, thereby improving the generalization performance of the model for objects in the library.

[0038] (3) The present invention uses a hierarchical open set recognition classifier to classify targets in sequence at each non-leaf node in the category tree hierarchy. Each classification node can not only distinguish whether the target is a category of its connected child node or an out-of-library category, but also detect unknown observations and provide contextual information about their relationship with known categories through hierarchical connections. This can provide information about the relationship between unknown categories and known categories while identifying unknown categories, which can greatly improve the performance of radar target open set recognition in actual applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The flowchart of the radar HRRP target open set recognition method constrained by the hierarchical prototype loss function;

[0040] Figure 2 Build a diagram for a tree hierarchy;

[0041] Figure 3 It is a tree-like category hierarchy diagram;

[0042] Figure 4 Schematic diagram of the loss function;

[0043] Figure 5 Schematic diagram of network training based on hierarchical prototype learning;

[0044] Figure 6 Flowchart for label prediction;

[0045] Figure 7 HRRP diagram of the truck in eight different postures;

[0046] Figure 8 It is a schematic diagram of the results of the comparison method;

[0047] Fig. 9 This is a tsne visualization of the feature space learned based on different loss functions;

[0048] Fig.10 Recognition graph of two unknown classes. DETAILED DESCRIPTION

[0049] The present invention is described in detail below with reference to the accompanying drawings and based on examples.

[0050] The present invention provides a radar HRRP target open set recognition method constrained by a hierarchical prototype loss function. The flowchart of the method is as follows: Figure 1 shown.

[0051] Step 1: Obtain HRRP data of radar targets: Use high-resolution radar to obtain HRRP echo data of the target, perform preprocessing such as data slicing and segmentation and amplitude normalization, and form a data set by sorting and adding labels. The data set is divided into training and test sets to facilitate subsequent model training and radar target recognition performance testing.

[0052] Step 2: Construct a tree hierarchy: Using the prior knowledge in RATR, all input HRRPs are labeled at multiple levels to guide the construction of the hierarchical tree T, where the categories of leaf nodes are represented by c i Indicates that the nodes along the real path are represented by c i j Indicates that j represents the level from 0 (root) to L (leaf).

[0053] 2.1. Based on the experts’ understanding of the RATR task, the prior knowledge of the recognition process includes three parts: target discrimination, i.e. identifying the target group; target classification, i.e. overall classification of the targets; and target recognition, i.e. multi-level labeling of all input HRRPs to construct a tree hierarchy including attributes, types, and models.

[0054] 2.2. Considering the posture sensitivity of HRRP, different collection postures have significant differences in HRRP. A posture layer is introduced below the recognition layer to capture posture information in the learned feature space and enhance the consistency between the distribution of sample points in the feature space and the original waveform. This ensures that the subtle differences in the collected postures can be systematically integrated, thereby providing a more detailed and accurate target representation.

[0055] Based on the above prior knowledge, all input HRRPs are labeled in multiple levels. Figure 2 As shown, the final tree hierarchy is as follows Figure 3 shown.

[0056] Step 3. Construct a deep learning network based on hierarchical prototype learning: The deep learning network mainly includes two modules: feature extraction module and hierarchical prototype learning module. The deep features of the original HRRP are extracted through the backbone network, and the hierarchical prototype learning method is used to learn the class prototype for each category node, and the hierarchical prototype loss is combined to optimize the feature space distribution.

[0057] 3.1. During the training phase, the original HRRPs are processed using the backbone network ResNet-50 for feature extraction. The process is represented as f(u,θ), where u is the original HRRP input and θ represents the parameters of the network, resulting in a set of deep features represented as X.

[0058] 3.2. Using the prototype learning method, for each category node Learn a class prototype, denoted as

[0059] Hierarchical prototype loss is used to ensure that the feature distribution of the target in the feature space is consistent with the hierarchical category structure, so as to learn a better feature space and enhance the generalization ability of the model. The original HRRP input u is denoted as c i , and its true value path within the classification standard T is recorded as P i , based on the prototype loss, the hierarchical prototype loss L defined HPL as follows:

[0060]

[0061] Among them, l CE and l PLThey correspond to cross entropy loss and prototype loss, respectively, as shown below:

[0062]

[0063] Here, if There are N child nodes, then the child nodes are recorded as They share a common parent node. is the feature representation f(u;θ) and the prototype In addition, X k Representative Category deep features, i.e., features of sibling nodes that share the same parent node in the hierarchy.

[0064] like Figure 4 As shown, CE The goal is to expand The distance between its siblings in the feature space, and l PL Try to compress The feature distribution of the class is reduced, thereby reducing the feature distance between connected child nodes. Weight factor ω PL (j) and ω CE (j) is used to adjust the importance of each level of loss. Specifically, ω PL (j) increases with the depth of the layer, aiming to consolidate the features of more fine-grained categories; while ω CE (j) increases with decreasing levels, to ensure a wider distribution gap between features between more distantly related categories.

[0065] Step 4: Network training and parameter updating: The feature extractor and prototype are jointly trained with the data based on the hierarchical prototype learning strategy. For each class node in the hierarchy, the corresponding prototype is calculated and updated according to the feature extraction results.

[0066] 4.1. Based on the hierarchical prototype learning strategy in 202, the feature extractor f(u; θ) and the prototype Joint training with data, such as Figure 5 shown.

[0067] 4.2. For each node represented by a class in the hierarchy, find its prototype based on the feature extraction results. The prototype of each class represented is represented as:

[0068]

[0069] Step 5, network testing: The test sample is classified through the trained hierarchical classifier model, step by step, by the local classifiers associated with each non-leaf node in the hierarchical structure. Each local classifier determines whether the sample belongs to its sub-node category or an unknown category according to the set rejection threshold and the distance to the sub-class prototype, until it reaches the type layer or is identified as an unknown category, and finally obtains the prediction result.

[0070] 5.1. In the testing phase, each local classifier F i j Associated with a non-leaf node in the hierarchy T to distinguish the N child nodes connected to it, F i j can be expressed as:

[0071]

[0072] Among them, η(j,τ) represents the rejection threshold of the jth layer, g k (u) indicates category The discriminant function of is:

[0073]

[0074] In addition, the threshold η(j,τ) varies with the level index j. The lower the level, the higher the rejection threshold. This is because the categories closer to the root of the word are more consistent with the basic semantic distinction and more likely to be classified as known categories. τ is defined as the percentage of training samples identified as known categories to all samples, which determines the value of η(j,τ) at different levels.

[0075] Among them, F i j The role of is to distinguish nodes By measuring the test sample and The sample is assigned to the class of the closest prototype based on the proximity of the subclass prototype. If the distance to the nearest prototype exceeds the set threshold, the sample is classified as unknown.

[0076] 5.2. In the testing phase, when a test sample passes through the node When , the associated local classifier Determine the next node that the test sample passes through. This process continues until one level above the leaf node - the type layer. In practical applications, it is often sufficient to identify the type of the target. If the sample has not yet reached a type node, take over the classifier F connected to the current node i j Find the subsequent nodes until you reach this type of node. The operation process is as follows Figure 6 shown.

[0077] Example

[0078] This section describes the effect of the present invention in combination with the measured data experiment. In order to evaluate the performance of the proposed radar HRRP target open set recognition method constrained by the hierarchical prototype loss function, two types of target data are collected: vehicles and clutter for experiments.

[0079] Dataset and parameter settings:

[0080] The dataset was collected by a wideband radar operating in the W-band frequency range. The radar system uses a frequency-stepped, linear frequency modulated signal with a synthesis bandwidth of 1250MHz to acquire high-resolution HRRP of the target.

[0081] The dataset includes HRRP data reflecting different target poses. For each vehicle, HRRP samples are taken from 8 different azimuths, each azimuth is 45° apart, and these angles cover a comprehensive range: left front oblique (LFO), left side (LS), left rear oblique (LRO), tailgating (TC), right rear oblique (RRO), right side (RS), right front oblique (RFO) and head-on (HO). Figure 7 The HRRP of the truck at different poses is shown.

[0082] The known categories of the experiment consist of 9 types of targets, including 6 types of vehicle targets and 3 types of clutter. Each category has 8000 training samples and 4000 test samples. The specific settings are shown in Table 1.

[0083] Table 1 Description of experimental dataset types

[0084]

[0085] In addition, the test set is expanded to 7 unknown categories, including 5 vehicle categories and 2 clutter categories.

[0086] The parameter settings of the proposed model are as follows: ResNet-50 network is used as the classifier, and the Stochastic Gradient Descent-Momentum (SGD-M) optimizer is used for optimization. The network is initialized with a learning rate of 0.1, and the learning rate is reduced to one-tenth of its initial value every 30 rounds. The training process has a total of 100 rounds. In addition, the hyperparameter λ is set to 0.1. For comparative analysis, the parameters selected by the discussed methods adhere to their recommended values.

[0087] Experimental results:

[0088] Based on the network parameters, a radar HRRP target open set recognition method constrained by a hierarchical prototype loss function is constructed, and classification experiments are carried out on several vehicles and clutter targets using the measured data set to evaluate the effectiveness of the method proposed in the present invention.

[0089] The proposed method is compared with SVDD+SVM, One-Class SVM (OCSVM), Forest+SVM, SoftMax, OpenMax, OSRCI, complementary point learning (RPL), generalized convolutional prototype learning (GCPL), prototype learning (PL) and adversarial complementary point learning (ARPL) based on the same dataset. PL and GCPL are used as ablation studies. PL omits the l existing in the model. PL The components of GCPL ignore the hierarchical prototype loss structure. Three indicators, ACC (precision), AUROC (area under the receiver operating characteristic curve) and OSCR (open set classification rate), are used to evaluate the effectiveness of the open set recognition (OSR) model.

[0090] The experimental results are as follows Figure 8 As shown, it can be seen that the classification accuracy of the present invention is the highest, and compared with the closed set distribution method, it has a better ability to reject unknown classes, can effectively distinguish known classes from unknown classes, and maintain optimal performance under different degrees of openness.

[0091] Further, to evaluate the performance of the method of the present invention, Fig. 9 As shown in Figure 2, visualizations of feature spaces learned using different loss functions visually compare the present invention with the GCPL approach. It can be seen that in the feature space generated by GCPL, the types of each class are tightly arranged, which leads to suboptimal generalization performance. In contrast, in HPL, the features of the types are neatly aligned according to the class they belong to. In addition, HPL creates a more compact feature space with clearer boundaries, which greatly enhances the generalization ability of the model.

[0092] The recognition results of the unknown class are shown as follows: Fig.10 As shown. The red circle represents the recognition made by the method of the present invention, while the actual category of the unknown class is described in the gray rectangle. Compared with simply rejecting the unknown class, the present invention is able to predict more information about the unknown class and provide a more detailed understanding of the unknown class.

[0093] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A radar HRRP target open set recognition method constrained by a hierarchical prototype loss function, characterized in that , including the following steps: Step 1: Obtain HRRP data of radar targets: Use high-resolution radar to obtain HRRP echo data of the target, perform preprocessing such as data slicing and segmentation, amplitude normalization, and form a data set by arranging and adding labels. The data set is divided into training set and test set to facilitate subsequent model training and radar target recognition performance testing. Step 2: Construct a tree hierarchy: Using the prior knowledge in RATR, all input HRRPs are labeled at multiple levels to guide the construction of the hierarchical tree T. Step 3: Construct a deep learning network based on hierarchical prototype learning: the deep learning network mainly includes two modules: feature extraction module and hierarchical prototype learning module; extract the deep features of the original HRRP through the backbone network, and use the hierarchical prototype learning method to learn the class prototype for each category node, and optimize the feature space distribution in combination with the hierarchical prototype loss; Step 4: Network training and parameter updating: The feature extractor and prototype are jointly trained based on the hierarchical prototype learning strategy and data. For each class node in the hierarchy, the corresponding prototype is calculated and updated according to the feature extraction results. Step 5, network testing: The test sample is classified through the trained hierarchical classifier model, step by step, by the local classifiers associated with each non-leaf node in the hierarchical structure. Each local classifier determines whether the sample belongs to its sub-node category or an unknown category according to the set rejection threshold and the distance to the sub-class prototype, until it reaches the type layer or is identified as an unknown category, and finally obtains the prediction result.

2. The method according to claim 1, characterized in that In step 2, the prior knowledge of the recognition process includes three parts: target discrimination, i.e., identifying the target group; target classification, i.e., classifying the targets in general; Target recognition is to perform multi-level labeling on all input HRRPs and build a tree hierarchy including attributes, types, and models.

3. The method according to claim 1, characterized in that In step 2, a gesture layer is introduced below the recognition layer to capture the gesture information in the learned feature space and enhance the consistency between the distribution of sample points in the feature space and the original waveform.

4. The method according to claim 1, characterized in that In step 3, the deep learning network mainly includes two modules: feature extraction module and hierarchical prototype learning module. The deep features of the original HRRP are extracted through the backbone network, and the hierarchical prototype learning method is used to learn the class prototype for each category node, and the feature space distribution is optimized in combination with the hierarchical prototype loss.

5. The method according to claim 1, characterized in that In step 3, the hierarchical prototype loss is used to ensure that the feature distribution of the target in the feature space is consistent with the hierarchical category structure, so as to learn a better feature space and enhance the generalization ability of the model.

6. The method according to claim 1, characterized in that In step 3, the category of the leaf node is c i Indicates that the nodes along the real path are represented by denotes, j denotes the layer from 0 (root) to L (leaf), and the original HRRP input u is denoted as c i , and its true value path within the classification standard T is recorded as P i , based on the prototype loss, the hierarchical prototype loss L defined HPL as follows: Among them, ω PL (j) and ω CE (j) is the weight factor used to adjust the importance of each level of loss, l CE and l PL They correspond to cross entropy loss and prototype loss, respectively, as shown below: if There are N child nodes, then the child nodes are recorded as p∈1,...N, they share a common parent node; is the feature representation f(u;θ) and the prototype The Euclidean distance between them; λ is a hyperparameter, X k Representative Category deep features, i.e., features of sibling nodes that share the same parent node in the hierarchy.

7. The method according to claim 1, characterized in that In step 4, the node The prototype of each class represented is: Among them, f(u;θ) is the feature extractor.

8. The method according to claim 1, characterized in that In step 5, the local classifier F i j can be expressed as: Among them, η(j,τ) represents the rejection threshold of the jth layer, g k (u) indicates category The discriminant function of is:

9. The method according to claim 1, characterized in that In step 5, when a test sample passes through node When , the associated local classifier Determine the next node for the test sample to pass; this process continues until one level above the leaf nodes - the type level.

Images