Radar HRRP deformation target identification method based on feature correction

Abstract

The invention discloses a radar HRRP deformation target recognition method based on feature correction, and the method comprises the steps: carrying out the preprocessing of the radar one-dimensional HRRP echo data of a to-be-recognized target, and obtaining the HRRP original target data; superposing block sparse signals with different sparseness and Gaussian white noise on the HRRP original target data to obtain HRRP deformation target data; constructing a lightweight feature extraction network, and performing pre-training by using HRRP original target data; a feature correction network comprising a multi-dimensional parallel feature correction module, a feature fusion module and a de-noising post-processing module is constructed, and training is carried out after the feature correction network is cascaded with the feature extraction network; and combining the feature extraction network with the feature correction network to obtain a deformation target identification model so as to realize identification and classification of a deformation target. According to the method, the intrinsic feature distribution of the deformed target can be effectively recovered, and the recognition accuracy and generalization ability of the non-cooperative target in a low signal-to-noise ratio environment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of radar target detection and recognition technology, specifically relating to a radar HRRP deformation target recognition method based on feature correction. Background Technology

[0002] High-resolution range profiles (HRRPs), as one-dimensional electromagnetic scattering features, can accurately depict the projection distribution and scale information of the target's scattering center along the radar line of sight. With its advantages of convenient data acquisition, low dimensionality, and mature processing algorithms, HRRP has become a key data source in the field of Radar Automatic Target Recognition (RATR).

[0003] Currently, most existing recognition models are based on ideal detection assumptions, namely, a high signal-to-noise ratio environment, a complete and unobstructed target structure, and consistent distribution between the training and test sets. However, in actual non-cooperative target detection tasks, these ideal conditions are often difficult to meet. Due to the complexity of the battlefield environment and the variability of the target itself, measured data often exhibit distortion or loss of some range cell scattering information, leading to severe feature mismatch between the test samples and the preset template.

[0004] Such target deformation mainly stems from two factors: first, non-rigid changes in target mounting configuration, such as adjustments to weapon mountings or external components due to mission requirements, directly altering the target's electromagnetic scattering topology; and second, environmental obstruction or camouflage interference causing information truncation of the target's local structure in the radar line-of-sight direction. Especially in low signal-to-noise ratio environments, the additive interference of noise and the structural deformation of the target exhibit a complex coupling effect, further obscuring effective scattering characteristics, posing a severe challenge to the robustness of traditional identification methods.

[0005] To address the aforementioned issues, achieving accurate interpretation of HRRP under the dual constraints of low signal-to-noise ratio and deformation interference has become a critical technical challenge that urgently needs to be overcome in the field of radar signal processing.

[0006] In recent years, research on HRRP (High-Resolution Range Image Recognition) for deformable targets has gradually attracted attention from the academic community. Existing representative methods mainly fall into two categories: one is the high-resolution range image recognition method for variant targets based on block sparse Bayes. For example, patent application number 201810978483.0 discloses a "high-resolution range image recognition method for variant targets based on block sparse Bayes." This method constructs a mathematical model of the variant target, defines the prior probabilities of variables and their parameter distributions, and uses a block sparse Bayes iterative algorithm to solve and separate the variant components in the echo. Subsequently, the system removes the variant component and uses an adaptive Gaussian classifier to classify and recognize the reconstructed clean range image. Although this scheme reduces the interference of variant factors on recognition accuracy to some extent, it is essentially still a step-by-step processing architecture of "recovery first, recognition later." This cascaded processing flow makes the final recognition performance highly dependent on the recovery quality of the front end. Once there are residuals in the recovery stage, it will directly lead to the distortion of subsequent classification features. Furthermore, phased independent optimization cannot achieve end-to-end global optimization, easily leading to the accumulation and propagation of system errors, thus limiting further improvement in the overall recognition rate. Another type is the high-resolution range profile method for identifying variant aircraft based on convolutional neural networks. For example, patent application number 201910201317.4 discloses a "high-resolution range profile method for identifying variant aircraft based on convolutional neural networks." This method aims to mine deep common features between variant and non-variant target HRRP data using convolutional neural networks, utilizing these invariant features to alleviate feature mismatch problems caused by shape changes, thereby improving the robustness of recognition. However, this algorithm has significant limitations in practical applications: on the one hand, its feature extraction process highly depends on the completeness of target information; once the key scattering center region is occluded, the extraction of common features will be severely interfered with, leading to a significant decrease in recognition accuracy; on the other hand, the model is only trained based on samples under ideal conditions, without fully considering the impact of complex battlefield environments such as low signal-to-noise ratios, resulting in limited generalization ability of the model under non-ideal observation conditions.

[0007] In summary, existing HRRP identification methods for deformable targets have the following main shortcomings: 1. Error accumulation and propagation: Existing methods mostly adopt a cascaded step-by-step processing architecture of "recover first, then identify", which makes the final identification performance highly dependent on the quality of the front-end recovery. Once there are residuals in the recovery stage, it will directly cause distortion of subsequent classification features, and independent optimization in each stage cannot achieve end-to-end global optimization.

[0008] 2. Difficulty in extracting key features: The feature extraction process of existing algorithms is highly dependent on the completeness of target information. Under non-cooperative observation, once the key scattering center region is occluded or deformed, the extraction of common features will be severely interfered with, making it difficult for the model to handle structural defects and significantly reducing recognition accuracy.

[0009] 3. Poor generalization ability: Existing models are mostly trained on samples under ideal conditions, without fully considering the impact of complex battlefield environments such as low signal-to-noise ratios. This results in limited generalization ability of the models when faced with non-ideal observation conditions, making it difficult to adapt to the variable noise interference in real-world scenarios. Summary of the Invention

[0010] To address the problem that existing deformable target recognition methods experience rapid performance degradation in low signal-to-noise ratio (SNR) environments, this invention provides a radar HRRP deformable target recognition method based on feature correction. The technical problem to be solved by this invention is achieved through the following technical solution: This invention proposes a radar HRRP deformation target recognition method based on feature correction, comprising: S1. Acquire the radar one-dimensional HRRP echo data of the target to be identified, and perform preprocessing operations on the radar one-dimensional HRRP echo data, and use the preprocessed data as the original HRRP target data. S2. Add block sparse signals of different sparsity as deformation components to the original HRRP target data, and further add Gaussian white noise to obtain HRRP deformable target data; associate the HRRP deformable target data, the original HRRP target data and the corresponding target category labels to construct a pairwise sample set for subsequent network training. S3. Construct a lightweight feature extraction network. Use the original target data of HRRP to pre-train the feature extraction network to obtain a pre-trained feature extraction network. S4. Construct a feature correction network that includes a multi-dimensional parallel feature correction module, a feature fusion module based on a multi-branch attention mechanism, and a denoising post-processing module based on an improved Pureformer network, and cascade it with the feature extraction network; fix the parameters of the feature extraction network, use paired sample sets as input data, and optimize the parameters of the feature correction network through joint training to obtain a trained feature correction network. S5. Combine the feature extraction network obtained from the pre-training in step S3 with the feature correction network obtained from the training in step S4 to obtain a deformable target recognition model based on feature correction and Pureformer post-processing, so as to use the model to recognize and classify deformable targets.

[0011] The beneficial effects of this invention are: 1. The radar HRRP deformable target recognition method based on feature correction provided by this invention first constructs a sample set including HRRP original target data and HRRP deformable target data based on the radar echo data of the target to be identified; then, a lightweight feature extraction network is constructed and pre-trained using the HRRP original target data; next, a feature correction network is constructed including a multi-dimensional parallel feature correction module, a feature fusion module based on a multi-branch attention mechanism, and a denoising post-processing module based on an improved Pureformer network, and is cascaded with the feature extraction network. By fixing the parameters of the feature extraction network, joint training is performed using the sample set; finally, the trained feature extraction network and feature correction network are combined to obtain a deformable target recognition model based on feature correction and Pureformer post-processing, so as to facilitate the recognition and classification of deformable targets using the model. This method introduces a correction module after the feature extraction network. Through a multi-dimensional correction mechanism, it preserves the integrity of the target features, effectively avoids the loss of key information, and significantly improves the robustness of the recognition algorithm. At the same time, this method can operate directly at the feature level, avoiding errors that may be introduced by signal reconstruction, and ensuring a high degree of consistency between the corrected deformed target features and the original target features in the discrimination space. It has achieved excellent recognition accuracy under low signal-to-noise ratio and complex deformation conditions.

[0012] 2. Based on the multidimensional feature correction and recognition framework, this invention introduces a feature fusion module based on a three-branch attention mechanism. This module can adaptively recalibrate the fusion ratio of multi-branch features. Compared with simple feature fusion methods that only use convolution, the TAFF module proposed in this invention dynamically adjusts the weights through the attention mechanism, effectively avoiding the problem of performance degradation caused by the model's over-reliance on a specific feature, and ensuring the balance and effectiveness of feature fusion. At the same time, this adaptive weight allocation strategy significantly enhances the robustness of the network model, enabling it to adapt to more complex application scenarios. Ablation experiments show that the introduction of this module significantly improves the recognition accuracy and reliability of this invention for HRRP deformable targets.

[0013] 3. Building upon multidimensional feature correction and feature fusion based on a three-branch attention mechanism, this invention proposes an efficient denoising post-processing mechanism based on an improved Pureformer network for deep denoising. On one hand, this network fully leverages the powerful global modeling capabilities of the Transformer architecture, effectively addressing the long-distance dependency problem in HRRP signal denoising. On the other hand, by designing a multi-head dynamic transpose attention mechanism and introducing a local enhancement module, this invention effectively overcomes the technical shortcomings of the standard Transformer model, such as high computational complexity and weak ability to capture local details. While ensuring computational efficiency, it achieves accurate preservation of local subtle features of HRRP, significantly enhancing the robustness of the entire network architecture in feature correction under low signal-to-noise ratio environments.

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 A schematic flowchart of the radar HRRP deformation target recognition method based on feature correction provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a deformable target recognition model based on feature correction and Pureformer post-processing provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of a one-dimensional high-resolution range echo of a non-deformable aircraft in a simulation experiment. Figure 4 This is a schematic diagram of the deformation components of a non-deformable aircraft in a simulation experiment. Figure 5 for Figure 3 Correspondingly, a high-resolution echo diagram of a deformed aircraft at a signal-to-noise ratio of 5dB is shown. Figure 6 To obtain feature distribution maps of various targets before and after feature correction of deformed and original targets using the method of the present invention at a signal-to-noise ratio of 30dB; Figure 7 To illustrate the feature distribution of various targets before and after feature correction of deformed and original targets using the method of this invention at a signal-to-noise ratio of 5dB. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The first aspect of the present invention provides a radar HRRP deformable target recognition method based on feature correction. In a low signal-to-noise ratio environment, the method corrects the feature dimensions of deformable targets, making the corrected features more consistent with the features of the original target (excluding the deformable component), thereby improving the recognition of deformable targets.

[0018] The approach to achieving the objective of this invention is as follows: First, a baseline feature space is constructed using a pre-trained lightweight backbone network with an integrated channel attention mechanism. Initial feature extraction is performed on the original target and variant targets to serve as input for subsequent processing. Second, a feature correction network is constructed, utilizing parallel channel correction and spatial correction modules to perform multi-dimensional correction on the input original features from the channel and spatial domains, respectively, to obtain corresponding channel-corrected features and spatial-corrected features. Third, a three-branch attentional feature fusion (TAFF) module is used to adaptively weight and fuse the original features, channel-corrected features, and spatial-corrected features. This fusion method effectively models the relative importance between different features, improving the expressive power of the fusion result. The fused features are then input into an improved Pureformer network for deep denoising and refined reconstruction. This network utilizes the Transformer... Leveraging the long-range modeling capabilities and the local perception advantages of multi-scale convolution, this method aims to accurately reconstruct target features in the latent space that are close to the original noise-free state. Finally, by constructing a supervision mechanism through the joint multi-constraint loss function, the distribution consistency between the corrected variant target features and the original target features is maximized. This results in a deformable target recognition method based on feature correction and Pureformer post-processing that has high recognition performance, low computational complexity, and meets real-time requirements in low signal-to-noise ratio environments.

[0019] For details, please see Figure 1 , Figure 1 This is a flowchart illustrating the radar HRRP deformation target recognition method based on feature correction provided in an embodiment of the present invention. The method mainly includes the following steps: S1. Acquire the radar one-dimensional HRRP echo data of the target to be identified, and perform preprocessing operations on the radar one-dimensional HRRP echo data, and use the preprocessed data as the original HRRP target data.

[0020] Specifically, for the acquired one-dimensional HRRP echo data from the radar, preprocessing operations such as pulse compression, amplitude extraction, range alignment, amplitude normalization, and noise reduction are performed sequentially to obtain smoothed, aligned, and normalized echo amplitude data. This processed data is then used as the original HRRP target data, denoted as... .

[0021] S2. Add block sparse signals of different sparsity as deformation components to the original HRRP target data, and further add Gaussian white noise to obtain HRRP deformed target data; associate the HRRP deformed target data, the original HRRP target data and the corresponding target category labels to construct a pairwise sample set for subsequent network training.

[0022] It should be noted that since deformation data is difficult to obtain in real-world scenarios, this embodiment utilizes the sparsity characteristics of the deformation target block to randomly generate block sparse signals with different sparsity, which are then superimposed onto the original target (excluding deformation components) as deformation data. At the same time, in order to simulate noise in real-world scenarios, noise of different intensities is superimposed on the deformation data to realistically simulate deformation data in the environment.

[0023] Specifically, based on the structured sparsity characteristics of deformation components, a block sparse model is used to simulate target deformation. On the basis of the original HRRP target data, block sparse signals with different sparsities are superimposed as deformation components, and Gaussian white noise is further superimposed to simulate the actual low signal-to-noise ratio electromagnetic environment, thereby generating HRRP deformed target data containing noise and deformation, denoted as . Finally, the HRRP deformed target data, the original HRRP target data, and the corresponding target category labels are associated to construct a pairwise sample set for subsequent network training.

[0024] S3. Construct a lightweight feature extraction network. Use the original target data from HRRP to pre-train the feature extraction network to obtain a pre-trained feature extraction network.

[0025] Optionally, as an implementation method, the feature extraction network in this embodiment adopts an Ultra-Lightweight Residual Squeeze-and-Excitation Network (UL-ResSENet) structure, or adopts a network structure based on ResNet or DenseNet, or adopts a network structure based on channel attention or coordinate attention.

[0026] It is understood that, in addition to the network structures listed above, this embodiment can also use other one-dimensional convolutional neural networks with residual structures and attention mechanisms as the benchmark feature extractor. As long as they can effectively extract the discriminative features of high-resolution distance images, they can achieve the same or similar feature expression effects.

[0027] Preferably, this embodiment uses the UL-ResSENet network structure as the feature extraction network. The UL-ResSENet network includes an input layer, an enhancement residual layer, and an output layer. The feature extraction network is pre-trained using the original HRRP target data, including: S31. Input the original target data of HRRP into the input layer of the feature extraction network, and perform preliminary feature extraction and normalization processing including batch normalization, nonlinear activation and random deactivation regularization in sequence to obtain the first feature.

[0028] The specific processing procedure is as follows: 31a) Convolutional Feature Extraction Specifically, the input layer first performs a one-dimensional convolution operation on the original target data, using a kernel size of 7 and a stride of 1, and symmetrical padding on both sides of the input data to ensure that the feature dimension remains consistent before and after convolution. Through this convolution operation, local structural features in the HRRP signal are extracted using the local receptive field mechanism, thereby obtaining a low-level feature representation, laying the foundation for deep feature extraction in subsequent network layers.

[0029] 31b) Batch normalization processing After the convolution operation, the input data undergoes batch normalization. Batch normalization normalizes each feature dimension, reducing internal covariate bias during training. Specifically, it adjusts the mean of each feature to 0 and the variance to 1, making the input distribution of each layer more stable, thereby accelerating training and improving the model's stability and convergence speed. After the convolution operation, batch normalization is performed on the convolution output features. By standardizing the feature distribution, its mean approaches zero and its variance approaches one, effectively reducing internal covariate bias during training and improving the stability and convergence speed of the network training process.

[0030] 31c) Nonlinear activation processing The batch-normalized features are further nonlinearly mapped using the ReLU activation function, introducing nonlinear expressive power so that the network can characterize more complex feature patterns in the HRRP signal.

[0031] 31d) Random inactivation regularization At the end of the input layer, a Dropout random deactivation mechanism is introduced, which randomly discards some neuron outputs during training to reduce the network's over-reliance on local features, suppress overfitting, and thus improve the model's generalization ability and robustness.

[0032] This embodiment inputs the raw target data of HRRP into the input layer of the feature extraction network UL-ResSENet to perform preliminary feature extraction and normalization processing on the raw data, so as to provide stable and effective input features for subsequent deep feature learning.

[0033] S32. The first feature is input into the enhancement residual layer of the feature extraction network, and multi-layer convolution processing, feature enhancement processing based on channel attention weighting mechanism, and residual connection fusion processing are performed in sequence to obtain the second feature.

[0034] The specific processing procedure is as follows: 32a) Multi-layer convolution feature enhancement Specifically, after the first feature is input into the enhancement residual layer, the input feature is first processed through multiple convolutional operations. Each convolutional layer combines batch normalization and ReLU activation function to continuously enhance the local structural features of the HRRP signal, enabling the network to extract richer and more complex discriminative features step by step.

[0035] 32b) Channel attention weighting mechanism To enhance the network's ability to focus on key feature channels, a channel attention mechanism is introduced into the residual enhancement module. This mechanism first aggregates global information from the features of each channel using global average pooling to obtain statistics reflecting the overall response strength of the channels. Then, it models the correlation between channels through two-stage convolutional mapping, generating importance weights for each channel. Finally, it maps the weights to the 0, 10, 10, 1 interval using a sigmoid activation function to achieve adaptive weighting of features from different channels, thereby enhancing key channel features and suppressing redundant or irrelevant features.

[0036] 32c) Residual Connectivity Fusion Specifically, the feature output processed by convolution and channel attention mechanisms is fused with the input features through residual connections. This allows the network to retain the original feature information while learning feature transformations, thereby effectively alleviating the gradient vanishing problem in deep network training and improving the stability and expressive power of the feature extraction process.

[0037] This embodiment improves the network's ability to represent key target features by inputting the preprocessed features from the input layer into the enhancement residual module for further in-depth feature extraction and optimization.

[0038] S33. Input the second feature into the output layer of the feature extraction network, perform feature aggregation and mapping, and generate a primary feature representation.

[0039] The specific processing procedure is as follows: First, the temporal dimension features are compressed using adaptive average pooling to obtain a global feature vector with fixed dimensions. Then, a multi-layer fully connected network is used to map and fuse the feature vectors, and a non-linear activation function is used to constrain the feature distribution. Finally, the output features are normalized using the Tanh activation function to obtain the final output result for target classification or recognition.

[0040] S34. Use the cross-entropy loss function to backpropagate the feature extraction network, optimize the parameters of the feature extraction network iteratively, until the pre-training is completed.

[0041] Thus, the pre-trained UL-ResSENet network, i.e. the feature extraction network, is obtained.

[0042] S4. Construct a feature correction network that includes a multi-dimensional parallel feature correction module, a feature fusion module based on a multi-branch attention mechanism (Temporal Amplitude Feature Fusion, TAFF), and a denoising post-processing module based on an improved Pureformer network, and cascade it with the feature extraction network; fix the parameters of the feature extraction network, use paired sample sets as input data, and optimize the parameters of the feature correction network through joint training to obtain a trained feature correction network.

[0043] As one implementation method, the multi-dimensional parallel feature correction module constructed in this embodiment can adopt a parallel spatial correction and channel correction structure, or a multi-scale self-attention mechanism, a non-local attention mechanism, or a feature relationship modeling method based on graph neural networks to perform correlation modeling and correction on the feature representation output by the feature extraction network to obtain multi-dimensional corrected features.

[0044] It should be noted that, without changing the core idea of ​​the invention, other forms of attention mechanisms or feature recalibration methods can be used as alternatives, which can also achieve compensation and repair of deformed or occluded features.

[0045] Furthermore, the feature fusion module based on the multi-branch attention mechanism can use a three-branch attention fusion network structure to fuse multi-dimensional corrective features to obtain fused features.

[0046] In addition, the denoising post-processing module based on the improved Pureformer network includes a feature enhancement module, a multi-head deep separable self-attention module, and a gated feedforward network, which are used to perform contextual relationship modeling and global dependency enhancement on the fused features to obtain the final feature representation for target recognition and classification.

[0047] In another embodiment of the present invention, a Transformer variant optimized for one-dimensional sequences, a hybrid network of convolution and attention, or a self-attention network based on temporal modeling can also be used as an alternative to the improved Pureformer network. As long as noise can be effectively suppressed under low signal-to-noise ratio conditions and the consistency of feature structure can be maintained, similar technical effects as the present invention can be achieved.

[0048] By cascading and combining the UL-ResSENet network, spatial correction and channel correction modules, a feature fusion module based on a multi-branch attention mechanism, and a denoising post-processing module based on an improved Pureformer network, and through training, the deformable target recognition model based on feature correction and Pureformer post-processing of this invention can be obtained. Please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shows the structural block of the deformable target recognition model based on feature correction and Pureformer post-processing provided in this embodiment of the invention.

[0049] The following is based on Figure 2 The model framework shown introduces the training process of the feature correction network. In S4, paired sample sets are used as input data, and the parameters of the feature correction network are optimized through joint training, including: S41. Analyze the HRRP deformation target data from the paired sample sets respectively. Compared with HRRP original target data After feature extraction using a feature extraction network, the corresponding deformation feature representation is obtained. Compared with the original feature representation .

[0050] S42, respectively and Spatial dimension correction is performed using a spatial attention network to generate positional importance weights along the distance dimension, resulting in spatially corrected features. and The spatial attention network consists of two layers of one-dimensional convolution, a non-linear activation function, and a normalization function.

[0051] 42a) respectively and In the input spatial attention network, the number of channels of the input features is compressed by the first layer of one-dimensional convolution, and local spatial structure information is extracted while keeping the distance dimension unchanged to obtain the first spatial features.

[0052] Here, the number of channels of the input features is compressed by the first layer of one-dimensional convolution, and local spatial structure information is extracted while keeping the distance dimension unchanged. This allows for the integration of common expressions related to spatial location in multi-channel features, thereby improving feature representation.

[0053] 42b) The first spatial features are processed using the ReLU nonlinear activation function to obtain the second spatial features.

[0054] This embodiment uses the ReLU activation function to introduce nonlinear discrimination capability, which can enhance the ability to distinguish between stable structural regions and abnormal disturbance regions.

[0055] 42c) The second spatial features are mapped to single-channel spatial responses through a second-layer one-dimensional convolution to obtain the third spatial features.

[0056] In this embodiment, the compressed spatial features are mapped into a single-channel spatial response through a second-layer one-dimensional convolution to represent the importance of each distance unit, thereby improving the representation of important features.

[0057] 42d) The spatial response of the third spatial feature is normalized to the 0-1 interval using the Sigmoid function, thus obtaining the spatially corrected feature. and .

[0058] This embodiment normalizes the spatial response to the 0-1 range using the Sigmoid function, forming spatial attention weights with clear physical meaning, thereby enhancing the features of the target stable region and suppressing the features of the occlusion interference region in subsequent processing.

[0059] S43, respectively and Channel attention networks are used to correct the channel dimension, modeling the importance of each feature channel along the channel dimension, thus obtaining the channel-corrected features. and .

[0060] 43a) respectively and In the input channel attention network, the features of each channel are globally aggregated through adaptive average pooling to obtain the features of the first channel.

[0061] This embodiment uses adaptive average pooling to globally aggregate the features of each channel, compressing the distance dimension information into a single statistic, thereby obtaining a global description that reflects the overall response strength of the channel.

[0062] 43b) One-dimensional convolution is used to compress the channel dimension of the first channel feature, and then combined with the ReLU activation function to obtain the second channel feature.

[0063] This embodiment uses one-dimensional convolution to compress the channel dimension and combines it with the ReLU activation function to introduce nonlinear modeling capabilities, thereby uncovering the correlation between different channels.

[0064] 43c) The second channel features are mapped back to the original number of channels through a second layer of one-dimensional convolution to generate the response weights of each channel, thus obtaining the third channel features.

[0065] 43d) The weights of the third channel features are normalized to the 0-1 interval using the Sigmoid function, thus obtaining the channel-corrected features. and .

[0066] This embodiment uses the Sigmoid function to normalize the weights to the 0-1 range, forming channel attention coefficients with clear physical meaning. These coefficients are used to enhance feature channels that contribute significantly to target discrimination and suppress redundant or occluded channel features in subsequent processing.

[0067] The channel attention network described above can be used to model the importance of each feature channel in the channel dimension of the feature map.

[0068] This invention introduces a correction module after the feature extraction network to adaptively correct the extracted features in both spatial and channel dimensions. This minimizes the distribution distance between the deformed target and the original target in the corrected feature space, enabling efficient identification of both HRRP deformed and original targets simultaneously. Compared to traditional common feature extraction methods, this invention does not simply discard deformation components but preserves the integrity of target features through a correction mechanism, effectively avoiding the loss of key information and significantly improving the robustness of the recognition algorithm. Compared to recognition methods based on target reconstruction, this invention operates directly at the feature level, avoiding errors that may be introduced by signal reconstruction and ensuring a high degree of consistency between the corrected deformed target features and the original target features in the discrimination space. Experimental results show that the unified framework proposed in this invention can effectively construct a unified feature space applicable to both deformed and original targets, achieving excellent recognition accuracy under low signal-to-noise ratio and complex deformation conditions.

[0069] S44. Simultaneously process the two feature paths. and Adaptive weighted fusion is performed using a three-branch attention network to obtain the fused deformation HRRP features. Compared with the original HRRP features .

[0070] 44a) Two-way features and The data are input into a multi-scale convolutional attention network, where a fusion model is performed, including global average pooling and local one-dimensional convolution, resulting in two first fusion features.

[0071] Specifically, the core of the TAFF network consists of a Multi-Scale Convolutional Attention Network (MSCAM), which first fuses and models the two input features: on the one hand, it extracts channel-level global context information through global average pooling to characterize the importance of each feature channel as a whole; on the other hand, it directly models the local response on the original features through a local one-dimensional convolutional structure, preserving fine-grained feature differences.

[0072] 44b) The two first fusion features are processed by channel compression convolution, batch normalization and ReLU activation respectively, and then channel recovery convolution is used to generate an attention response with the same number of channels as the original, thus obtaining two second fusion features.

[0073] This involves introducing nonlinear expressive capabilities and reducing redundant information.

[0074] 44c) The two second fusion features are weighted and fused using the Sigmoid function to obtain the fused deformation HRRP features. Compared with the original HRRP features .

[0075] The Sigmoid function outputs fusion weights between 0 and 1, which are used to weight and combine the two input features. This allows the network to adaptively balance the roles of different feature branches at the spatial and channel levels based on feature reliability and discriminative contribution, thus achieving more robust and effective feature fusion and correction.

[0076] The feature fusion module based on the multi-branch attention mechanism ensures adaptive weighted fusion of features from different branches or different enhancement methods, thus avoiding information redundancy and noise amplification caused by simple addition or splicing.

[0077] This invention introduces a feature fusion module based on a three-branch attention mechanism on the foundation of the feature correction and recognition framework. This module can adaptively recalibrate the fusion ratio of multi-branch features. Compared with simple feature fusion methods that only use convolution, the TAFF module proposed in this invention dynamically adjusts the weights through the attention mechanism, effectively avoiding the problem of performance degradation caused by the model's over-reliance on a specific feature, and ensuring the balance and effectiveness of feature fusion. At the same time, this adaptive weight allocation strategy significantly enhances the robustness of the network model, enabling it to adapt to more complex application scenarios. Ablation experiments show that the introduction of this module significantly improves the recognition accuracy and reliability of this invention for HRRP deformable targets.

[0078] S45. Analyze the deformation HRRP features respectively. Compared with the original HRRP features After feature denoising based on an improved Pureformer network, the final feature representation for target recognition and classification is obtained.

[0079] Specifically, the Pureformer network's role is to model the contextual relationships and enhance global dependencies of the fused features, compensating for the shortcomings of convolutional structures in long-range dependency modeling, thereby improving the reconstruction and correction capabilities of occluded features. It is composed of a Feature Enhancement Block (FEB), a Multi-Head Depthwise Separable Self-Attention Module (MDTA), and a Gated Deep Feed-Forward Network (GDFN).

[0080] Unlike the high computational complexity and sensitivity to noise of the standard Transformer, this invention designs an improved Pureformer structure for one-dimensional radar sequences. It replaces traditional self-attention with a multi-head dynamic transposed attention mechanism (MDTA), significantly reducing computation and enhancing global modeling capabilities by calculating the attention map along the distance dimension rather than the sequence length. Furthermore, it introduces a gated convolutional feedforward network (GDFN) specifically designed to suppress residual noise introduced during feature reconstruction.

[0081] Specifically, S45 includes the following steps: 45a) Deformation HRRP features respectively Compared with the original HRRP features In the input feature enhancement module, the input features are modeled at multiple scales through a parallel one-dimensional convolutional structure, and the features at different scales are fused and compressed to obtain the first denoised features.

[0082] Specifically, FEB models the input features at multiple scales through parallel one-dimensional convolutional structures. 1×1 convolutions are used to extract channel-level combined features, while 3×1 and 5×1 convolutions are used to capture local and larger-scale distance structure information, respectively. Then, the multi-scale features are fused and compressed to enhance the spatial expressiveness of the features.

[0083] 45b) Input the first denoising feature into the multi-head depth-separable self-attention module, use the multi-head self-attention mechanism to explicitly model the correlation between different positions in the distance dimension, and perform depth-separable convolution and feature normalization operations to obtain the second denoising feature.

[0084] Specifically, the MDTA module utilizes a multi-head self-attention mechanism to explicitly model the correlation between different locations in the distance dimension, and efficiently extracts long-range dependency information through depthwise separable convolution and feature normalization operations, thereby making up for the shortcomings of convolutional structures in global modeling.

[0085] 45c) The second denoised feature is input into the gated feedforward network. Nonlinear feature transformation and inter-channel interaction are achieved through channel expansion, depthwise separable convolution and gate mechanism to obtain the final feature representation for target recognition and classification.

[0086] Specifically, GDFN achieves nonlinear feature transformation and inter-channel interaction through channel expansion, depthwise separable convolution, and gating mechanisms, further enhancing the discriminativeness and robustness of feature representation.

[0087] It should be noted that all the above modules are fused with the input features through residual connections. While ensuring stable feature transmission, this enhances key structural information and suppresses occlusion interference, providing a more complete and consistent feature representation for subsequent residual feature correction.

[0088] This invention, building upon multidimensional feature correction and feature fusion based on a three-branch attention mechanism, proposes an efficient denoising post-processing mechanism based on an improved Pureformer network for deep denoising. On one hand, this network fully leverages the powerful global modeling capabilities of the Transformer architecture, effectively addressing the long-distance dependency problem in HRRP signal denoising. On the other hand, by designing a multi-head dynamic transpose attention mechanism and introducing a local enhancement module, this invention effectively overcomes the technical shortcomings of the standard Transformer model, such as high computational complexity and weak ability to capture local details. While ensuring computational efficiency, it achieves accurate preservation of local subtle features of HRRP, significantly enhancing the robustness of the entire network architecture in feature correction under low signal-to-noise ratio environments.

[0089] S46. Perform forward propagation on the feature correction network through steps S41 to S45, and calculate a composite loss function including multiple losses during the forward propagation process to iteratively optimize the parameters of the feature correction network until training is completed.

[0090] Optionally, in S46, the composite loss function includes at least two of the following: identity consistency loss, triple loss, classification loss, center loss, or contrastive loss. Furthermore, other loss functions based on metric learning can be used depending on the specific application scenario, as long as they can constrain the consistency of the corrected features with the original features in terms of distribution and discriminability, without affecting the achievement of the technical objective of this invention.

[0091] As one implementation method, this embodiment preferably uses a conformance loss function that includes identity consistency loss, triple loss, and classification loss for network training. These three losses are described below.

[0092] 1. Loss of identity consistency Identity consistency loss aims to ensure that features for the same HRRP target (regardless of occlusion) remain consistent after correction. It consists of two parts: feature representation consistency and self-similarity pattern consistency. Feature representation consistency requires that the corrected features be close to the original unoccluded features in terms of feature values, and its specific formula is as follows: ; Self-similarity pattern consistency requires that the corrected features and the original unoccluded features maintain consistency in the self-similarity matrix, and the specific formula is as follows: ; in, and Let represent the self-similarity matrices of the original features and the corrected features of the unoccluded samples, respectively. This represents the self-similarity matrix of the features after occlusion correction. The formula for calculating the self-similarity matrix is: ; in, and These are obtained by expanding the features of the occluded samples after correction. and Indicates the length and width in a single channel; This indicates the inner product.

[0093] 2. Triple loss The purpose of triple loss is to further refine the features of samples with the same identity but different occlusion states, while pushing away features of different identities. Its specific formula is: ; ; ; in, This represents the corrected features of an occluded sample. With the original features of the unoccluded sample The cosine distance between them; and This represents the corrected features of an occluded sample. With the original features of the occluded sample The cosine distance between them; This indicates the interval, and is usually set to 0.1; This indicates taking a positive value, that is .

[0094] 3. Classification of Losses Classification loss can improve the intra-class compactness and inter-class separability of the corrected features. This invention chooses CosFace as the classification loss because the boundary constraints of CosFace are more suitable for learning discriminative corrected features in feature correction tasks. The specific classification loss function is defined as follows: ; in, It is the first The true label of each sample It is the first The characteristics of the sample and the first Cosine similarity between class weight vectors It is a scaling factor. Represents the cosine boundary.

[0095] In summary, the total loss is expressed as: ; This invention proposes a composite loss function system that forces the alignment of "corrected features" with "original features". In particular, the identity consistency loss not only constrains the consistency of feature values ​​(L2 norm), but also innovatively constrains the consistency of self-similarity patterns, ensuring that the corrected features retain the internal structural relationships of the original target; combined with the CosFace classification loss, it achieves the minimization of intra-class distance and the maximization of inter-class distance in low signal-to-noise ratio environments.

[0096] Through steps S41 to S45 above, the feature correction module is propagated forward. At the same time, the loss function of the generator during the forward propagation process is calculated using step S46. The parameters of the feature correction network are iteratively optimized in turn to obtain the trained feature correction network.

[0097] S5. Combine the feature extraction network obtained from the pre-training in step S3 with the feature correction network obtained from the training in step S4 to obtain a deformable target recognition model based on feature correction and Pureformer post-processing, so as to use the model to recognize and classify deformable targets.

[0098] In summary, the radar HRRP deformation target recognition method based on feature correction provided by this invention has the following advantages: 1. This invention employs an end-to-end feature correction framework, abandoning the time-consuming physical signal iterative reconstruction process and directly repairing distorted features in the feature space using a parallel correction module. This not only avoids interference from front-end recovery residuals on subsequent recognition but also achieves global optimization of feature extraction and correction through joint optimization, significantly improving the system's real-time performance and recognition accuracy.

[0099] 2. This invention introduces a feature correction approach, a Pureformer post-processing module, and a three-branch attention feature fusion mechanism. Pureformer utilizes an improved MDTA mechanism to effectively capture long-range dependencies and filter out background noise while reducing computational cost; TAFF adaptively fuses spatial and channel correction information. This enables the model to reconstruct highly discriminative features even when key scattering centers are missing or there is strong noise interference, exhibiting stronger resistance to deformation and noise.

[0100] 3. This invention, by combining the UL-ResSENet network with an efficient feature correction mechanism, significantly improves recognition performance under non-cooperative observation while maintaining low computational cost. Especially in adverse conditions such as low signal-to-noise ratio and structural deformation of the target, this invention, through the constraint of a composite loss function, can effectively pull the deformed features back to a standard distribution, greatly improving the target recognition rate compared to existing technologies and possessing extremely high practical application value.

[0101] The beneficial effects of the present invention will be verified and explained through simulation experiments below.

[0102] I. Experimental Parameter Settings This experiment used one-dimensional high-resolution range echoes of three types of aircraft to detect the recovery effect of the method of this invention. The specific parameters of the radar that recorded the high-resolution echoes of the three types of aircraft targets and the parameters of the three types of aircraft targets are shown in Table 1.

[0103] Table 1. Specific parameters of radar and three types of aircraft

[0104] The An-26's flight path was divided into 6 segments, the Citation's flight path into 7 segments, and the Yak-42's flight path into 5 segments. Only the fifth segment of the Yak-42's data had 17,950 samples; the remaining segments each had 26,000 samples. Regarding the selection of training samples, to ensure the completeness of the template library, the training samples should cover as many attitude angles as possible. Therefore, we selected the fifth and sixth segments of the An-26's flight path, the sixth and seventh segments of the Citation's flight path, and the second and fifth segments of the Yak-42's flight path as training samples, and the data from the remaining segments as test samples.

[0105] The experiments related to this invention were conducted on the NVIDIA GeForce RTX 4060 graphics processor platform manufactured by NVIDIA Corporation. The experimental program was written in the open-source Python language, using Python version 3.9, and implemented based on the PyTorch deep learning framework. The specific experimental environment configuration is shown in Table 2.

[0106] Table 2 Experimental Environment Configuration Table

[0107] The system of this invention employs the AdamW optimizer in both the feature extraction network training stage and the feature correction module training stage, with an initial learning rate set to 0.0002 and a cosine annealing scheduler, and a training period of 20 epochs. However, the difference lies in the feature correction stage, which utilizes a three-stage dynamic weight adjustment mechanism to collaboratively optimize the classification loss, identity consistency loss, and triple loss. Within the 20-epoch training period, this guides the model to smoothly transition from feature space alignment to high-precision classification and recognition, significantly improving the system's robustness in complex occlusion environments.

[0108] II. Experimental Content and Result Analysis The method of this invention is used to classify and identify high-resolution echoes from deformed aircraft.

[0109] Specifically, HRRP echo data from the Citation aircraft in its undeformed state were selected as the original sample. Please refer to [link to relevant documentation]. Figure 3 , Figure 3 This is a one-dimensional high-resolution range echo diagram of a non-deformed aircraft in a simulation experiment. To simulate a deformed target, a deformation component is superimposed within the 70-90 range cell of the HRRP. The specific waveform of this deformation component is shown below. Figure 4 As shown in the figure. Based on this, further noise is added to achieve a synthesized HRRP signal SNR of 5dB. The HRRP waveform after the above deformation component superposition and noise injection processing is shown in the figure. Figure 5 As shown. By Figure 5It is evident that the signal waveform has undergone significant distortion, especially in the 70-90 range of the distance unit, where the amplitude fluctuations are obvious and deviate significantly from the original characteristics.

[0110] Subsequently, the feature correction method proposed in this invention was applied to process the aforementioned damaged signal, and the corrected feature distribution effect is as follows: Figure 6 and Figure 7 As shown. Among them, Figure 6 To use the method of this invention to generate feature distribution maps of various targets before and after feature correction for deformed targets and original targets at a signal-to-noise ratio of 30dB, Figure 7 To illustrate the feature distribution of various targets before and after feature correction of deformed and original targets using the method of this invention at a signal-to-noise ratio of 5dB. Figure 6 and Figure 7 The left side represents the uncorrected feature distributions, while the right side represents the feature distributions after feature correction and Pureformer post-processing. Analysis Figure 6 and Figure 7 As can be seen, after processing by the method of this invention, the damaged variant features are effectively remapped and converged to the vicinity of the original target feature distribution region. This result shows that the method of this invention greatly enhances the clustering performance of features, thereby ensuring that the model can maintain stable classification ability even in low signal-to-noise ratio environments.

[0111] To further verify the effectiveness of the method of this invention, a comparative experiment was conducted with the existing "High-Resolution Range Image Recognition Method for Variant Aircraft Based on Convolutional Neural Networks". The experiment set the deformation component ratio to 30% and tested the recognition accuracy of both methods under different signal-to-noise ratios. The statistical results are shown in Table 3. Furthermore, to quantitatively evaluate the feature correction performance of the method of this invention for HRRP deformable targets, the experiment statistically analyzed and compared the feature consistency index of "deformed target / original target features" and "initial original target features" before and after feature correction. The results are shown in Table 4. The calculation formula for feature consistency is as follows.

[0112] ; in, Indicates the first The original target features of each sample Indicates the first The variant target features of each sample. The value of this indicator ranges from [-1, 1]. The closer the value is to 1, the more similar the original target features are to the variant target features.

[0113] Table 3

[0114] Analysis of the data in Table 3 shows that as the signal-to-noise ratio (SNR) decreases, the recognition accuracy of the method in this invention exhibits a relatively gradual downward trend, demonstrating excellent noise resistance and robustness. In contrast, the convolutional neural network-based comparison method, which focuses on extracting common features between the deformed target and the original target, has limited feature extraction capabilities under low SNR conditions, leading to a significant decrease in its recognition performance. Furthermore, if the deformed region is a crucial area for feature recognition, the common features extracted by the convolutional neural network may discard features from that region, resulting in a decline in recognition results.

[0115] Table 4

[0116] As shown in Table 4, after correction using the method of this invention, the consistency between the features of the deformed target and the original target features is significantly enhanced. Meanwhile, for the original target that has not undergone deformation, after the same feature correction process, its feature information is highly preserved, with no obvious feature loss. This further confirms that the method of this invention can repair damaged features while maintaining the integrity of the original features, demonstrating good algorithmic robustness.

[0117] This verifies the effectiveness of the present invention.

[0118] 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 radar HRRP deformation target recognition method based on feature correction, characterized in that, include: S1. Acquire radar one-dimensional HRRP echo data of the target to be identified, and perform preprocessing operation on the radar one-dimensional HRRP echo data, and use the preprocessed data as the original HRRP target data. S2. Add block sparse signals of different sparsity as deformation components to the original HRRP target data, and further add Gaussian white noise to obtain HRRP deformation target data. The HRRP deformed target data, the HRRP original target data, and the corresponding target category labels are associated to construct a pairwise sample set for subsequent network training; S3. Construct a lightweight feature extraction network, and pre-train the feature extraction network using the original target data of HRRP to obtain a pre-trained feature extraction network. S4. Construct a feature correction network including a multi-dimensional parallel feature correction module, a feature fusion module based on a multi-branch attention mechanism, and a denoising post-processing module based on an improved Pureformer network, and cascade it with the feature extraction network; fix the parameters of the feature extraction network, use the pairwise sample set as input data, and optimize the parameters of the feature correction network through joint training to obtain a trained feature correction network. S5. Combine the feature extraction network obtained from the pre-training in step S3 with the feature correction network obtained from the training in step S4 to obtain a deformable target recognition model based on feature correction and Pureformer post-processing, so as to use the model to recognize and classify deformable targets.

2. The radar HRRP deformation target recognition method based on feature correction according to claim 1, characterized in that, In S3, the feature extraction network adopts an ultra-lightweight residual compression-excitation network structure, or a network structure based on ResNet or DenseNet, or a network structure based on channel attention or coordinate attention.

3. The radar HRRP deformation target recognition method based on feature correction according to claim 1, characterized in that, In S3, the feature extraction network adopts an ultra-lightweight residual compression-activation network structure, including an input layer, an enhancement residual layer, and an output layer; then, the feature extraction network is pre-trained using the original HRRP target data, including: S31. Input the original HRRP target data into the input layer of the feature extraction network, and perform preliminary feature extraction and normalization processing including batch normalization, nonlinear activation and random deactivation regularization in sequence to obtain the first feature; S32. The first feature is input into the enhancement residual layer of the feature extraction network, and multi-layer convolution processing, feature enhancement processing based on channel attention weighting mechanism and residual connection fusion processing are performed in sequence to obtain the second feature; S33. Input the second feature into the output layer of the feature extraction network to perform feature aggregation and mapping, and generate a primary feature representation; S34. Backpropagate the feature extraction network using the cross-entropy loss function to optimize and iterate the parameters of the feature extraction network until pre-training is complete.

4. The radar HRRP deformation target recognition method based on feature correction according to claim 1, characterized in that, In S4, the multi-dimensional parallel feature correction module adopts a parallel spatial correction and channel correction structure, or a multi-scale self-attention mechanism and a non-local attention mechanism, or a feature relationship modeling method based on graph neural networks to perform correlation modeling and correction on the feature representation output by the feature extraction network to obtain multi-dimensional corrected features. The feature fusion module based on the multi-branch attention mechanism uses a three-branch attention fusion network structure to fuse multi-dimensional corrective features to obtain fused features; The denoising post-processing module based on the improved Pureformer network includes a feature enhancement module, a multi-head deep separable self-attention module, and a gated feedforward network, which are used to perform contextual relationship modeling and global dependency enhancement on the fused features to obtain the final feature representation for target recognition and classification.

5. The radar HRRP deformation target recognition method based on feature correction according to claim 4, characterized in that, In S4, the parameters of the feature correction network are optimized through joint training, using the paired sample set as input data, including: S41. The HRRP deformation target data in the paired sample sets are respectively... Compared with HRRP original target data After feature extraction by the aforementioned feature extraction network, the corresponding deformation feature representation is obtained. Compared with the original feature representation ; S42, respectively and Spatial dimension correction is performed using a spatial attention network to generate positional importance weights along the distance dimension, resulting in spatially corrected features. and The spatial attention network comprises two layers of one-dimensional convolution, a nonlinear activation function, and a normalization function. S43, respectively and Channel attention networks are used to correct the channel dimension, modeling the importance of each feature channel along the channel dimension, thus obtaining the channel-corrected features. and ; S44. Simultaneously process the two feature paths. and Adaptive weighted fusion is performed using a three-branch attention network to obtain the fused deformation HRRP features. Compared with the original HRRP features ; S45. Analyze the deformation HRRP features respectively. Compared with the original HRRP features After feature denoising based on an improved Pureformer network, the final feature representation for target recognition and classification is obtained. S46. Perform forward propagation on the feature correction network through steps S41 to S45, and calculate a composite loss function including multiple losses during the forward propagation process to iteratively optimize the parameters of the feature correction network until training is completed.

6. The radar HRRP deformation target recognition method based on feature correction according to claim 5, characterized in that, S42 includes: 42a) respectively and In the input spatial attention network, the number of channels of the input features is compressed by the first layer of one-dimensional convolution, and local spatial structure information is extracted while keeping the distance dimension unchanged to obtain the first spatial features; 42b) The first spatial features are processed using the ReLU nonlinear activation function to obtain the second spatial features; 42c) The second spatial feature is mapped to a single-channel spatial response through a second layer of one-dimensional convolution to obtain the third spatial feature; 42d) The spatial response of the third spatial feature is normalized to the 0-1 interval using the Sigmoid function, thus obtaining the spatially corrected feature. and .

7. The radar HRRP deformation target recognition method based on feature correction according to claim 5, characterized in that, S43 includes: 43a) respectively and In the input channel attention network, the features of each channel are globally aggregated through adaptive average pooling to obtain the features of the first channel. 43b) One-dimensional convolution is used to compress the channel dimension of the first channel feature, and then combined with the ReLU activation function to obtain the second channel feature; 43c) The second channel features are mapped back to the original number of channels through a second layer of one-dimensional convolution to generate the response weights of each channel, thus obtaining the third channel features; 43d) The weights of the third channel features are normalized to the 0-1 interval using the Sigmoid function, thus obtaining the channel-corrected features. and .

8. The radar HRRP deformation target recognition method based on feature correction according to claim 5, characterized in that, S44 includes: 44a) Two-way features and The data are input into a multi-scale convolutional attention network, and a fusion model is performed, including global average pooling and local one-dimensional convolution, to obtain two first fusion features. 44b) The two first fusion features are processed by channel compression convolution, batch normalization and ReLU activation respectively, and then channel recovery convolution is used to generate an attention response with the same number of channels as the original, thus obtaining two second fusion features; 44c) The two second fusion features are weighted and fused using the Sigmoid function to obtain the fused deformation HRRP features. Compared with the original HRRP features .

9. The radar HRRP deformation target recognition method based on feature correction according to claim 5, characterized in that, In S45, the improved Pureformer network includes a feature enhancement module, a multi-head deep separable self-attention module, and a gated feedforward network; therefore, S45 includes: 45a) Deformation HRRP features respectively Compared with the original HRRP features In the input feature enhancement module, the input features are modeled at multiple scales through a parallel one-dimensional convolutional structure, and the features at different scales are fused and compressed to obtain the first denoised feature; 45b) Input the first denoising feature into the multi-head depth-separable self-attention module, use the multi-head self-attention mechanism to explicitly model the correlation between different positions in the distance dimension, and perform depth-separable convolution and feature normalization operations to obtain the second denoising feature; 45c) The second denoised feature is input into the gated feedforward network, and nonlinear feature transformation and inter-channel interaction are achieved through channel expansion, depthwise separable convolution and gate mechanism to obtain the final feature representation for target recognition and classification.

10. The radar HRRP deformation target recognition method based on feature correction according to claim 5, characterized in that, In S46, the composite loss function includes at least two of the following: identity consistency loss, triple loss, classification loss, center loss, or contrast loss.

Images