A high-resolution range profile increment identification method based on domain condition feature calibration

Abstract

The application relates to a high-resolution range profile (HRRP) increment recognition method based on domain condition feature calibration, which comprises the following steps: collecting an HRRP signal to be recognized; inputting the HRRP signal to be recognized into a trained HRRP increment recognition network to obtain a recognition result of a target category; wherein the HRRP increment recognition network comprises a shared feature extractor, a domain condition feature calibrator, a cross-domain feature aligner and a classifier which are connected in sequence, wherein: the shared feature extractor is used for performing feature extraction on the HRRP signal to be recognized to obtain an HRRP feature vector; the domain condition feature calibrator is used for outputting a domain feature calibration vector; and the cross-domain feature aligner is used for outputting a domain-invariant feature vector which is robust to posture changes, wherein the domain-invariant feature vector is used for recognition by the classifier, so that the recognition result is obtained. The method can significantly improve the generalization ability and memory efficiency in a serialized posture domain.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a high-resolution range image incremental recognition method based on domain condition feature calibration. Background Technology

[0002] High-resolution range profiles (HRRPs), as an important feature in radar target recognition, can reflect the fine distribution of scattering points of a target in the radial range dimension. With the development of deep learning technology, HRRP target recognition methods based on deep neural networks have achieved high recognition accuracy under ideal conditions of complete training data and complete attitude angle coverage. However, in practical engineering scenarios such as non-cooperative target recognition, there is often a continuous and dynamic data acquisition environment: the target attitude angle range cannot be fully covered at once, and data is often collected and supplemented in batches and stages according to different attitude angle ranges. This leads to a continuous learning problem in a "domain expansion" scenario, where the target category set remains fixed, but the attitude domain (data distribution) of the data gradually expands over time.

[0003] For incremental learning in this domain, existing technologies mainly fall into two categories. One category is based on regularization, such as elastic weight consolidation, which mitigates the overwriting of knowledge from old tasks by imposing importance constraints on network parameters. However, as the number of tasks increases, the constraints of different tasks are prone to conflict, leading to a significant performance degradation. The other category is based on contextual memory, such as gradient contextual memory, which avoids the increase in loss from old tasks by storing a small number of historical samples and constraining the gradient direction during training for new tasks. Although this method provides certain theoretical guarantees, it has obvious limitations in the incremental pose domain scenario of HRRP: First, its computational cost increases linearly with the number of tasks, requiring the solution of a quadratic programming problem for each old task, resulting in high engineering deployment costs; second, it lacks an explicit alignment mechanism for feature distribution shifts between different pose domains, making it difficult to effectively learn domain-invariant features; third, its memory samples usually adopt a random selection strategy, failing to fully utilize limited storage space to retain the most representative boundary samples, resulting in low anti-forgetting efficiency.

[0004] However, existing technologies, when applied to incremental learning in the HRRP pose domain, fail to address the feature distribution shift caused by continuous changes in pose angles, thus limiting their generalization ability and memory efficiency in the serialized pose domain. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a high-resolution range image increment recognition method based on domain conditional feature calibration. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a high-resolution range image increment recognition method based on domain condition feature calibration, comprising: Acquire the HRRP signal to be identified; The HRRP signal to be identified is input into the trained HRRP incremental recognition network to obtain the recognition result of the target category; The HRRP incremental recognition network comprises a shared feature extractor, a domain conditional feature calibrator, a cross-domain feature aligner, and a classifier connected in sequence, wherein: The shared feature extractor is used to extract features from the HRRP signal to be identified, and obtain the HRRP feature vector. The domain condition feature calibrator is used to estimate the domain probability distribution of the HRRP feature vector through a built-in domain detector, and adaptively fuse the calibration parameters specific to each attitude domain according to the domain probability distribution estimation result, so as to perform an adaptive affine transformation on the initial feature vector, alleviate the feature distribution offset caused by attitude differences, and output the domain feature calibration vector. The cross-domain feature aligner is used to suppress pose-domain related information in the domain feature calibration vector and enhance class discrimination information through its trained parameter mapping, and output a domain-invariant feature vector that is robust to pose changes. The domain-invariant feature vector is used by the classifier for recognition, thereby obtaining the recognition result.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the limitation of existing technologies in incremental learning of the HRRP pose domain, which fail to specifically handle feature distribution shifts caused by continuous changes in pose angles, thus restricting their generalization ability and memory efficiency in the sequential pose domain, this invention provides a high-resolution distance image incremental recognition method based on domain-conditional feature calibration. This method constructs a recognition network comprising a shared feature extractor, a domain-conditional feature calibrator, a cross-domain feature aligner, and a classifier. During the network's sequential learning of new pose domains, the domain-conditional feature calibrator first performs an affine transformation on the input features to adapt to the pose domain, explicitly mitigating the distribution shift. Then, the cross-domain feature aligner, combined with adversarial training and distribution alignment loss, learns domain-invariant discriminative features. Simultaneously, a domain boundary-aware strategy is used to maintain the memory bank, and a feature distillation loss weighted by pose angle differences is employed to efficiently preserve historical knowledge. Furthermore, the model strictly controls model size growth when introducing new domains by incrementally expanding the domain calibration parameters and freezing old parameters. This method unifies domain adaptation and anti-forgetting under the same optimization objective, thereby maintaining stable recognition performance of learned old domains while continuously introducing new pose domains. It significantly improves generalization ability and memory efficiency in serialized pose domains and has good engineering deployability. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the high-resolution distance image incremental recognition method based on domain condition feature calibration provided in this embodiment of the invention. Figure 2 This is a flowchart of the training process of the HRRP incremental recognition network provided in an embodiment of the present invention; Figure 3 This is a comparison chart of the recognition accuracy of the proposed method and the gradient context memory method in a 10-domain incremental learning scenario, provided by an embodiment of the present invention. Detailed Implementation

[0008] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0009] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0010] The high-resolution distance image increment recognition method based on domain condition feature calibration proposed in this invention will now be described in detail with reference to the accompanying drawings.

[0011] Figure 1 This is a flowchart illustrating the high-resolution range image increment recognition method based on domain conditional feature calibration provided in this embodiment of the invention. Figure 1 As shown, it includes: S1: Acquire the HRRP signal to be identified; S2: Input the HRRP signal to be identified into the trained HRRP incremental recognition network to obtain the recognition result of the target category; The HRRP incremental recognition network comprises a shared feature extractor, a domain-conditional feature calibrator, a cross-domain feature aligner, and a classifier connected in sequence, wherein: A shared feature extractor is used to extract features from the HRRP signal to be identified, and obtain the HRRP feature vector. The domain condition feature calibrator is used to estimate the domain probability distribution of HRRP feature vectors through a built-in domain detector, and adaptively fuses the fixed calibration parameters corresponding to each attitude domain according to the domain probability distribution estimation results, so as to perform an adaptive affine transformation on the initial feature vector, cancel the feature distribution shift caused by attitude differences, and output the domain feature calibration vector. A cross-domain feature aligner is used to suppress pose-domain-related information in the domain feature calibration vector and enhance class discrimination information through its trained parameter mapping, outputting a domain-invariant feature vector that is robust to pose changes. The domain-invariant feature vector is used by the classifier for recognition, thereby obtaining the recognition result.

[0012] Here, the shared feature extractor is implemented using a one-dimensional residual network. The domain-conditional feature calibrator includes a domain detector and a feature calibrator. The domain detector consists of a three-layer fully connected network used to estimate the domain probability distribution of the input features. The feature calibrator maintains a set of learnable scaling and offset parameters for each pose domain, used to perform affine transformations on the features based on the domain identifier. The cross-domain feature aligner includes a feature projector, a gradient inversion layer, and a domain discriminator. The feature projector maps the input features to a low-dimensional alignment space. The gradient inversion layer and the domain discriminator work together to implement adversarial training, driving the network to learn domain-invariant features.

[0013] It should be noted that the above model structures operate differently during the training and inference phases: (1) During the training phase, all components are activated and participate in optimization. Specifically, the gradient reversal layer works in conjunction with the domain discriminator to guide feature learning through adversarial loss; at the same time, the domain detector is trained under supervision using real domain labels and is jointly optimized with the feature calibrator.

[0014] (2) During the inference phase (i.e., when the model is actually deployed and used), the network performs forward propagation computation. At this time, the gradient inversion layer is in identity mode and does not change the gradient flow direction; the domain discriminator does not participate in the computation. The model mainly relies on the pre-trained feature projector, domain detector, and feature calibrator parameters to extract features from the input signal, perform domain adaptive calibration and alignment, and finally output the recognition result by the classifier.

[0015] For example, the HRRP signal to be identified is input into a pre-trained HRRP incremental recognition network. First, the shared feature extractor extracts the initial feature vector; then, the domain detector in the domain conditional feature calibrator estimates its domain probability distribution. It should be noted that the domain conditional feature calibrator comprises two core sub-modules: the domain detector and the feature calibrator. They work together to achieve normalized calibration of features in different pose domains, but their roles and data interaction methods differ fundamentally at different stages (training / inference).

[0016] Specifically, the domain detector, as a discriminative network, has the core function of estimating the pose domain to which the input sample belongs. The feature calibrator's core function is to perform feature transformation. It maintains an independent set of learnable parameters, i.e., scaling parameters, for each learned pose domain. ) and offset parameter ( ).

[0017] During the inference phase, a "soft hybrid strategy" is employed. First, the trained domain detector estimates the domain attributes of the input samples, outputting a domain probability distribution. Then, the feature calibrator weights and fuses the calibration parameters of all domains based on this probability distribution, and performs an affine transformation using the fused parameters. For example, the parameters for weighting and fusing the calibration parameters of all domains are represented as follows: ;in, The sample estimated by the domain detector belongs to the domain. The probability (satisfying) ), , Domains The corresponding learnable scaling and offset parameters.

[0018] During the model training phase, a "hard selection strategy" is adopted; the system directly indexes and calls the corresponding data from the feature calibrator based on the true domain labels of the samples. , The parameter set performs precise transformations on the features. This process does not depend on the output probabilities of the domain detector. In parallel, the same feature vector is input into the domain detector, and its estimated probability distribution is compared with the true domain labels. By calculating the domain classification loss, the parameters of the domain detector are optimized in reverse, continuously improving its predictive ability.

[0019] Subsequently, the data is processed in the cross-domain feature aligner. Specifically, the feature projector maps the calibrated features to a low-dimensional alignment space (e.g., 256-dimensional), outputting low-dimensional features (domain-invariant feature vectors). Finally, the classifier outputs the target category recognition result based on these low-dimensional features.

[0020] The training process of the HRRP incremental recognition network is explained in detail below.

[0021] Figure 2 This is a flowchart illustrating the training process of the HRRP incremental recognition network provided in an embodiment of the present invention. Figure 2 As shown, the HRRP incremental recognition network is trained in the following way: S100: D attitude domains are pre-divided according to the attitude angle range.

[0022] In real-world radar target recognition scenarios, high-resolution range image data often cannot cover the entire attitude angle range of a target in one go. Instead, it is acquired in batches according to different time periods and observation angles. By systematically dividing the continuous attitude angle range (e.g., 10° to 90°) into multiple non-overlapping attitude domains (e.g., each domain is 8°), each attitude domain is treated as a "new task" that needs to be learned sequentially. Moreover, when learning a new domain, the complete data of the old domain is usually no longer available (due to storage limitations). This provides a clear object and boundary for the application of all subsequent mechanisms of this invention (such as anti-forgetting and cross-domain alignment).

[0023] S200: Perform sequential training on the D attitude domains in order. For the d-th attitude domain, 1≤d≤D, execute S210-S260: S210: When d is greater than 1, obtain the network trained in the (d-1)th pose domain as the current model, and expand the domain condition feature calibration module in the current model to add learnable calibration parameters corresponding to the dth pose domain, while freezing the calibration parameters that have been learned in the previous (d-1)th pose domains.

[0024] It should be noted that when d=1, the HRRP incremental recognition network obtained after training the first pose domain is obtained by training the initialized HRRP incremental recognition network using a training batch consisting only of samples from the first pose domain; wherein, the first joint loss includes at least the current domain classification loss.

[0025] It should be understood that freezing the calibration parameters of all old domains (1 to d-1) ensures that these parameters, which represent old knowledge, will not be changed by new gradient updates when training new domains, thus structurally preventing direct destruction of old knowledge.

[0026] It should be understood that adding a learnable calibration parameter corresponding to the d-th pose domain can be understood as adding a new row or group to the data structure storing calibration parameters in the feature calibrator to represent the d-th new pose domain. Furthermore, when expanding the domain conditional feature calibration module in the current model, the initial value of the learnable calibration parameter corresponding to the d-th pose domain is determined based on the statistical values ​​of the calibration parameters that have been learned for the previous d-1 pose domains. Additionally, the number of neurons in the output layer of the domain detector in the domain conditional feature calibration module is expanded from d-1 to d, enabling it to output the probability of the new domain.

[0027] S220: Input the d-th training batch into the current model, and process it sequentially through the shared feature extractor, domain conditional feature calibrator, cross-domain feature aligner and classifier in the current model to obtain various outputs.

[0028] Here, before executing S220, it also includes: building a memory bank, which is initialized to empty before training begins and updated according to the domain boundary awareness selection strategy after training in each pose domain is completed. It is used to store representative samples in the current pose domain whose pose angles are located at the boundary and close to the center of each category, as historical pose domain data.

[0029] Here, the d-th training batch is obtained by saving a copy of the parameters of the current model as a teacher model and extracting some historical pose domain data from the memory bank as replay samples to suppress catastrophic forgetting; the replay samples are mixed with new samples in the d-th pose domain to form the d-th training batch.

[0030] Based on this, S220 specifically includes: S221: Extract features from the samples in the d-th training batch using the current shared feature extractor to obtain the initial feature vector for each sample.

[0031] For example, the initial feature vector for each sample is a 512-dimensional feature vector. .

[0032] S222: In the current domain conditional feature calibrator, based on the true domain label of each sample, select the corresponding, learned and fixed calibration parameters, perform an affine transformation on the initial feature vector of the sample, and obtain the domain-calibrated feature vector.

[0033] For example, in S222, the feature calibrator directly uses the true domain label of each sample to perform an affine transformation on the initial feature vector of that sample, obtaining a domain-calibrated feature vector (still 512-dimensional). Here, the mathematical expression for the domain-calibrated feature vector is:

[0034] in, , Domains The corresponding scaling and offset parameters, This indicates element-wise multiplication.

[0035] While performing feature calibration in step S222, the model performs the following two loss calculations in parallel for training supervision: 1. Domain classification loss calculation: Input the same initial feature vector obtained in S221 into the domain detector to obtain its estimated domain probability distribution, and use the cross-entropy loss function to calculate the difference between this distribution and the true domain label of the sample, which is used as the domain classification loss to optimize the domain detector.

[0036] 2. Replay classification loss calculation: For replay samples from the memory bank in the training batch, the difference between the recognition result (output by the domain detector) and the corresponding true category label is calculated using the cross-entropy loss function, which serves as the replay classification loss to consolidate the model's mastery of historical category knowledge.

[0037] The two losses mentioned above, together with other loss terms, will constitute the joint training loss used to optimize the model in subsequent steps. The specific calculation method will be explained in a unified manner when introducing the joint training loss.

[0038] S223: Input the feature vector of each domain-calibrated feature vector into the feature projector of the current cross-domain feature aligner for mapping processing, so as to output the corresponding domain-invariant feature vector.

[0039] It should be noted that during the model training phase, the mapping process described in S223 is a core step in achieving the goal of learning "domain-invariant features," but it is not the entirety of this step. To achieve this goal, this step actually includes a multi-branch data processing and preparation process during training, designed to provide the necessary computational input for the subsequent joint optimization of the loss function.

[0040] Specifically, while and after obtaining the projected features, the system performs the following key data processing in parallel: 1. Adversarial Training Data Stream Preparation: After projection, the features are fed into the domain discriminator through the gradient inversion layer. The system must simultaneously provide corresponding domain source labels (labeling features as coming from "current new domain" or "historical old domain") to prepare aligned supervision signals for the calculation of domain adversarial loss.

[0041] 2. Explicitly Aligned Data Organization: To calculate the maximum mean difference loss, the system needs to collect and organize two types of feature sets: one is the projected feature set of the current new domain (domain d), and the other is the projected feature set of historical old domain samples sampled from the memory. These two sets will serve as direct inputs for measuring distribution difference.

[0042] 3. Feature distillation data caching (if d>1): For replay samples, the system needs to cache two sets of feature data simultaneously: one set of normalized features extracted by the current model and the other set of corresponding normalized features extracted by the teacher model, in order to calculate the feature similarity loss later.

[0043] Therefore, during the training phase, the output includes: a. Domain-invariant feature vectors (main data stream) used by the classifier.

[0044] b. Feature-domain label pairs for use by the domain discriminator (for domain adversarial loss).

[0045] c. Feature sets of the old and new domains (used for distribution alignment loss).

[0046] d. Feature pairs of the old and new models (used for feature distillation loss).

[0047] The specific calculation functions and fusion methods for each loss will be explained in detail in the subsequent section on joint training loss. This design ensures the coherence of the training process description and concentrates the technical details in a dedicated chapter.

[0048] S224: Input the domain-invariant feature vector of each domain into the current classifier to obtain the corresponding recognition result.

[0049] The calculation of losses at each of the above stages will now be described in detail.

[0050] S230: Based on various outputs, calculate the d-th joint training loss, which is a weighted sum of the current domain classification loss, replay classification loss, domain adversarial loss, distribution alignment loss, and feature distillation loss.

[0051] Specifically, the current domain classification loss, replay classification loss, domain adversarial loss, and distribution alignment loss in the d-th joint training loss are obtained through the following methods (S231-S237): S231: Input the initial feature vector of each sample in the d-th training batch into the domain detector of the current domain conditional feature calibrator to estimate the corresponding domain probability distribution.

[0052] S232: Using the cross-entropy loss function, calculate the difference between the domain probability distribution of the new sample in the d-th training batch and the corresponding true domain label, as the current domain classification loss.

[0053] Here, the mathematical expression for the current domain classification loss is: ; in, This represents the number of samples in the current domain. For category indexing, For the first The one-hot representation of the true label of the nth sample is shown in the nth... The components of the dimension, The first output of the domain detector in the current domain condition feature calibrator The predicted category of the sample belongs to the _th ... The probability of a class This represents the number of categories.

[0054] S233: The cross-entropy loss function is used to calculate the difference between the domain probability distribution corresponding to the replay sample in the d-th training batch and the corresponding true domain label, which is used as the replay classification loss.

[0055] Here, the mathematical expression for the replay classification loss is: ; in, This represents the number of replay samples in the current batch. For the first The one-hot representation of the true label of the replay sample at the _th replay sample. The components of the dimension, The first output of the domain detector in the current domain condition feature calibrator The predicted replay sample belongs to the _th The probability of a class.

[0056] S234: Input the initial feature vector of each sample in the d-th training batch into the current cross-domain feature aligner to estimate the corresponding domain discrimination result.

[0057] S235: Quantify the difference between the domain discrimination result of each sample and the corresponding real domain source label, as the domain adversarial loss.

[0058] Here, the mathematical expression for the adversarial loss in this domain is: ; in, Let be the expected function. The true domain source label for the sample. This represents the domain discrimination result predicted by the current cross-domain feature aligner.

[0059] S236: Obtain the domain-invariant feature vectors of the new samples and replay samples in the d-th training batch after they have been processed by the cross-domain feature aligner.

[0060] S237: Use the maximum mean difference loss function to calculate the distance between the two sets of domain-invariant eigenvectors on the feature distribution, and obtain the distribution alignment loss.

[0061] Here, the mathematical expression for the distribution alignment loss is: ; in, For Gaussian kernel mapping, The number of samples for replay. The number of samples for the new sample. This is the set of domain-invariant feature vectors obtained after the replay samples have been processed by the cross-domain feature aligner. This is the set of domain-invariant feature vectors obtained after processing the new sample using a cross-domain feature aligner. Let the square of the norm in the reproducing kernel Hilbert space be . For the regenerated nucleus Hilbert space.

[0062] Furthermore, as a major innovation of this invention, the feature distillation loss in the d-th joint training loss is calculated in the following manner: If d > 1, perform the following steps: Normalized feature vectors of replay samples in the normalized feature space are extracted from the current model and the teacher model corresponding to the current model, respectively. Based on the difference between the attitude angle of each replay sample and the attitude center angle of its historical attitude domain, the corresponding distillation weights are determined. Using the distillation weights, the feature similarity loss between the two sets of normalized feature vectors is calculated to obtain the feature distillation loss.

[0063] It's important to note that knowledge distillation is performed in the normalized feature space rather than the output layer or the original feature space because traditional distillation methods provide limited constraint information in this output layer or original feature space, making it difficult to capture feature representation changes caused by domain shifts within the model. Furthermore, the magnitude (norm) of unnormalized feature vectors is easily affected by factors such as input scale and the range of network layer activation values. Directly calculating the distance between two feature vectors (e.g., Euclidean distance) introduces both magnitude and direction differences. By projecting features onto a unit hypersphere using normalization, magnitude information can be removed, allowing focus on comparing feature directions. Moreover, ensuring that all sample features are at the same starting point when calculating similarity helps mitigate internal covariate shifts, making the calculation of feature similarity losses (e.g., cosine distance or mean squared error) more robust, and gradient backpropagation more stable, thereby improving the convergence and reliability of the entire distillation training process.

[0064] Here, the distillation weight is determined based on the difference between the attitude angle of each replay sample and the attitude center angle of its corresponding historical attitude domain. The attitude angle of each replay sample refers to the specific attitude angle value (e.g., 23°) corresponding to a single sample belonging to a certain historical attitude domain, retrieved from the memory. The attitude center angle of the historical attitude domain refers to the statistical center of the attitude angles of the entire historical attitude domain to which the replay sample belongs (e.g., the "20°-30°" domain), which is the mean, median, or a predefined nominal value of the attitude angles of all samples belonging to that domain. The larger the difference between the two attitude angles, the smaller the distillation weight.

[0065] Therefore, the mathematical expression for this characteristic distillation loss is: ; in, The number of replay samples that participated in the distillation. For the first Distillation weights of each replay sample To utilize the current model to extract the first A set of normalized feature vectors of each replay sample. To utilize the teacher model corresponding to the current model to extract the first A set of normalized feature vectors of each replay sample. for Norm, , .

[0066] Based on the above, the calculation expression for each joint training loss is as follows: ;in, , , , and All of these are hyperparameters.

[0067] S240: Based on the d-th joint loss, update all trainable parameters in the current network except for the frozen parameters through the backpropagation algorithm to obtain the network trained in the d-th pose domain.

[0068] Specifically, the scope of parameter updates clearly includes: 1) Share all parameters of the feature extractor; 2) All parameters of the domain detector in the domain condition feature calibrator; 3) All parameters of the feature projector and domain discriminator in the cross-domain feature aligner; 4) All parameters of the classifier; 5) In the domain condition feature calibrator, only the set of learnable calibration parameters corresponding to the d-th pose domain currently being trained are used.

[0069] The scope of parameter freezing is clearly defined as follows: 6) In the domain condition feature calibrator, all calibration parameters (scaling and offset parameters) that have been learned in the first d-1 historical pose domains are included. These parameters are marked as untrainable, their gradients are masked, and their values ​​remain unchanged in the current and subsequent training iterations.

[0070] S250: Determine if d is equal to D.

[0071] S260: If not, perform a new round of network training in the (d+1)th pose domain until sequential training of all D pose domains is completed, and then obtain the trained HRRP incremental recognition network.

[0072] To address the limitation of existing technologies in incremental learning of the HRRP pose domain, which fail to specifically handle feature distribution shifts caused by continuous changes in pose angles, thus restricting their generalization ability and memory efficiency in the sequential pose domain, this invention provides a high-resolution distance image incremental recognition method based on domain-conditional feature calibration. This method constructs a recognition network comprising a shared feature extractor, a domain-conditional feature calibrator, a cross-domain feature aligner, and a classifier. During the network's sequential learning of new pose domains, the domain-conditional feature calibrator first performs an affine transformation on the input features to adapt to the pose domain, explicitly mitigating the distribution shift. Then, the cross-domain feature aligner, combined with adversarial training and distribution alignment loss, learns domain-invariant discriminative features. Simultaneously, a domain boundary-aware strategy is used to maintain the memory bank, and a feature distillation loss weighted by pose angle differences is employed to efficiently preserve historical knowledge. Furthermore, the model strictly controls model size growth when introducing new domains by incrementally expanding the domain calibration parameters and freezing old parameters. This method unifies domain adaptation and anti-forgetting under the same optimization objective, thereby maintaining stable recognition performance of learned old domains while continuously introducing new pose domains. It significantly improves generalization ability and memory efficiency in serialized pose domains and has good engineering deployability.

[0073] The effectiveness of the proposed method is now verified using simulation software. The proposed method is compared with the gradient context memory method. Here, a high-resolution range image dataset containing 30 aircraft models, generated by electromagnetic simulation software, is used. The radar carrier frequency is 5.2 Hz, the bandwidth is 400 MHz, and the generated high-resolution range image dimension is 200. The attitude angle range is 10° to 90°, divided into 10 domains, each domain covering an 8° attitude angle. Each target class contains 1600 training samples and 1600 test samples, for a total of 48,000 training samples and 48,000 test samples.

[0074] Figure 3 This is a comparison chart of the recognition accuracy of the proposed method and the gradient episodic memory method in a 10-domain incremental learning scenario, provided by an embodiment of this invention. The gradient episodic memory method (GEM), as the closest existing method to this invention, its core idea is to store a subset of samples from old tasks as episodic memory, and constrain the gradient update direction when learning new tasks to ensure that updates do not increase the loss of old tasks. Its gradient constraint form is:

[0075] in The gradient of the current task. For the first The gradient of the old task is calculated on the memorized samples. When the constraints are not satisfied, the projected gradient that is closest to the original gradient and satisfies the constraints is solved by quadratic programming. The advantage of this method is that it is theoretically guaranteed not to increase the loss of the old task and can suppress catastrophic forgetting to some extent.

[0076] Table 1 shows a comparison of the advantages and disadvantages of the two methods.

[0077] Table 1

[0078] like Figure 3 As shown, this experiment divides the data into 10 sequential tasks (numbered 0-9) at equal intervals according to attitude angles. The experiment covers an attitude range of 10° to 90°, with each task corresponding to an 8° sampling interval (e.g., task 0 is 10°-17°, task 1 is 18°-25°, and so on). Each task contains 2400 training samples and 2400 test samples. As the domain increases, the recognition accuracy of both methods decreases, but the decrease in the method of this invention is significantly smaller. After completing the learning of all 10 domains, the overall recognition accuracy of the method of this invention remains above 90%, while the accuracy of the gradient contextual memory method decreases to about 60%. Experimental results show that the proposed method has a significantly higher recognition accuracy than the gradient contextual memory method in multi-domain incremental learning scenarios and can effectively suppress catastrophic forgetting.

[0079] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A high-resolution range image increment recognition method based on domain conditional feature calibration, characterized in that, include: Acquire the HRRP signal to be identified; The HRRP signal to be identified is input into the trained HRRP incremental recognition network to obtain the recognition result of the target category; The HRRP incremental recognition network comprises a shared feature extractor, a domain conditional feature calibrator, a cross-domain feature aligner, and a classifier connected in sequence, wherein: The shared feature extractor is used to extract features from the HRRP signal to be identified, and obtain the HRRP feature vector. The domain condition feature calibrator is used to estimate the domain probability distribution of the HRRP feature vector through a built-in domain detector, and adaptively fuse the calibration parameters specific to each attitude domain according to the domain probability distribution estimation result, so as to perform an adaptive affine transformation on the initial feature vector, alleviate the feature distribution offset caused by attitude differences, and output the domain feature calibration vector. The cross-domain feature aligner is used to map the domain feature calibration vector through a built-in feature projector, suppress the pose domain-related information in the domain feature calibration vector and enhance the class discrimination information, and output a domain-invariant feature vector that is robust to pose changes. The domain-invariant feature vector is used by the classifier for recognition, thereby obtaining the recognition result.

2. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 1, characterized in that, The HRRP incremental recognition network was trained in the following way: D attitude domains are pre-divided according to the attitude angle range; Sequential training is performed on the D pose domains in sequence. For the d-th pose domain, 1≤d≤D, the following is executed: When d is greater than 1, the network trained in the (d-1)th pose domain is obtained as the current model, and the domain condition feature calibration module in the current model is expanded to add learnable calibration parameters corresponding to the dth pose domain, while freezing the calibration parameters that have been learned in the previous (d-1)th pose domains. The d-th training batch is input into the current model and processed sequentially by the shared feature extractor, domain conditional feature calibrator, cross-domain feature aligner and classifier in the current model to obtain various outputs. Based on the various outputs, the d-th joint training loss is calculated, which is a weighted sum of the current domain classification loss, replay classification loss, domain adversarial loss, distribution alignment loss, and feature distillation loss. Based on the d-th joint loss, all trainable parameters in the current network except for the frozen parameters are updated by the backpropagation algorithm to obtain the network trained in the d-th pose domain. Determine if d is equal to D; If not, a new round of network training is performed in the (d+1)th pose domain until sequential training of all D pose domains is completed, and the trained HRRP incremental recognition network is obtained.

3. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 2, characterized in that, When d=1, the HRRP incremental recognition network obtained after training the first pose domain is obtained by training the initialized HRRP incremental recognition network using a training batch consisting only of samples from the first pose domain. The first joint loss includes at least the current domain classification loss.

4. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 2, characterized in that, Training the HRRP incremental recognition network also includes: A memory bank is constructed. The memory bank is initialized to empty before training begins and updated according to the domain boundary awareness selection strategy after training in each pose domain is completed. It is used to store representative samples in the current pose domain whose pose angles are located at the boundary and close to the center of each category, as historical pose domain data.

5. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 4, characterized in that, The d-th training batch is obtained in the following way: A copy of the parameters of the current model is saved as a teacher model, and a portion of the historical pose domain data is extracted from the memory bank as a replay sample to suppress catastrophic forgetting. The replay samples are mixed with the new samples in the d-th pose domain to form the d-th training batch.

6. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 2, characterized in that, The d-th training batch is input into the current model and processed sequentially through the shared feature extractor, domain conditional feature calibrator, cross-domain feature aligner, and classifier in the current model to obtain various outputs, including: The current shared feature extractor is used to extract features from the samples in the d-th training batch to obtain the initial feature vector of each sample; In the current domain conditional feature calibrator, based on the true domain label of each sample, the corresponding, learned and fixed calibration parameters are selected, and an affine transformation is performed on the initial feature vector of the sample to obtain the domain-calibrated feature vector. Each domain-calibrated feature vector is input into the feature projector of the current cross-domain feature aligner for mapping processing to output the corresponding domain-invariant feature vector. Each domain-invariant feature vector is input into the current classifier to obtain the corresponding recognition result.

7. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 6, characterized in that, The current domain classification loss, replay classification loss, domain adversarial loss, and distribution alignment loss in the d-th joint training loss are obtained in the following way: The initial feature vector of each sample in the d-th training batch is input into the domain detector of the current domain conditional feature calibrator to estimate the corresponding domain probability distribution; The cross-entropy loss function is used to calculate the difference between the domain probability distribution of the new sample in the d-th training batch and the corresponding true domain label, which is used as the current domain classification loss. The cross-entropy loss function is used to calculate the difference between the domain probability distribution and the corresponding real domain label of the replay sample in the d-th training batch, which is used as the replay classification loss. The initial feature vector of each sample in the d-th training batch is input into the current cross-domain feature aligner to estimate the corresponding domain discrimination result; The difference between the domain discrimination result of each sample and the corresponding real domain source label is quantified and used as the domain adversarial loss; Obtain the domain-invariant feature vectors of the new samples and replay samples in the d-th training batch after they have been processed by the cross-domain feature aligner. The maximum mean difference loss function is used to calculate the distance between the two sets of domain-invariant feature vectors on the feature distribution, thus obtaining the distribution alignment loss.

8. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 7, characterized in that, The feature distillation loss in the d-th joint training loss is calculated in the following way: If d > 1, perform the following steps: Normalized feature vectors in the normalized feature space of the replay samples in the d-th training batch are extracted using the current model and the teacher model corresponding to the current model, respectively. The corresponding distillation weight is determined based on the difference between the attitude angle of each playback sample and the attitude center angle of its historical attitude domain. Using the distillation weights, the feature similarity loss between the two sets of normalized feature vectors is calculated in a weighted manner to obtain the feature distillation loss.

9. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 8, characterized in that, The calculation expression for each joint training loss is as follows: ； ； ； ； ； ； in, , , , and All of these are hyperparameters. The classification loss for the current domain is... For the domain adversarial loss, The playback classification loss, The distribution alignment loss is... The characteristic distillation loss; This represents the number of samples in the current domain. For category indexing, For the first The one-hot representation of the true label of the nth sample is shown in the nth... The components of the dimension, The first domain detector output in the current domain condition feature calibrator The predicted category of the sample belongs to the _th ... The probability of a class For the number of categories, Let be the expected function. The true domain source label for the sample. The domain discrimination result predicted by the current cross-domain feature aligner. This represents the number of replay samples in the current batch. For the first The one-hot representation of the true label of the replay sample at the _th replay sample. The components of the dimension, The first domain detector output in the current domain condition feature calibrator The predicted replay sample belongs to the _th The probability of a class For Gaussian kernel mapping, The number of samples in the replay sample. The number of samples in the new sample. This refers to the set of domain-invariant feature vectors obtained after the replay samples have been processed by the cross-domain feature aligner. The set of domain-invariant feature vectors obtained after the new sample has been processed by the cross-domain feature aligner. Let the square of the norm in the reproducing kernel Hilbert space be . For the regenerated nucleus Hilbert space, The number of replay samples that participated in the distillation. For the first Distillation weights of each replay sample To utilize the extracted data from the current model, the first... A set of normalized feature vectors of each replay sample. To utilize the teacher model corresponding to the current model to extract the first... A set of normalized feature vectors of each replay sample. for Norm.

10. The high-resolution range image increment recognition method based on domain condition feature calibration according to claim 2, characterized in that, When expanding the domain condition feature calibration module in the current model, the initial value of the learnable calibration parameter corresponding to the d-th attitude domain is determined based on the statistical values ​​of the calibration parameters that have been learned in the previous d-1 attitude domains.

Images