An air target radar high-resolution range profile few-sample class incremental identification method

Abstract

This invention discloses a method for incremental recognition of high-resolution range profiles of airborne targets using radar with few sample classes, relating to the field of radar target recognition. Without requiring retraining with old class samples, it generates pseudo-class data by simulating changes in the physical scale of the target using temporal scaling, thus mitigating the risk of model overfitting caused by the single aircraft type of the actual cooperative target. Simultaneously, it employs a Markov probability matrix to model the HRRP amplitude sequence as a robust discrete state transition process. This invention requires no complex coordinate transformations, has low computational overhead, and involves only one-dimensional signal interpolation, pooling, and lightweight convolution operations, enabling seamless integration into existing radar target recognition systems.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition, and more particularly to an incremental recognition method for low-sample-value aerial target radar high-resolution range images. Background Technology

[0002] High-Resolution Range Profile (HRRP) is a one-dimensional projection of the radial distance of a target's scattering point along the radar's line-of-sight. It provides detailed scattering information and enables rapid imaging, and has been widely used in practical radar automatic target identification tasks such as airborne target surveillance and early warning. Existing research typically operates under the assumptions of closed sets and sufficient data. However, in real-world radar deployment and continuous detection scenarios, target identification systems usually begin with basic training using a large amount of data from cooperative targets (such as frequently appearing civilian airliners), and then continuously encounter unknown, non-cooperative incremental targets during actual operation. Since it is difficult to obtain a large amount of labeled data for non-cooperative targets, the system must quickly learn new categories without forgetting previously learned targets, even with only a small number of samples. This scenario constitutes the few-sample incremental learning problem for radar HRRP.

[0003] Existing incremental methods for few-shot classes mostly adopt a paradigm of "training and freezing the feature extractor in the basic stage, and applying a nonparametric classifier in the incremental stage." However, when directly applying this general algorithm to real-world radar aerial target recognition scenarios, it faces two prominent physical and technical limitations: (1) Insufficient diversity of basic categories leads to poor generalization ability. In actual engineering, the basic classes are usually civil aircraft with similar structural features, resulting in a lack of diversity in the feature space of the basic training set. Traditional data augmentation methods (such as flipping and translation) cannot effectively expand the distribution range between classes. When the feature extraction network is frozen, the system's ability to extract features from incremental new targets weakens, which can easily lead to feature aliasing between old and new categories.

[0004] (2) Attitude sensitivity leads to instability in scattering information. Changes in the attitude of an aerial target during flight can alter the radar line of sight, causing drastic fluctuations in the position and amplitude of the scattering center in HRRP, i.e., attitude sensitivity. Traditional coding directly extracts amplitude or position features, lacking a robust characterization of the statistical structure of scattering. Under conditions of few samples, it is prone to generating unreliable feature representations, exacerbating performance degradation during the incremental process.

[0005] Therefore, those skilled in the art are dedicated to developing an incremental recognition method for high-resolution range images of airborne targets using radar with few samples. By deeply integrating with the physical characteristics of radar detection and fusing prior knowledge of physical scales and statistical laws at the algorithmic level, this method addresses the specific engineering problems of insufficient training diversity and attitude changes leading to decreased recognition rates in practical radar target recognition systems. Summary of the Invention

[0006] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is the problem of declining recognition rate of radar target recognition systems.

[0007] To achieve the above objectives, the present invention provides a method for incremental identification of air target radar high-resolution range images with few samples, including radial scale-based category enhancement and using temporal scaling to simulate changes in the physical scale of the target to generate pseudo-category data.

[0008] Furthermore, the physical size of the target is scaled proportionally, and the spacing between strong scattering centers is changed without altering its topological sequence; by performing downsampling and upsampling on the HRRP signal, a pseudo-class with a different scale from the original target is constructed.

[0009] Furthermore, the HRRP signal is downsampled to simulate a small-sized target.

[0010] Furthermore, the HRRP signal is upsampled to simulate a large target.

[0011] Furthermore, Markov residual transition probability coding is included to model the HRRP amplitude sequence as a robust discrete state transition process.

[0012] Furthermore, the HRRP signal is divided into strong scattering, weak scattering, and clutter, and the transition probability between adjacent points in different physical states is statistically analyzed using Markov chains.

[0013] Furthermore, the Markov residual transition probability encoding serves as a spatial attention mechanism, forcing the network to suppress clutter and focus on stable, strongly physics-scattering regions.

[0014] Furthermore, it includes the following steps: Step 1: High-resolution range image data preprocessing and category enhancement; Step 2: Calculate the Markov residual transition probability encoding; Step 3: Feature extraction and stable attention fusion; Step 4: Incremental inference with few samples from nearest-neighbor prototypes.

[0015] Furthermore, during the training phase, preprocessing and class enhancement operations are first performed on the collected basic cooperative target HRRPs. Pseudo-classes representing different scales are generated through temporal linear interpolation and average pooling, and added to the basic training set to expand the inter-class distribution boundaries. Secondly, the training sequence is simultaneously fed into the deep extraction network and the Markov coding branch. The coding branch determines the physical state of the signal sampling points through amplitude thresholds, calculates the global first-order Markov state transition probability matrix, and outputs the residual transition probability distribution with respect to clutter states. Subsequently, the distribution is mapped to a spatial attention weight matrix, and residual addition is performed to fuse it with the shallow feature map of the backbone network to enhance the signal representation in regions resistant to amplitude fluctuations. Finally, the feature extractor is iteratively optimized using cross-entropy as the loss function.

[0016] Furthermore, during the deployment phase, after obtaining the unknown non-cooperative target HRRP, the system directly maps it to high-dimensional features through the frozen feature extractor and calculates the class mean prototype. Based on the Euclidean distance in the feature space, the system establishes the nearest class mean classification model to achieve the identification of both basic cooperative targets and incremental non-cooperative targets.

[0017] In real-world radar target recognition systems, the similar structures of basic cooperative targets (such as commercial airliners) lead to a sharp decline in model performance when generalizing to unknown non-cooperative targets. Traditional data augmentation methods fail to generate effective inter-class differentiation. This invention proposes a radial scale-based class augmentation (CA) method, combining the algorithm with the physical scale features of radar targets and generating pseudo-classes of different scales through temporal scaling. In real-world physical scenarios, radial scale is a key discriminative feature for distinguishing radar targets. Scaled-up the target's physical size proportionally is equivalent to changing the strong scattering center spacing without altering its topological sequence. By downsampling (simulating small targets) and upsampling (simulating large targets) the one-dimensional HRRP signal, pseudo-classes with significantly different scales from the original targets are constructed. These pseudo-classes are used as independent classes in basic training, forcing the network to learn a wider feature distribution boundary. This invention expands the feature distribution coverage of basic training, significantly improving the generalization ability of the frozen backbone network for unknown incremental targets. In real-world radar missions, it effectively reduces feature aliasing between old and new classes.

[0018] The attitude changes of aerial targets cause drastic fluctuations in the amplitude and position of the radar HRRP scattering center. Features extracted from a small number of samples are highly unstable, leading to severe forgetting in incremental learning performance. This invention proposes a Markov Residual Transition Probabilistic Encoding (MTC) method to model the HRRP amplitude sequence as a robust discrete state transition process. The HRRP of an aerial target physically consists of strong scattering, weak scattering, and clutter. After discretizing the HRRP into these three states, the state transition probabilities of adjacent positions are calculated. The "residual transition probability between a position and the clutter state" is used to quantify the probability that the position belongs to the true physical scattering center. The larger the residual transition probability, the more stable the region and the greater its discriminative contribution. This encoding is used as spatial attention weighting on the feature map, guiding the network to focus on stable scattering regions resistant to amplitude fluctuations. This invention overcomes the feature drift problem caused by directly relying on position or amplitude, improving the stability of feature representation. When identifying new targets with very few samples (e.g., 5 samples), it can extract more reliable prototype-like features, thereby significantly suppressing the performance degradation index (PD).

[0019] Technical Advantages: This invention directly addresses the continuous target identification needs in real-world scenarios such as airborne target surveillance and early warning. Without requiring retraining with old class samples, it generates pseudo-class data by simulating changes in the physical scale of the target through temporal scaling, thus mitigating the risk of model overfitting caused by the single aircraft type of the actual cooperative target. Simultaneously, it employs a Markov probability matrix to model the HRRP amplitude sequence as a robust discrete state transition process. This invention requires no complex coordinate transformations, has low computational overhead, and involves only one-dimensional signal interpolation, pooling, and lightweight convolution operations, enabling seamless integration into existing radar target identification systems.

[0020] Performance metrics: The core evaluation metrics are class-average accuracy and performance degradation (PD, i.e., the difference between the class-average accuracy of the base session and the final session). In the 1-way 5-shot incremental task of the HRRP eight-class target dataset for actual radar testing, such as... Figure 4 As shown, the class-average accuracy of this invention across five sessions was 98.57%, 87.45%, 84.89%, 81.42%, and 80.86%, with a PD of only 17.72%. This represents a 9.94% reduction compared to the baseline method (without CA and MTC, PD of 27.66%), and a further reduction of at least 3.96% compared to the state-of-the-art few-shot incremental class method (S3C, PD of 21.68%).

[0021] Production Implementation and Industrial Application Prospects: This invention is compatible with existing radar platforms and RATR systems. In production implementation, the basic stage involves offline training and solidification of the backbone model. When the deployed radar automatic target identification equipment encounters a small number of new aircraft target samples during surveillance and early warning missions, it can quickly calculate feature prototypes online and expand the target feature library, eliminating the burden of historical data communication backhaul and real-time retraining computational costs. In terms of industrial application, this method is suitable for identifying rapidly expanding new aircraft types in key area air defense and low-altitude economic scenarios. With the accumulation of radar data and the continuous expansion of the target library, this invention is conducive to building a sustainably evolving target identification capability system, possessing clear engineering implementation value and prospects for large-scale promotion.

[0022] Compared with the prior art, the present invention has the following obvious substantive features and significant advantages: 1. Addressing the limitations of computing power and data communication in actual operational deployments, this invention eliminates the need to reset or fine-tune the feature extractor during the incremental operation phase. It calculates the category feature prototype and expands the target library solely through forward inference using only a very small number of new samples. This mitigates the severe overfitting caused by catastrophic forgetting and small-sample fine-tuning, while significantly reducing the storage and computational overhead of system edge deployments.

[0023] 2. This approach overcomes the limitations of generalization ability under closed datasets. A class augmentation algorithm combining HRRP physical characteristics is proposed, utilizing temporal scaling to generate large and small pseudo-classes for augmented training. This method effectively widens the inter-class decision boundary for structurally similar aircraft targets.

[0024] 3. Attitude sensitivity interference caused by target flight maneuvers is suppressed. By using Markov residual transition probabilities, the amplitude fluctuations caused by attitude sensitivity are transformed into stable discrete transition statistics, locating the core physical scattering region that contributes robustly to HRRP discrimination. In the measured radar HRRP 1-way 5-shot incremental test of eight types of airborne targets, the performance degradation index (PD) decreased to 17.72%, demonstrating the effectiveness of this method for registering and identifying multiple rounds of continuously unknown targets.

[0025] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0026] Figure 1 This is a preferred embodiment of the scattering state division of the present invention; Figure 2 This is a preferred embodiment of the present invention: MTC statistical coding and fusion. Figure 3 This is an overall block diagram of a preferred embodiment of the present invention; Figure 4 This is a performance comparison experiment result of a preferred embodiment of the present invention. Detailed Implementation

[0027] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0028] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0029] This embodiment details the complete implementation process based on Class Augmentation (CA) and Markov Residual Transition Probabilistic Coding (MTC) for the few-shot incremental high-resolution range image recognition task in airborne target radar target identification. The scheme is divided into a basic training phase and a few-shot incremental inference phase.

[0030] 1) Overall technical solution

[0031] Step 1: High-resolution range image data preprocessing and category enhancement (CA).

[0032] Function and Principle: In radar target recognition systems, the basic targets available for training are mostly similar civil airliners, resulting in a limited sample category distribution. Utilizing the physical meaning of the radial scale of HRRP, scaling HRRP is equivalent to changing the physical size of the target in the time domain. Let the discrete time-domain HRRP vector be... .right Perform category enhancement to generate pseudo-category samples with enlarged and reduced physical sizes, thus amplifying the differences between base classes.

[0033] Specific method: Input length is Basic HRRP Discrete Sequence .

[0034] 1. Simulating large-size targets (upsampling): For Length is obtained by performing 2x linear interpolation Given a sequence, calculate the center of the sequence. Extract a truncation of length [value missing]. The fragments obtained .

[0035] 2. Simulating small-sized targets (downsampling): For Length is obtained by performing 2x average pooling. The sequence is zero-padded to restore its length to [the specified value]. get .Will and Marked as a new independent category label, compared with the true sequence Together they form the expanded base training set.

[0036] Step 2: Markov Residual Transition Probability Code (MTC) calculation.

[0037] Function and Principle: To overcome the drastic fluctuations in echo amplitude caused by changes in the attitude of aerial targets, this step does not directly use the absolute amplitude value of HRRP, but instead extracts a robust discrete state transition process resistant to amplitude fluctuations. The signal is divided into strong scattering, weak scattering, and clutter, and the transition probabilities between different physical states of adjacent points are statistically analyzed using a Markov chain. The higher the residual probability, the stronger the certainty that the location belongs to a stable physical scattering structure.

[0038] Specific methods: like Figure 1 As shown, state partitioning: for the sequence Each sampling point The amplitude threshold is mapped to the scattering state. ,in Corresponding to clutter, Corresponding to weak scattering, Corresponds to strong scattering.

[0039] Transition matrix calculation: Statistical sequence state transition matrix , of which elements Indicates from state Transition to state The probability of:

[0040] Here Given an indicator function, calculate the frequency of functions that satisfy the conditions. The transition matrix can then be obtained.

[0041] Extracting residual transfer coding: based on clutter state Calculate the bidirectional residual transition probability as a stability weight:

[0042] And spliced ​​together

[0043] Step 3: Feature extraction and stable attention fusion (basic training).

[0044] Function and Principle: During the training phase, the backbone neural network learns feature extraction through supervised learning. In this process, the MTC encoding generated in step two is used as a spatial attention mechanism, forcing the network to suppress clutter and focus on stable regions with strong physical scattering.

[0045] like Figure 2 As shown, the specific method is: Input Features are extracted from a one-dimensional ResNet network. After one-dimensional convolution Activation and Average Pooling The generated length is attention mask :

[0046] In the first residual block ( Afterwards, residual fusion connections are performed to refine the feature maps, and finally, the network parameters are updated using a multilayer perceptron (MLP) and cross-entropy loss. After basic training, the MLP classification head is discarded, and the backbone weights of the network are frozen.

[0047] Step 4: Incremental inference with few samples from nearest-neighbor prototypes.

[0048] Function and principle: In actual deployment, only a small number of HRRP samples are needed to complete the identifiable category amplification for unknown new targets, without the need to retrain the network, thus avoiding catastrophic forgetting.

[0049] Specific method: For the first For the new target data in the next incremental task, the feature vectors are extracted using the frozen backbone network, and the mean is calculated to obtain the class mean prototype. Facing the query sample to be tested Using Euclidean distance metric Execute the nearest neighbor algorithm and output the probability that the sample belongs to the known base class or each incremental class:

[0050] Workflow of this invention: During the training phase, the system first performs preprocessing and class enhancement operations on the collected HRRPs of basic cooperative targets (such as civil aircraft). Pseudo-classes representing different scales are generated through temporal linear interpolation and average pooling, and added to the basic training set to expand the inter-class distribution boundaries. Next, the training sequence is simultaneously fed into the deep extraction network and the Markov coding branch. The coding branch determines the physical state (clutter / weak scattering / strong scattering) of the signal sampling points using amplitude thresholds, calculates the global first-order Markov state transition probability matrix, and outputs the residual transition probability distribution for clutter states. Subsequently, this distribution is mapped to a spatial attention weight matrix and fused with the shallow feature map of the backbone network using residual addition to enhance signal representation in regions resistant to amplitude fluctuations. Finally, the feature extractor is iteratively optimized using cross-entropy as the loss function. During the testing (actual deployment) phase, after obtaining a very small number of unknown non-cooperative target HRRPs, the system directly maps them to high-dimensional features through the frozen feature extractor and calculates the class mean prototype. Based on the Euclidean distance in the feature space, a nearest class mean classification model is established to achieve the identification of both basic cooperative targets and incremental non-cooperative targets.

[0051] Application examples of this invention: This invention is directly applied to the continuous aerial target type discrimination task in an automatic radar target identification system. The system's initial target database contains only four frequently occurring civil aircraft types (Session 0): Boeing and Airbus. Historical radar HRRPs of these aircraft are used for offline model training. During the basic training phase, the system performs temporal interpolation and pooling to generate multi-scale pseudo-classes from the aircraft HRRP sequences for training. A Markov residual transfer model is used to encode the HRRPs, which are then weighted as spatial attention onto the feature map, guiding the network to focus on stable scattering regions resistant to amplitude fluctuations. Subsequently, the automatic radar target identification system is deployed. The radar initially collects HRRPs of unregistered unknown propeller aircraft, cargo planes, or helicopters. Due to their non-cooperative nature, only five HRRP samples are obtained for each target (1-way 5-shot). The automatic radar target identification system uses the previously frozen feature extractor to calculate the class mean prototype of these five samples, directly registering them as prototypes for the new target type. As new categories continue to emerge, the system continuously uses the nearest mean classifier to classify these new categories. Experimental results show that the system's performance degradation index for all subsequent new and old Type 8 targets is only 17.72%.

[0052] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and the scope of protection of the present invention is not limited to these specific embodiments.

[0053] This invention is applicable to aerial target classification and recognition tasks, and is particularly suitable for few-shot incremental learning scenarios. Taking the experimental HRRP dataset (containing 8 types of aerial targets: passenger planes, cargo planes, helicopters, etc.) as an example, a specific implementation of the network under the incremental learning setting is given. First, the overall framework is determined to be a two-stage process: a basic training stage and a few-shot incremental inference stage. The feature extractor adopts a hybrid structure of "one-dimensional residual network backbone + MTC attention branch".

[0054] The classification network and related modules are mainly divided into four key levels, covering the complete process from enhancement of raw physical features to statistical coding fusion: First layer: Data input and category augmentation layer (CA layer). During the basic training phase, this layer is responsible for processing the original baseline class HRRP sequence. (length is) Physical scale expansion is performed. The change in the target's physical size is simulated through time-domain scale transformation: 1. Upsampling path: for Performing double linear interpolation yields a length of The sequence, with a center cut length of... fragments Simulates the features of large-sized targets.

[0055] 2. Downsampling path: for Performing double average pooling yields a compressed sequence, which is then zero-padded to restore a length of [length missing]. fragments This simulates the features of small-sized targets. The resulting pseudo-classes... and They are assigned independent labels and fed into the subsequent network along with the original samples to expand the inter-class boundaries of the basic training.

[0056] The second layer: Markov Statistical Coding (MTC) layer: This layer runs in parallel with the backbone network and is responsible for extracting statistical structural features that are resistant to amplitude fluctuations. First, the input sampling points are mapped to clutter. Weak scattering Strong scattering Three scattering states were identified. The first-order Markov transition matrix was then calculated. And calculate the residual transition probability encoding, and concatenate the encoding matrix. (shape is) This characterizes the certainty that each sampling point belongs to a stable physical scattering center.

[0057] The third layer: Feature extraction backbone and attention fusion layer: The backbone network adopts a one-dimensional ResNet18 structure. To incorporate physical statistical laws into deep features, this invention designs a shallow fusion mechanism: 1. Attention generation: The attention obtained from the second layer After one-dimensional convolution, sigmoid gating, and average pooling, shape and shallow feature maps of the backbone (such as...) are generated. Consistent spatial attention mask .

[0058] 2. Fusion Mechanism: In the first residual block of the backbone network ( After that, apply the attention mask. and The output feature maps are fused using addition or weighted multiplication. This shallow fusion ensures that the model accurately captures stable physical scattering centers before feature semantic abstraction.

[0059] Fourth layer: Feature mapping and incremental classification layer: The backbone network is subsequently composed of... to Further extract high-dimensional semantic features and output the final feature vector. Training phase: In The MLP classification head is then connected, and cross-entropy loss is used to drive model learning. After training, the MLP is discarded and the entire feature extractor is frozen. Inference phase: For new categories in incremental sessions (e.g., 1-way 5-shot), the mean features of a small number of samples are extracted from the frozen network as class prototypes. Query samples are then evaluated using Euclidean distance metrics with all known (base + incremental) class prototypes to achieve the final discriminative output.

[0060] Aside from CA-enhanced input during the basic training phase, the model relies solely on the frozen feature extractor during inference. Compared to traditional methods, this invention guides the network to focus on stable physical scattering points through the MTC attention mechanism, achieving robust recognition results in pose-sensitive scenes without complex coordinate transformations. An overall block diagram of the example is shown below. Figure 3 As shown.

[0061] This embodiment presents an implementation process based on category augmentation and Markov residual transition probability encoding for incremental high-resolution distance image recognition tasks with few sample classes. In the basic stage, the scheme expands the training distribution through "category augmentation." In the encoding stage, it constructs positional weights and refines feature representations through "Markov residual transition probability encoding." In the incremental stage, replay retraining is not performed; instead, a prototype is constructed using only a small number of new class samples to complete category expansion. This reduces performance degradation during the incremental process and improves the final recognition stability.

[0062] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An air target radar high-resolution range profile few-sample class incremental identification method, characterized in that, This includes radial scale-based category augmentation, which uses temporal scaling to simulate changes in the physical scale of the target to generate pseudo-class data.

2. The airborne target radar high resolution range profile few sample class incremental identification method of claim 1, wherein, The target's physical size is scaled proportionally, and the spacing between strong scattering centers is changed without altering its topological sequence. By downsampling and upsampling the HRRP signal, a pseudo-class with a different scale from the original target is constructed.

3. The airborne target radar high resolution range profile few sample class incremental identification method of claim 2, wherein, The HRRP signal is downsampled to simulate a small target.

4. The airborne target radar high resolution range profile few sample class incremental identification method of claim 2, wherein, Upsampling of the HRRP signal is performed to simulate a large target.

5. The airborne target radar high resolution range profile few sample class incremental identification method of claim 1, wherein, Including Markov residual transition probability coding, the HRRP amplitude sequence is modeled as a robust discrete state transition process.

6. The airborne target radar high resolution range profile few sample class incremental identification method of claim 5, wherein, The HRRP signal is divided into strong scattering, weak scattering and clutter, and the transition probability between adjacent points in different physical states is statistically analyzed using Markov chains.

7. The airborne target radar high resolution range profile few sample class incremental identification method of claim 5, wherein, The Markov residual transition probability coding serves as a spatial attention mechanism, forcing the network to suppress clutter and focus on stable, strongly physics-scattering regions.

8. The method for incremental identification of small sample classes in high-resolution range images of airborne targets by radar as described in claim 1, characterized in that, Includes the following steps: Step 1: High-resolution range image data preprocessing and category enhancement; Step 2: Calculate the Markov residual transition probability encoding; Step 3: Feature extraction and stable attention fusion; Step 4: Incremental inference with few samples from nearest-neighbor prototypes.

9. The airborne target radar high resolution range profile few sample class incremental identification method of claim 1, wherein, During the training phase, preprocessing and class augmentation operations are first performed on the collected basic cooperative target HRRPs. Pseudo-classes representing different scales are generated through temporal linear interpolation and average pooling, and added to the basic training set to expand the inter-class distribution boundaries. Next, the training sequence is simultaneously fed into the deep extraction network and the Markov coding branch. The coding branch determines the physical state of the signal sampling points through the amplitude threshold, calculates the global first-order Markov state transition probability matrix, and outputs the residual transition probability distribution with respect to the clutter state. Then, the distribution is mapped to the spatial attention weight matrix and fused with the shallow feature map of the backbone network by residual addition to enhance the signal representation in the region resistant to amplitude fluctuations. Finally, the feature extractor is iteratively optimized using cross-entropy as the loss function.

10. The airborne target radar high resolution range profile few sample class incremental identification method of claim 9, wherein, During the deployment phase, after obtaining the unknown non-cooperative target HRRP, the system directly maps it to high-dimensional features through the frozen feature extractor and calculates the class mean prototype. Based on the Euclidean distance in the feature space, the nearest class mean classification model is established to achieve the identification of both basic cooperative targets and incremental non-cooperative targets.

Images