Target classification method based on HRRP sequence and related equipment

Abstract

The invention discloses a target classification method based on an HRRP sequence and related equipment, and relates to the technical field of target classification, and the method comprises the steps: constructing a recognition model based on an attention mechanism and a bidirectional long short-term memory network, the recognition model comprises an angle fusion module, a local feature extraction module, a time sequence feature extraction module, a feature fusion module and a classifier; training the recognition model by using an HRRP sequence sample with a label; and inputting a to-be-classified HRRP sequence into the trained recognition model, and outputting a target classification result corresponding to the to-be-classified HRRP sequence. According to the method, the space and time information in the high-resolution range profile sequence can be utilized more fully and cooperatively, so that the accuracy and stability of target classification are improved.

Description

Technical Field

[0001] This invention relates to the field of target classification technology, and in particular to a target classification method and related equipment based on HRRP sequences. Background Technology

[0002] Automatic target recognition using high-resolution range profiles (HRRP) is a crucial technology. In practical applications, high-resolution range profile sequences acquired by radar not only contain local scattering features determined by the target's own structure but also contain temporal evolution patterns caused by target motion or changes in radar observation perspective. Therefore, an effective recognition model needs to possess three key capabilities: first, accurately extracting local structural features robust to attitude changes from each frame of range profile; second, modeling the long-range dependencies of the entire sequence over time to capture dynamic evolution patterns; and third, distinguishing the differences in importance of different regions in the feature map to the final recognition, thereby suppressing noise and focusing on discriminative regions. How to synergistically improve these three capabilities is a significant challenge currently facing the field of high-resolution range profile sequence recognition.

[0003] To address these challenges, researchers have conducted extensive research. Initially, convolutional neural network (CNN)-based methods were introduced, primarily leveraging their powerful local feature extraction capabilities to capture structural patterns in high-resolution range images. To compensate for the limitations of CNNs in modeling temporal relationships, recurrent neural networks (RNNs) were introduced to this task, capturing temporal dependencies by processing sequential data segments. To further enhance the modeling ability for long-range dependencies, more advanced sequence models such as Long Short-Term Memory (LSTM) networks and gated recurrent units (ROUs) were employed, whose gating mechanisms better control information flow. Furthermore, combining CNNs and RNNs to construct hybrid models has become a mainstream approach. This architecture attempts to integrate the former's local feature extraction capabilities with the latter's temporal modeling abilities, aiming to jointly improve recognition performance.

[0004] However, the aforementioned existing technical solutions still have several drawbacks. First, the simple combination of convolutional neural networks and recurrent neural networks has limited ability to model the global context of a sequence, especially struggling to effectively handle complex dependencies in long sequences. Second, standard long short-term memory networks or gated recurrent units are typically unidirectional, learning only from historical information and unable to utilize future contextual information, which limits the completeness of temporal feature representation. More critically, existing methods typically assume that all extracted local features are equally important during feature extraction, lacking an adaptive mechanism to evaluate and strengthen key discriminative regions while suppressing redundant or noisy regions, which may lead the model to focus on non-critical information. Finally, most methods do not explicitly utilize pose angle information acquired simultaneously with high-resolution range images, which is crucial for explaining feature distortions caused by changes in target pose.

[0005] In summary, existing high-resolution range image sequence recognition methods have limitations in global temporal dependency modeling, bidirectional context utilization, adaptive feature weighting, and pose information fusion. To systematically address these issues, a new model architecture is needed that can integrate angle information, achieve adaptive local feature enhancement, and collaboratively perform bidirectional temporal feature extraction (both global and local), ultimately making robust decisions through flexible feature fusion strategies. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, specifically by providing a target classification method and related equipment based on HRRP sequences, as detailed below: 1) In a first aspect, the present invention provides a target classification method based on HRRP sequences, the specific technical solution of which is as follows: A recognition model based on an attention mechanism and a bidirectional long short-term memory network is constructed. The recognition model includes an angle fusion module, a local feature extraction module, a temporal feature extraction module, a feature fusion module, and a classifier. The angle fusion module is used to synthesize the original elevation angle and original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, and to obtain enhanced features based on the synthetic attitude angle and the initial features obtained from the HRRP sequence. The local feature extraction module is used to obtain weighted local features based on the enhanced features. The temporal feature extraction module is used to obtain temporal features that integrate local temporal dynamics based on the enhanced features. The feature fusion module is used to fuse the weighted local features and the temporal features that integrate local temporal dynamics to obtain fused features or fused decision scores. The classifier is used to classify the target based on the fused features or fused decision scores. The recognition model is trained using labeled HRRP sequence samples. The HRRP sequence to be classified is input into the trained recognition model, and the target classification result corresponding to the HRRP sequence to be classified is output.

[0007] The beneficial effects of the target classification method based on HRRP sequences provided by this invention are as follows: The angle fusion module explicitly synthesizes and fuses attitude angles, enhancing the model's adaptability to target attitude changes and mitigating feature instability caused by varying observation angles in high-resolution range image sequences. The local feature extraction module employs an embedded attention mechanism to adaptively weight enhanced features, generating weighted local features that effectively focus on highly discriminative scattering center regions and suppress noise interference. The temporal feature extraction module integrates the global modeling capabilities of the Transformer encoder with the bidirectional local temporal capture capabilities of the bidirectional Long Short-Term Memory network, outputting temporal features that integrate local temporal dynamics to more comprehensively represent the sequence's evolution. The feature fusion module provides multiple fusion strategies, allowing adaptive integration at the level of fused features or fused decision scores based on actual needs, improving the flexibility of feature utilization and the robustness of decision-making. These improvements work together to enable the recognition model to more fully and synergistically utilize the spatial and temporal information in high-resolution range image sequences, thereby enhancing the accuracy and stability of target classification.

[0008] 2) In a second aspect, the present invention also provides a target classification system based on HRRP sequences, the specific technical solution of which is as follows: The system comprises a model building module, a model training module, and a model application module. The model building module constructs a recognition model based on an attention mechanism and a bidirectional long short-term memory network. This model includes an angle fusion module, a local feature extraction module, a temporal feature extraction module, a feature fusion module, and a classifier. The angle fusion module synthesizes the original elevation and azimuth angles of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, and obtains enhanced features based on the synthetic attitude angle and initial features obtained from the HRRP sequence. The local feature extraction module obtains weighted local features based on the enhanced features. The temporal feature extraction module obtains temporal features integrating local temporal dynamics based on the enhanced features. The feature fusion module fuses the weighted local features and the temporal features integrating local temporal dynamics to obtain fused features or fused decision scores. The classifier classifies the target based on the fused features or fused decision scores. The model training module trains the recognition model using labeled HRRP sequence samples. The model application module inputs the HRRP sequence to be classified into the trained recognition model and outputs the target classification result corresponding to the HRRP sequence.

[0009] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to enable the electronic device to implement any of the above-mentioned target classification methods based on HRRP sequences.

[0010] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned target classification methods based on HRRP sequences.

[0011] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a target classification method based on HRRP sequences according to an embodiment of the present invention. Figure 2 A schematic diagram of the network structure for identifying the model; Figure 3 This is a schematic diagram of the network structure of the dual attention module; Figure 4 This is a schematic diagram of the structure of a target classification system based on HRRP sequences according to an embodiment of the present invention. Detailed Implementation

[0013] like Figure 1 As shown, an embodiment of the present invention provides a target classification method based on HRRP sequences, comprising the following steps: S1. Construct a recognition model based on attention mechanism and bidirectional long short-term memory network. The recognition model includes an angle fusion module, a local feature extraction module, a temporal feature extraction module, a feature fusion module, and a classifier. The angle fusion module is used to synthesize the original elevation and azimuth angles of the target relative to the radar line of sight in the HRRP sequence into synthetic attitude angles. Based on these synthetic attitude angles and initial features obtained from the HRRP sequence, enhanced features are derived. Specifically, the original elevation and azimuth angles of the target relative to the radar line of sight in the HRRP sequence are synthesized into synthetic attitude angles. These synthetic attitude angles are then encoded to generate angle encoding vectors, and these angle encoding vectors are fused with the initial features obtained from the HRRP sequence to obtain enhanced features. By synthesizing the original elevation and azimuth angles into physically meaningful synthetic attitude angles and converting them into angle encoding vectors for fusion with the initial features, the enhanced features directly imply the spatial geometric relationships of the target. This process provides the model with explicit modeling of target attitude changes, mitigating the differences in feature distribution of high-resolution range image sequences caused by different observation angles, thereby improving the adaptability and stability of the recognition system to target attitude changes.

[0014] The angle fusion module consists of three sequentially connected network layers: a mathematical synthesis layer, a fixed encoding layer, and a learnable linear projection fusion layer. The specific implementation process for obtaining enhanced features is as follows: 1) The angle fusion module receives two inputs: the raw elevation angle and raw azimuth angle of the target relative to the radar line of sight in the HRRP sequence. For each time step index in the HRRP sequence... This layer reads the corresponding raw pitch angle. and original azimuth The computational operation of this layer is to execute a predefined geometric synthesis formula to combine the two angles into a single composite attitude angle. The calculation formula is: ,in, Represents the time step index of the HRRP sequence; Indicates at time step At that time, the original elevation angle of the target in the HRRP sequence relative to the radar line of sight; Indicates at time step At that time, the original azimuth angle of the target in the HRRP sequence relative to the radar line of sight; This indicates the calculated result at time step The resultant attitude angle at time; function It is the four-quadrant arctangent function, used to ensure the calculated angle Value range coverage is complete Degree range; function and These represent the sine and cosine trigonometric functions, respectively. The mathematical synthesis layer does not contain learnable parameters; its output is a synthesized attitude angle sequence with the same time length as the HRRP sequence. ,in This represents the total number of time steps in the sequence. The structural characteristic of this layer is that it performs deterministic mathematical transformations, mapping two input angles to an attitude angle with a more comprehensive physical meaning.

[0015] 2) The synthesized attitude angles are encoded to generate angle encoding vectors. This step is completed by a fixed encoding layer. This layer encodes the synthesized attitude angle sequence output by the mathematical synthesis layer. As input, it is transformed into a fixed-dimensional high-dimensional vector representation, namely the angle-encoded vector. This layer's design borrows from the positional encoding concept in the Transformer architecture, and its structure consists of a set of pre-defined, non-trainable sine and cosine functions. First, an encoding dimension needs to be pre-defined for this layer. This dimension determines the length of the angular encoding vector. For each time step in the sequence... Synthetic attitude angle This layer generates a trigonometric function of different frequencies by computing them in parallel. A 3D vector. The encoding formula is: For dimension indexes , , ,in, This indicates the angle encoding vector generated by the fixed coding layer at time step. , No. The value of the dimension; It is a time step The composite attitude angle; It is a pre-defined encoding dimension; It is a dimension index variable; It is a constant basis used to adjust the frequency distribution. The fixed coding layer has no trainable parameters; its structure is essentially a set of fixed nonlinear transformation functions. Through this formula, each synthesized pose angle... Converted into a length of The angle encoding vector. Finally, for the entire length of The sequence, with a fixed coding layer outputting a shape of An angular encoding vector matrix, where each row corresponds to the angular encoding of a time step.

[0016] 3) The angle-coded vector is fused with the initial features obtained from the HRRP sequence to obtain enhanced features, which is accomplished by a learnable linear projection fusion layer. This layer receives two inputs: the first input is the output of the fixed coding layer, with a shape of... The first input is the angle-encoded vector matrix; the second input is the initial features obtained from the HRRP sequence, which are extracted by a separate preprocessing layer (e.g., a one-dimensional convolutional layer) at the front end of the model, and their shape is assumed to be... ,in It is the number of time steps. This refers to the number of feature channels. The core structure of the linear projection fusion layer is a fully connected layer used to solve the dimension alignment problem. The weight matrix of this fully connected layer... The shape is Bias vector The shape is These are all trainable parameters. For the time step... This layer first performs a linear projection on the angle encoding vector, the operation of which is defined as: ,in, Indicates at time step An angular encoding vector, which is of shape... The row vector; Represents the learnable weight matrix; Represents a learnable bias vector; This represents the pose feature vector, aligned with the initial feature dimensions after linear projection, and its shape is... .

[0017] The linear projection fusion layer performs this operation in parallel across all time steps, resulting in a shape of The pose feature matrix. The key to this layer structure is the introduction of learnable parameters. and This allows the model to automatically learn how to map the high-dimensional encoding of the attitude angles to the semantic space most relevant to the radar features. Finally, this layer performs a feature fusion operation, adding the projected attitude feature matrix element-wise to the initial feature matrix obtained from the HRRP sequence. For each time step... The addition operation is defined as follows: ,in, Indicates at time step Initial features obtained from HRRP sequences; Indicates at time step The enhanced features obtained after fusion. This addition operation is part of the linear projection fusion layer structure, which directly injects the learned attitude information into the feature representation of the radar echo. At this point, the angle fusion module has completed the forward computation of all network layers and outputs the enhanced feature tensor. This is for use by subsequent local feature extraction modules and temporal feature extraction modules.

[0018] The HRRP sequence is a complete representation of the high-resolution range image sequence. It refers to a series of high-resolution one-dimensional range images arranged in chronological order, obtained through continuous observation of moving targets using broadband radar. Each high-resolution range image reflects the intensity distribution of the target's scattering center along the radar's line-of-sight at a specific moment. This sequence contains dynamic information about the target's structure and attitude changes over time.

[0019] In the HRRP sequence, targets refer to entities observed by radar and attempted to be classified using an identification model. Typical targets in radar target identification include ships and aircraft. The physical structure, size, and material properties of these targets determine the distribution pattern of their scattering centers in high-resolution range images, which serves as the basis for identification and classification.

[0020] In the HRRP sequence, the raw elevation angle of the target relative to the radar line of sight is an angular parameter describing the spatial geometric relationship between the target and the radar. It is defined as the angle between the radar line of sight direction and a certain reference horizontal plane. This angular information is usually provided by the radar's measurement system or external sensors along with the raw echo data, and it directly affects the projection relationship of the target's scattering center in the range dimension.

[0021] Among them, the raw azimuth angle of the target relative to the radar line of sight in the HRRP sequence is another angular parameter describing the spatial geometric relationship between the target and the radar. It is defined as the angle between the projection of the radar line of sight direction onto a reference horizontal plane and a certain reference direction. Like the raw elevation angle, this information is also provided by the radar system and together they determine the complete field of view of the radar observing the target.

[0022] The initial features obtained from the HRRP sequence are an intermediate data representation. These refer to the features obtained after the original HRRP sequence data has undergone a preliminary transformation through a basic feature extraction layer before being input into the angle fusion module. This basic feature extraction layer can be a one-dimensional convolutional layer or a fully connected layer. Its function is to transform the original, possibly preprocessed, HRRP numerical sequence into a feature representation more suitable for further processing by the neural network. This initial feature serves as the basic carrier for the fusion of angle information.

[0023] The mathematical synthesis layer is a parameter-free computational layer within the angle fusion module. Its structure consists of predefined mathematical formulas and contains no trainable weights or biases. Its function is to synthesize two input angle parameters—the original pitch angle and the original azimuth angle—into a new angle parameter, the synthesized attitude angle, through a defined geometric transformation formula. This layer performs a physically meaningful feature construction.

[0024] The fixed-encoding layer is a parameter-free mapping layer within the angle fusion module. Its structure consists of a set of sine and cosine functions with preset frequencies arranged in parallel. Its function is to map the input synthetic attitude angle scalar values ​​into a fixed-length high-dimensional vector. This mapping method encodes the periodic variation patterns inherent in the angle values ​​into a vector representation, providing the model with structured prior knowledge. Furthermore, all mapping rules for this layer are determined before model training begins and will not change during training.

[0025] The learnable linear projection fusion layer is a parameterized network layer in the angle fusion module. Essentially, this layer is a fully connected layer or linear transformation layer, containing a trainable weight matrix and bias vector. It has two functions: first, to project the high-dimensional vector output from the fixed encoding layer to a dimension matching the initial radar features; second, to fuse the projected attitude information vector with the initial radar features element-wise. This layer learns the optimal projection and fusion methods through training.

[0026] The local feature extraction module is used to obtain weighted local features based on the enhanced features. Specifically, the local feature extraction module includes a one-dimensional convolutional residual network and a dual attention module. Specifically, the local feature extraction module extracts spatial features from the enhanced features through the one-dimensional convolutional residual network, and then applies channel and spatial attention weights to the spatial features using the dual attention module, outputting the weighted local features. It also extracts deeper spatial features from the enhanced features through the one-dimensional convolutional residual network, and then applies sequential channel and spatial dimension weights to the spatial features using the dual attention module. This design allows the model to adaptively improve the feature response to key local regions in the recognition task while suppressing the activation intensity of redundant or interfering parts in the feature map. The final weighted local features have stronger discriminative power, improving the model's ability to capture local structural patterns in high-resolution range image sequences and its robustness.

[0027] The data processing procedure of the local feature extraction module is as follows: 1) Spatial features are extracted from the enhanced features using a one-dimensional convolutional residual network. This extraction is accomplished by the one-dimensional convolutional residual network. The one-dimensional convolutional residual network is a depth-feedforward structure containing multiple cascaded residual blocks. Its input is the enhanced features output by the angle fusion module, denoted as... Its shape is ,in, It is the number of time steps in the sequence. This refers to the number of input feature channels. The first layer of a one-dimensional convolutional residual network is a standalone one-dimensional convolutional layer used for initial feature transformation. This convolutional layer uses... The size is A one-dimensional convolution kernel with a stride of 1D For input Perform a convolution operation, add a bias, and then pass it through a non-linear activation function. The formula can be expressed as: ,in, This represents a one-dimensional convolution operation. This represents the weight matrix of the convolutional layer. This represents the bias vector. This represents the feature map after the initial transformation.

[0028] Subsequently, feature maps Sent to A series of consecutive one-dimensional residual convolutional blocks. The structure of each residual convolutional block is standardized. (The text then repeats the first two sentences, which are likely errors in the original Chinese.) Taking a block as an example, it contains two one-dimensional convolutional sub-layers, each followed by a batch normalization operation and... Activation function. The block calculation process is as follows: input features First, after passing through the first convolutional sub-layer, we obtain... ;then After the second convolutional sub-layer, we get .here This indicates a batch normalization operation. and These are the weights and biases of the two convolutional sub-layers, respectively.

[0029] Residual connections are key to the structure of this network layer. Each residual convolutional block converts the input of that block into a single layer. With the output of the second convolutional sub-layer Perform element-wise addition. This addition operation requires... and The dimensions must be the same. If the dimensions are different, the network will pass them through a shortcut connection layer containing one-dimensional convolution operations. Perform a projection transformation to match The dimension. Ultimately, the first The output of each residual block is .pass With such blocks, the network can extract deep spatial features. The final output of the one-dimensional convolutional residual network is denoted as... Its shape is This is the original spatial feature extracted from the enhanced features.

[0030] 2) Spatial features are weighted by channel attention and spatial attention using a dual attention module, and the weighted local features are output. The input of the dual attention module is the original spatial features. The dual attention module's internal structure consists of two sequentially connected sub-modules: a channel attention sub-module and a spatial attention sub-module. The module's output is a double-weighted feature, i.e., a weighted local feature. .

[0031] The channel attention submodule first processes the input features In the spatial dimension (i.e., the time step dimension) Information is aggregated on the surface to generate two different spatial context descriptors: one is a global average pooling descriptor. The other is the global max-pooling descriptor. The calculation method is as follows: ,in, The feature map represents the first time. All channel values ​​at each time step The operation is to retrieve the maximum value channel by channel. These two descriptors... and All shapes are Next, they are fed into a shared multilayer perceptron. This multilayer perceptron typically contains a hidden layer, using a dimensionality reduction ratio. To reduce the number of parameters, and use Activation function. The computation of a shared multilayer perceptron can be represented as: , Indicates input descriptor or , and These are the weights and biases corresponding to the two layers of the multilayer perceptron. express Activation function express Activation function. After calculating the activation function for each of the two descriptors, the resulting output vectors are summed element by element, and then... The function generates the final channel attention weight vector. : , The shape is Finally, the channel attention weights are... Compared with the original input features Perform channel-by-channel multiplication (broadcast mechanism) to obtain the channel-weighted features. : ,in, It is a time step index. It is a channel index.

[0032] The spatial attention submodule uses channel-weighted features As input, this submodule first aggregates information along the channel dimension, generating two channel context descriptors: one is a descriptor obtained through average pooling. The other is a descriptor obtained through max pooling. The calculation method is as follows: These two operations generated two shapes, both of which are The descriptor. Next, will and By splicing along the channel dimension, a shape is obtained. The feature map. Then, using a feature map of size... A one-dimensional convolutional layer performs a convolution operation on the stitched feature map, aiming to learn the importance relationships between spatial locations. This convolution operation is followed by a... Activation function to generate spatial attention weight matrix Its shape is The formula is expressed as: .

[0033] This indicates a splicing operation at the channel level. Indicates the kernel size as One-dimensional convolution, yes The function. Finally, the spatial attention weights. Channel weighted features Multiplication (broadcasting mechanism) is performed per spatial location (i.e. per time step) to obtain the final output after double weighting, which is the weighted local feature. : At this point, the local feature extraction module has completed the entire process from extraction to weighting, and outputs... It will be passed to the subsequent feature fusion module.

[0034] The temporal feature extraction module is used to: obtain temporal features that integrate local temporal dynamics based on enhanced features. Specifically, the temporal feature extraction module includes a temporal feature mapping layer, a Transformer encoder, and a bidirectional long short-term memory network. The module is specifically used to: extract initial temporal features from the enhanced features through the temporal feature mapping layer; model global dependencies on the initial temporal features through the Transformer encoder, outputting global enhanced features; and capture the forward and backward local temporal dynamics in the global enhanced features through the bidirectional long short-term memory network, obtaining temporal features that integrate local temporal dynamics. This hybrid architecture allows the extracted temporal features to simultaneously contain global contextual information of the long sequence and local evolution patterns of adjacent time points, thereby more comprehensively modeling the temporal dimension variation of high-resolution distance image sequences and improving the model's ability to handle time shift sensitivity and long-range dependencies.

[0035] The data processing procedure of the time series feature extraction module is as follows: 1) Initial temporal features are extracted from the enhanced features through a temporal feature mapping layer. This layer is a parameterized feature transformation layer, essentially an independent one-dimensional convolutional layer. The input to this layer is the enhanced features output by the angle fusion module, denoted as... Its shape is ,in, It is the time step of the HRRP sequence. It represents the number of channels in the input feature, corresponding to the dimension of the augmented feature.

[0036] The temporal feature mapping layer internally contains a set of trainable one-dimensional convolutional kernels. Assume this layer uses... The size is A one-dimensional convolution kernel with a stride of 1D Perform convolution operations and use the same padding method to preserve the time step dimension. The length remains unchanged. The computational operations for this layer are defined as follows: ,in, This represents a one-dimensional convolution operation; The weight parameters of the convolution kernel in the temporal feature mapping layer are represented by the following shape: ; This represents the corresponding bias parameter, and its shape is... ;function Represents the activation function of the Gaussian error linear unit; This represents the output feature obtained after this transformation, i.e., the initial temporal feature, and its shape is... .

[0037] The purpose of the temporal feature mapping layer is to perform a basic, non-linear spatial-to-temporal feature projection on the input augmented features. It maps the augmented features from the local feature extraction path to a new feature space, the dimension of which is... It is configured to match the input dimension expected by the subsequent Transformer encoder, preparing for subsequent complex global temporal modeling.

[0038] 2) The global dependencies of the initial temporal features are modeled using a Transformer encoder, outputting globally enhanced features. This step is performed by the Transformer encoder. The Transformer encoder is a... A deep network structure consisting of stacked identical encoder layers. Each encoder layer contains two main sublayers: a multi-head self-attention sublayer and a positional feedforward neural network sublayer. Residual connections and layer normalization operations are applied around each sublayer.

[0039] The input to the Transformer encoder is the initial temporal features. For the first The encoder layer is calculated as follows: First, the input features... The process then proceeds to the multi-head self-attention sublayer. This sublayer linearly projects the input features onto three spaces: query, key, and value, and partitions them into... Head. For the first The size is calculated as follows: in, , , They are the first The corresponding learnable query, key, and value projection matrix for each size, whose dimensions project the features onto the depth. and ,generally ; It is a scaling factor. The outputs of all heads are concatenated along the feature dimension and projected through a linear output matrix. The output of the multi-head self-attention sublayer is obtained by fusion. : .

[0040] Next, residual connections and layer normalization are applied: ,in The representation layer normalization operation is then performed. Entering the positional feedforward neural network sublayer. This sublayer consists of two linear transformations and an activation function, defined as follows: ,in, , and , These are the learnable parameters of the feedforward network; the dimensions of the intermediate layers are typically expanded to... Multiples of . Applying residual connectivity and layer normalization again, we obtain the . The final output of each encoder layer: .go through After processing by such an encoder layer, the final output of the Transformer encoder is denoted as... Its shape remains .this Known as global augmentation features, it encodes the global dependencies between any two time steps in the HRRP sequence.

[0041] 3) By capturing the forward and backward local temporal dynamics in the global enhancement features through a bidirectional long short-term memory network, temporal features integrating local temporal dynamics are obtained. The structure of the bidirectional long short-term memory network consists of two independent long short-term memory network units deployed in parallel: one is called the forward long short-term memory network, which processes the sequence in chronological order; the other is called the backward long short-term memory network, which processes the sequence in reverse chronological order.

[0042] The input to a bidirectional long short-term memory network is a global augmentation feature. For a basic long short-term memory network unit, at each time step It depends on the current input (This is) The (line), hidden state of the previous time step and cell state Through the input gate Forgotten Gate Output gate and candidate cell status To update the status. The calculation involves the following formula: in, express Activation function; Represents the hyperbolic tangent activation function; This represents element-wise multiplication; and These are the input weight matrix and cyclic weight matrix for different gates; It is the corresponding bias vector.

[0043] In bidirectional long short-term memory networks, the forward long short-term memory network... arrive Sequential processing sequence This generates a sequence of forward hidden states. Meanwhile, the backward long short-term memory network from arrive Reverse the same sequence to generate a set of backward hidden state sequences. For each time step Hide the forward state and backward hidden state Concatenate along the feature dimension: , This represents the concatenation operation. After performing this operation on all time steps, the complete output sequence is obtained, denoted as... Its shape is ,in It is the hidden layer dimension of the Long Short-Term Memory network in each direction. It integrates the temporal features of forward and backward local temporal dynamics. It includes global context information modeled by Transformer, and captures the local dependencies between each time point in the sequence and its immediate preceding and following context through a bidirectional loop structure.

[0044] Local temporal dynamics refer to the short-term dependencies and change patterns between adjacent or nearby time points in time series data. In high-resolution range image sequences, local temporal dynamics manifest as the evolution of the target scattering center distribution across several consecutive radar pulse echoes, such as the gradual change, appearance, or disappearance of scattering point intensity. This dynamic reflects the continuity of the target's local structural motion, distinguishing it from global trends or periodic patterns determined by the entire long sequence. Capturing local temporal dynamics is crucial for understanding the detailed evolution of the sequence.

[0045] The feature fusion module is used to fuse weighted local features and temporal features that integrate local temporal dynamics to obtain fused features or fused decision scores. Specifically, it employs at least one of the following fusion methods: feature-level adaptive weighted fusion, channel-level importance-weighted fusion, and decision-level fusion. It provides three selectable fusion strategies—feature-level, channel-level, and decision-level—allowing the model to adaptively select a fusion method from fine-grained to coarse-grained based on task requirements and data characteristics. This flexibility supports both pixel-level fine-tuning of weighted local features and temporal features integrating local temporal dynamics, as well as efficient channel-level overall weighting or decision-level score integration. The diverse fusion paths enhance the sufficiency of feature utilization and the adaptability of decision-making, improving the performance and stability of the recognition system in different scenarios.

[0046] The feature fusion module is a flexible network component responsible for effectively combining the weighted local features output by the local feature extraction module with the temporal features output by the temporal feature extraction module, which integrates local temporal dynamics. This module provides three optional fusion implementation methods with different network layer structures: feature-level adaptive weighted fusion, channel-level importance weighted fusion, and decision-level fusion. These methods differ in fusion granularity, parameter complexity, and computational efficiency, but all are based on learnable weight adjustments. The specific implementation process is as follows: 1) A fine-grained, parameterized fusion network layer is constructed using a feature-level adaptive weighted fusion approach. The input to this network layer consists of two feature tensors: one is the weighted local features from the local feature extraction module, denoted as... Its shape is The other is the temporal feature from the temporal feature extraction module, which integrates local temporal dynamics, denoted as... Its shape is also .here Indicates the number of time steps in the sequence. This represents the number of feature channels, and requires that the dimensions of the two input features be the same.

[0047] At the core of this fusion network layer is a learnable weight tensor. . Defined as a feature related to the input and Tensors with exactly the same shape, i.e. During model initialization, the weight tensor... All elements are set to the same scalar value. This tensor As parameters of the network layers, they will be optimized and updated during training using the backpropagation algorithm.

[0048] The forward computation operation of this network layer is an element-wise summation of Hadamard products. Specifically, the weight tensors are... Weighted local features Perform element-wise multiplication; simultaneously, calculate complementary weights. And combine it with temporal features that integrate local temporal dynamics. Perform element-wise multiplication. Finally, add the two products together to obtain the fused feature. This process can be expressed by the following formula: ,in, This represents element-wise multiplication; Represents a learnable weight tensor that is isomorphic to the feature map; Indicates the weighted local features; This indicates the integration of temporal features with local temporal dynamics; This represents the characteristics of the output after fusion.

[0049] This network layer structure allows the model to adaptively assign an independent fusion weight to each specific location (time step) and each feature channel in the feature map. For example, for regions with strong spatial discriminativeness, the corresponding... Element values ​​may approach This preserves more information about the weighted local features; for regions with strong discriminative power in temporal evolution, the corresponding... Element values ​​may approach This approach relies more heavily on information from temporal features that integrate local temporal dynamics. It enables fine-grained feature calibration and fusion at the pixel or vector level.

[0050] 2) A channel-level importance-weighted fusion method was used to construct a fusion network layer with extremely simple parameters. The input to this network layer is also the weighted local features. and time series features that integrate local temporal dynamics Their shapes are all .

[0051] Unlike feature-level fusion, the core of this network layer consists of two learnable scalar weight parameters: and .parameter The parameters are used to weight and integrate the overall importance of temporal feature channels that incorporate local temporal dynamics. Used to determine the overall importance of local feature channels after weighting. During initialization, and It is usually set to These two scalar parameters serve as all trainable parameters for the network layers.

[0052] The forward computation operations of this network layer are scalar multiplication and tensor summation. Specifically, scalar weights are used. Multiply by the entire temporal feature tensor scalar weights Multiply by the entire local feature tensor Then, the two weighted results are added together to obtain the fused features. This process can be expressed by the following formula: ,in, and This represents two learnable scalar weights; This represents the multiplication of a scalar and a tensor. This indicates the integration of temporal features with local temporal dynamics; Indicates the weighted local features; This represents the characteristics of the output after fusion.

[0053] During the training process, it is possible to... and Apply constraints This allows them to learn freely. This network layer structure treats each feature channel as a whole and performs weighted fusion, reflecting the model's evaluation of the discriminative contribution of the two types of features from a global perspective. Its advantages include introducing only two parameters into the network layer, extremely low computational complexity, almost no increase in model burden, effective avoidance of overfitting, and ease of optimization and deployment.

[0054] 3) The decision-level fusion approach constructs a fusion network layer that operates at the end of the model's decision-making process. This approach requires that the local feature extraction channel and the temporal feature extraction channel each have an independent classifier. Therefore, two parallel sub-steps are needed before fusion.

[0055] The first sub-step is to generate preliminary decision scores. Weighted local features. The data is fed into a dedicated fully connected classifier, which typically consists of one or more fully connected layers. The classifier ultimately outputs a preliminary decision score vector corresponding to each target class, denoted as [vector]. Similarly, it integrates temporal features that incorporate local temporal dynamics. It is fed into another independent fully connected classifier, which outputs its initial decision score vector, denoted as . The lengths of both score vectors are equal to the number of target categories. .

[0056] The second sub-step is decision score fusion. The network layer input for this step is the two preliminary decision score vectors. and At the core of this network layer is a learnable scalar weight parameter. Its range is usually in Between these two branches, the confidence levels are used to balance the decisions. The initial value is usually set to .

[0057] The forward computation operation of this network layer involves a weighted sum of the two score vectors to obtain the final fusion decision score. This process can be expressed by the following formula: ,in, Indicates a state between arrive Learnable scalar weights between them; This represents the initial decision score vector generated based on temporal features that integrate local temporal dynamics; This represents the initial decision score vector generated based on the weighted local features; This represents the final decision score vector output after fusion.

[0058] This network layer structure achieves the highest level of fusion. The two feature extraction branches are relatively independent in feature extraction and preliminary classification, effectively reducing the performance risk to the entire system caused by fluctuations in the feature quality of one branch. Learnable weights This allows the model to adaptively trust branches that perform more reliably on the current sample, based on the specific characteristics of the input sample. For example, for targets with drastic pose changes and more stable temporal features, It may learn larger values; for targets with unique structures and more prominent spatial characteristics, It is possible to learn smaller values. This approach improves the overall robustness and stability of the entire recognition system when facing different types of interference.

[0059] Depending on the selected fusion method, the feature fusion module's final output can be fused features used for final classification. (Corresponding to feature-level or channel-level fusion), or it can be a fusion decision score directly used to determine the category. (Corresponding to decision-level fusion). The specific network layer structure of the module is determined by the selected fusion strategy, providing users with flexibility to balance recognition accuracy and computational efficiency.

[0060] The classifier is used to classify targets based on fused features or fused decision scores. Its function is to determine the target category based on the output of the feature fusion module. The output of the feature fusion module has two forms: when using feature-level adaptive weighted fusion or channel-level importance weighted fusion, the output is the fused feature; when using decision-level fusion, the output is the fused decision score. The implementation of the classifier requires designing corresponding data processing paths for these two different input forms. The specific working process of the classifier is as follows: 1) When a classifier receives input, it first determines the nature of the input data. The input data may be a multidimensional tensor or a one-dimensional vector. If the input is a vector with shape... Or after being flattened tensors (where It refers to the batch size. If the total dimension of the features is [value], then the input is considered a fused feature. If the input is directly a shape [value]... vectors or matrices (where If the total number of target categories is a given value, then this input is determined as the fusion decision score. Based on the determination result, the classifier automatically selects the corresponding internal processing branch.

[0061] 2) When the input is determined to be a fusion feature In this case, the classifier needs to perform a transformation from the feature space to the class score space. This process is typically accomplished by a fully connected layer network structure, with optional feature aggregation operations added beforehand.

[0062] A common operation is global average pooling. This operation is useful for fusing features. In the time step dimension Compression is performed on the above, and the average value of each feature channel is calculated over all time steps, thus reducing the shape to The feature map is converted into a shape of The vector, denoted as . The calculation method is as follows: ,in, This indicates that a certain sample in the batch is at time step ,aisle eigenvalues, The aggregated vector is at the 1st position. The value of the dimension.

[0063] Next, the aggregated feature vectors It is fed into a fully connected layer. The weight matrix of this fully connected layer is... The bias vector is Here This represents the total number of categories of the target to be classified (e.g., different types of ships or aircraft). The linear transformation formula for the fully connected layer is: , This represents the original category score vector of the output, with the shape as follows: Each element Corresponding to the Unnormalized scores for each category.

[0064] 3) When the input is directly determined as a fusion decision score At this point, the data already represents a weighted sum of the classification results for each branch, which is essentially a preliminary class score. In this case, the classifier's processing is relatively straightforward. It is a shape of The matrix. In some designs, it can be directly... This serves as the final classification output. To obtain a standardized probability distribution, it is typically... Apply the Softmax function. The Softmax function is calculated as follows: ,in, The fusion decision score vector represents the value corresponding to the first... The scores for each category, This represents the result obtained after calculation using the Softmax function. The predicted probabilities for each category. Output It is a probability vector.

[0065] 4) For the original class score vector generated by the fully connected layer The classifier needs to transform the vector into a probability distribution. This is achieved by applying the Softmax activation function. The Softmax function applies a vector... Each element in the expression is exponentially operated on and normalized to ensure that the sum of all output values ​​is 1. The formula is as follows: ,in, This represents the first element in the original score vector output by the fully connected layer. Scores for each category; This indicates that the model predicts the input sample belongs to the first... The probability of each category; It represents the total number of categories; function This represents the natural exponential function. Ultimately, the classifier's output is a probability distribution vector. During the inference phase, the category index with the highest probability value is taken as the target classification result, i.e.: ,in, This represents the class label predicted by the model. The entire classifier implementation process completes the mapping from high-level features or decision scores to specific target classes.

[0066] S2. The recognition model is trained using labeled HRRP sequence samples. The specific implementation process is as follows: S20. A large, labeled HRRP sequence sample dataset is required. Each training sample is a triplet containing an HRRP sequence, the corresponding raw pitch and azimuth sequences, and a true class label. Assume the training dataset contains... The nth sample, the nth A sample can be represented as ,in, , This represents the total number of target categories. Before training begins, all learnable parameters of the recognition model need to be initialized. These parameters include the weight matrix of the learnable linear projection fusion layer in the angle fusion module. and bias vector The local feature extraction module includes the kernel weights and biases of the one-dimensional convolutional residual network, as well as the multilayer perceptron parameters in the dual attention module; the temporal feature extraction module includes the kernel weights of the temporal feature mapping layer, all linear projection matrices of the Transformer encoder, feedforward network parameters, and various gating weight matrices of the bidirectional long short-term memory network; and the learnable weights selected in the feature fusion module (such as...). , , or ); and the weights of the fully connected layer of the classifier. and bias Initialization methods typically employ Xavier initialization or He initialization to ensure gradient stability during the initial training phase.

[0067] S21. To quantify the difference between the model's predictions and the true labels, a differentiable loss function needs to be defined. In classification tasks, the cross-entropy loss function is typically used. For a batch containing... The loss for each sample prediction is calculated as follows: First, the model performs forward propagation on each sample in the batch to obtain the predicted probability distribution vector. Combine with real category labels The cross-entropy loss for this sample is: ,in, It is an indicator function; its value is true when the condition inside the parentheses is true. Otherwise ; The model predicts the sample. Category The probability of; This is the total number of categories. The average loss for the entire batch is: , This represents the average loss value of the current batch, which is the objective that the model needs to minimize. To minimize this loss, an optimization algorithm needs to be selected. This implementation uses the Adam optimizer. The Adam optimizer requires configuring the exponential decay rates of the first-moment and second-moment estimates for each trainable parameter in the model, typically set to 10 ... and And set a global learning rate. ,For example The optimizer's role is to update all model parameters using an adaptive algorithm based on the gradient calculated from the loss function.

[0068] S22. The training process is conducted over multiple rounds. In each training round, the entire training dataset is divided into several fixed-size batches in a random order. For each training batch, the following operations are performed: HRRP sequence data within the batch... Original pitch angle and original azimuth The data is input into the recognition model. The model performs forward propagation according to a predetermined structure: the angle fusion module processes angle information and fuses it with features; the local feature extraction module extracts weighted local features; the temporal feature extraction module extracts and integrates temporal features that incorporate local temporal dynamics; the feature fusion module fuses the two types of features; and the classifier outputs the final predicted probability. Based on the predicted probability and the true label within the batch... The average loss of the current batch is calculated using the defined cross-entropy loss formula. .

[0069] S23. After calculating the loss of the current batch Then, the backpropagation process is initiated. The backpropagation algorithm uses the chain rule, starting from the output of the loss function and calculating the gradient of the loss with respect to each learnable parameter in the model from top to bottom. Specifically, for any parameter... Calculate the gradient These gradients characterize the direction and magnitude of the effect of small changes in parameters on the loss value. After calculating the gradients of all parameters, the Adam optimizer begins to work. For each parameter... (in (Indicates the current optimization step), the optimizer executes the following update rules: In this formula, Indicates parameters In the gradient at each time step ; and These are the exponential moving averages of the first moment (mean) and the second moment (uncentered variance) of the gradient, respectively. and Yes and Deviation correction; and It is the exponential decay rate; It is the learning rate; It is an extremely small constant, for example, 10. -8 This is used to prevent division by zero. Through this update rule, all parameters in the model are updated. (Including but not limited to weights and biases) are adjusted in a way that reduces losses.

[0070] S24. After completing one or more training epochs, to evaluate the model's performance on unseen data and prevent overfitting, a separate validation dataset is needed for validation. The validation process only includes forward propagation of the model, without backpropagation or parameter updates. Metrics such as loss and accuracy on the validation set are calculated. The training process typically continues for multiple epochs until performance on the validation set no longer improves or a preset maximum number of epochs is reached. During training, learning rate scheduling strategies can be implemented, such as factorially reducing the learning rate when the validation loss stagnates. Finally, a snapshot of the model parameters that performs best on the validation set is saved as the completed recognition model.

[0071] S3. Input the HRRP sequence to be classified into the trained recognition model, and output the target classification result corresponding to the HRRP sequence to be classified. The specific implementation process is as follows: S30. Before starting classification, the parameter file of the recognition model saved after training needs to be loaded from the storage medium. These parameters include the weight matrix of the learnable linear projection fusion layer in the angle fusion module. and bias vector The local feature extraction module contains all convolutional kernel weights and biases of the one-dimensional convolutional residual network, as well as parameters in the dual attention module; the temporal feature extraction module contains convolutional kernel weights of the temporal feature mapping layer, all linear projection matrices of the Transformer encoder and feedforward network parameters, and various gating weight matrices of the bidirectional long short-term memory network; and the feature fusion module contains learnable weights (such as...). , , or ); and the weights of the fully connected layer of the classifier. and bias These parameters are loaded into the constructed recognition model network structure, and the model enters evaluation mode. Simultaneously, HRRP sequence samples to be classified are prepared. These samples should contain three parts of data: one is the HRRP sequence data to be classified. , shape ,in It is the sequence length. One is the original number of distance units; the other is the original pitch angle sequence corresponding to each frame of the sequence. Another is the original azimuth sequence. These data need to undergo the exact same preprocessing operations as in the training phase, such as magnitude normalization, to ensure consistency in the input distribution.

[0072] S31. Prepare the HRRP sequence data to be classified. Original pitch angle and original azimuth The values ​​are input as a whole into the loaded recognition model. The model performs forward computation according to its fixed network structure. The computation first flows through the angle fusion module. The angle fusion module converts the raw pitch angles into... and original azimuth Synthesized into synthetic attitude angles The formula is Next, the synthesized attitude angles are encoded to obtain angle encoding vectors, and then combined with the vectors obtained from the previous process using a learnable linear projection fusion layer. The extracted initial features are summed to output the enhanced features. .

[0073] Enhanced features The data is simultaneously fed into two parallel channels. In the local feature extraction channel, the enhanced features are sequentially passed through a one-dimensional convolutional residual network and a dual attention module. The one-dimensional convolutional residual network extracts deep spatial features, and then the dual attention module applies channel attention and spatial attention weights to these features, finally outputting the weighted local features. In the temporal feature extraction channel, the enhanced features are first transformed into initial temporal features through a temporal feature mapping layer. Then, the Transformer encoder models global dependencies and outputs global enhanced features. Finally, a bidirectional long short-term memory network captures the forward and backward local temporal dynamics, outputting temporal features that integrate the local temporal dynamics. .

[0074] Weighted local features and time series features that integrate local temporal dynamics The data is fed into the feature fusion module. Based on the fusion method (feature-level, channel-level, or decision-level) determined and fixed during model training, this module performs the corresponding fusion calculation. For example, if feature-level adaptive weighted fusion is used, then the calculation... If decision-level fusion is used, the features of the two branches will first generate preliminary decision scores through their respective classifiers. and , then calculate The output of the fusion module may be the fusion feature. It could also be a fusion decision score. .

[0075] S32. The output of the feature fusion module is passed to the classifier. If the input is fused features... The classifier processes the data (e.g., through global average pooling) and then performs a linear transformation through a fully connected layer to generate the original class score vector. The formula is Subsequently, regarding Apply the Softmax function to convert it into a probability distribution vector. ,in , This represents the total number of categories. If the input is directly the fusion decision score... The classifier then directly applies the Softmax function to obtain the probability distribution vector. .

[0076] Obtain the probability distribution vector Next, the recognition model needs to output the final target classification result. This is done by selecting the category index corresponding to the highest probability value. The specific operation is as follows: , The class label predicted by the model is a value between... arrive Integers between [a certain range]. The function returns the probability. Reaching the largest index This index This corresponds to the target category to which the HRRP sequence to be classified belongs (e.g., a specific type of ship or aircraft).

[0077] S33, The model will calculate the predicted category labels. This serves as the final output of this reasoning. In practical applications, this result can be displayed directly or compared with probability values. The results are output together to provide confidence levels, or fed into a larger system for further decision-making. This concludes the complete identification process for a single HRRP sequence sample to be classified. For multiple samples input in batches, the above process is executed in parallel in a vectorized manner, ultimately outputting a list containing the prediction results for all samples.

[0078] The targets can be ships, aircraft, ground vehicles, and man-made facilities, including wind turbines, bridges, towers, etc.

[0079] Further elaboration of the technical solution of the present invention is provided through another embodiment. Based on an attention mechanism and a bidirectional long short-term memory network (BSLM) recognition model, namely the ACT-BiLSTM model, deep local features and temporal features of high-resolution distance image sequences are extracted through dual feature channels, and recognition and classification are completed after feature fusion. The local feature channels use one-dimensional convolutional kernels to extract features, and a dual attention module mechanism is added to enhance feature extraction. After fully considering the bidirectional temporal correlation of the high-resolution distance image sequences, a bidirectional BSLM network is used for temporal feature extraction. The overall model architecture is as follows: Figure 2As shown in the diagram, the framework clearly depicts the complete path of data from input to output. The entire model mainly consists of four parts: angle fusion module, local feature extraction module, temporal feature extraction module, and feature fusion module. Figure 2 In this model, the input is the HRRP sequence. This data first enters the angle fusion module. This module serializes and integrates the broadband radar cross section data to generate a more easily processed high-resolution range image sequence representation, and deeply fuses the angle information into it. The output of the angle fusion module is simultaneously fed into two parallel channels.

[0080] Local feature extraction module. First, multi-scale one-dimensional convolutional kernels are used to perform preliminary feature fusion on the input. Figure 2 The annotations show 3x3, 5x5, and 7x7 convolutional kernels, indicating that the model uses convolutional kernels of different sizes in parallel to capture local patterns within different receptive fields. The initially fused features are then sequentially passed through four cascaded 1DResNet blocks, i.e., four one-dimensional residual network blocks, for deep spatial feature extraction. These four blocks... Figure 2 The deep features are explicitly labeled as 1DResNet, 1DResNet, 1DResNet, and 1DResNet. The extracted deep features are then passed through a CBAM (Convolutional Block Attention Module) module for adaptive feature enhancement, and finally output the weighted local features.

[0081] The temporal feature extraction module first performs layer normalization on the input. The normalized features then enter a multi-head attention layer, the core component of the Transformer encoder, used to model the global dependencies of the sequence. The output of the attention layer undergoes layer normalization again and is then fed into a feedforward layer for nonlinear transformation. This structure constitutes a Transformer encoder layer. Afterward, the features are input into a bidirectional long short-term memory network composed of multiple BiLSTM units for temporal modeling. The figure clearly shows the stacked structure of the BiLSTMs to capture the local temporal dynamics of the forward and backward directions, and the output is a temporal feature that integrates these local temporal dynamics.

[0082] Finally, the features from both the upper and lower channels are fed into the feature fusion module. This module fuses the weighted local features and the temporal features that integrate local temporal dynamics. The fused result enters the classification output section, i.e., the classifier, which ultimately produces the prediction result for the target category.

[0083] The angle fusion module is responsible for serializing and integrating broadband radar cross-section data to generate a continuous high-resolution range image sequence. The innovation of this module lies in constructing an angle representation that is highly adaptable to changes in target attitude from the original attitude-sensitive data, and deeply integrating it into the features using an effective encoding method.

[0084] Specifically, the angle fusion module first converts the original pitch angle... With azimuth Information, geometrically synthesized, forms a synthetic attitude angle with greater physical meaning. This synthetic attitude angle directly characterizes the overall spatial orientation of the target's scattering center relative to the radar line of sight, and its calculation formula is as follows: ,in, This represents the time step index of the high-resolution range image sequence. This is the synthesized angle. Even subtle changes in the scattering center distribution and intensity in high-resolution range images can lead to significant alterations, making it a key variable affecting recognition stability. This module explicitly models this key variable to improve the model's robustness to changes in target pose.

[0085] Secondly, to efficiently embed this continuous attitude angle information into the neural network, this module borrows the idea of ​​position encoding from the Transformer architecture and designs an angle encoding strategy. Specifically, the attitude angle is mapped through a set of sine and cosine functions with preset frequencies to generate a high-dimensional encoding vector with fixed dimensions that contains the periodic variation of the angle. This encoding method can provide the model with stable and discriminative prior knowledge of the angle.

[0086] Finally, through a learnable linear transformation layer, the generated high-dimensional angle encoding vector is projected to a dimension consistent with the one-dimensional high-resolution range image feature space. By adding features, the angle semantic information and radar echo features are effectively aligned and deeply fused at the depth level, thereby providing more comprehensive and discriminative fused features for subsequent recognition networks.

[0087] In the local feature extraction channel, a residual network structure based on one-dimensional convolution is adopted, combined with a bipartite attention mechanism for feature enhancement. One-dimensional convolutional neural networks, as an efficient feedforward architecture, can effectively process structured data, including one-dimensional temporal signals, through their unique convolutional kernel sliding, pooling downsampling, and fully connected layer cascading. The residual structure alleviates the gradient decay problem in deep networks through cross-layer connections, ensuring the stability of model training.

[0088] The dual attention mechanism, based on convolution, enhances the model's ability to perceive and select key features by sequentially introducing channel attention and spatial attention sub-modules. This mechanism effectively improves the quality of feature representation and the model's classification performance without significantly increasing computational complexity. The dual attention mechanism is also known as the CBAM attention module.

[0089] Figure 3 This demonstrates the network structure of the convolutional block attention module, which is also the specific structure of the dual attention module used in this invention. The module's input is the input feature matrix. This feature matrix first enters the channel attention module. The channel attention module generates channel-dimensional weight vectors by aggregating spatial information. In the diagram, Mc represents the input feature matrix. Next, the features obtained by multiplying the channel attention weights by the input feature matrix are input into the spatial attention module. The spatial attention module generates a spatial dimension weight matrix by aggregating channel information. , Figure 3 The term is denoted by Ms. Finally, the spatial attention weights are multiplied by the features, and the output is a double-weighted feature matrix.

[0090] In terms of specific structural design, the local feature extraction module first uses multi-scale convolutional kernels to perform preliminary feature fusion on the input data to capture local patterns under different receptive fields. Subsequently, the feature maps are sequentially passed through four cascaded one-dimensional residual convolutional blocks for deep spatial feature extraction. The residual structure ensures the effective propagation of gradients in the deep network through cross-layer connections, alleviating the gradient vanishing problem and promoting stable convergence during training.

[0091] Finally, the extracted features are adaptively weighted by a dual attention module to highlight local regions that significantly contribute to the target recognition task and suppress redundant or interfering responses, thereby outputting more discriminative deep spatial features. This integrated design improves the ability to model deep local structures in high-resolution range image data, thus enhancing the accuracy and robustness of the overall recognition system.

[0092] In high-resolution range image recognition tasks, the adaptive feature enhancement module introduces an attention mechanism to effectively enhance the model's ability to extract discriminative features of targets. Especially when processing high-resolution range image data containing a large amount of redundant echoes and noise interference, this mechanism can adaptively focus on the effective reflection area of ​​the target, suppressing the influence of non-critical components on the recognition results. This has significant research value and application prospects for improving the practicality and robustness of radar target recognition systems.

[0093] This invention introduces a classic channel-space dual-dimensional attention module, namely the bi-attention module. This module can sequentially model the channel relationships and spatial locations of input features. By processing the dependencies between channels and the spatial importance distribution in turn, it enhances the model's ability to perceive and select multi-dimensional discriminative features. The specific workflow of the bi-attention module is as follows: First, the input feature map enters the channel attention submodule. This submodule generates channel-dimensional weight vectors by aggregating spatial information, thereby modeling the interdependencies between feature channels and adaptively weighting each channel. Then, the weighted feature map enters the spatial attention submodule. This submodule generates a spatial-dimensional weight matrix by aggregating channel information, thereby analyzing and emphasizing key regions in the feature map in spatial location. Finally, the features output by the spatial attention submodule are integrated. This process, while maintaining the dimensionality of the output feature map, achieves sequential and refined enhancement of key features in both channel and spatial dimensions, effectively improving the quality and discriminative power of feature representation.

[0094] Based on the above design, this dual attention module can collaboratively utilize complementary information in both channel and spatial dimensions to achieve multi-angle focusing on discriminative regions in the input feature map. This mechanism can accurately enhance the response to key local features in the target recognition task, and effectively suppress the activation intensity of noise and background interference components. Thus, while improving the specificity of feature selection, it also enhances the model's generalization ability and recognition robustness in complex electromagnetic environments, ultimately improving the overall performance of the recognition system.

[0095] In terms of temporal feature extraction, traditional recurrent neural networks (RNNs) have significant limitations in sequence modeling, particularly in the tendency for gradient vanishing or exploding during training, which restricts their ability to model long sequence data. To overcome this bottleneck, researchers have proposed RNN variants with gated unit structures, such as Long Short-Term Memory (LSTM) networks and gated recurrent units (GRUs). These models effectively regulate the transmission and forgetting of sequence information by introducing gating mechanisms, thereby enhancing the network's ability to model long-range dependencies.

[0096] In addition to the improved recurrent neural networks mentioned above, the Transformer architecture demonstrates even superior performance in sequence data analysis. Its core lies in employing a self-attention mechanism, establishing global dependencies by directly calculating the association weights between any two positions in the sequence. This parallel processing approach breaks through the efficiency bottleneck of traditional recurrent neural network sequence computation, not only improving the processing speed for long sequence data but also enhancing the model's ability to capture complex dependency patterns.

[0097] To comprehensively improve the feature extraction efficiency of radar cross section data, this invention employs a hybrid architecture combining a Transformer encoder and a bidirectional long short-term memory (LSTM) network in the temporal feature extraction channel. The bidirectional LSTM network deploys two independent LSTM network units in parallel, modeling the temporal data from both forward and backward dimensions, achieving bidirectional encoding of sequence context information. This bidirectional learning mechanism enables the model to simultaneously capture historical and future context information, thereby obtaining richer temporal feature representations. This hybrid architecture retains the advantages of the bidirectional LSTM network in capturing local temporal dynamics while leveraging the Transformer's self-attention mechanism to enhance the modeling ability of global context information, thus effectively addressing the inherent time-shift sensitivity and long sequence dependencies in radar echoes.

[0098] Finally, in the feature fusion stage, the heterogeneous features extracted from the temporal and spatial channels are jointly encoded and integrated. To achieve optimal recognition performance, the key lies in designing an efficient feature fusion strategy. This invention proposes a flexible and adaptive multi-feature fusion framework. Specifically, this framework supports multiple fusion modes from the feature level to the decision level by constructing a learnable dynamic trade-off mechanism, thereby achieving optimal synergy and enhanced discriminative power of temporal and spatial features in the final decision. The main fusion methods include feature-level adaptive weighted fusion, channel-level importance-weighted fusion, and decision-level fusion.

[0099] Feature-level adaptive weighted fusion achieves the finest-grained fusion at the feature level. Define a feature with dual-channel output features. and Learnable parameter matrices with identical dimensions This is achieved by using element-wise multiplication to independently weight each specific location and channel in different feature maps. First, for local features... Perform direct weighting on time series features Complementary weighting (i.e.) Then sum the results to obtain the fusion features. : ,in, This represents element-wise multiplication. The initial value was set to 0.5 and optimized through backpropagation during model training. It is a tensor isomorphic to the feature map, rather than a single scalar. This allows the model to adaptively assign different fusion weights to the feature map at each spatial location and channel, depending on the specific context of the input samples. For regions with significant spatial features, The weights at corresponding positions may be close to 1 to preserve more local features; while in regions where temporal changes are critical, the weights may be close to 0 to rely more on temporal features. This dynamic and fine-grained weight allocation mechanism enables deep fusion and calibration of two heterogeneous features at the pixel or vector level.

[0100] Channel-level importance-weighted fusion performs coarse-grained fusion at the channel dimension, resulting in higher computational efficiency. Its core principle is to simplify the fusion weights into learnable scalar weights. and , representing the global importance of the temporal and spatial feature channels, respectively. For temporal features... and spatial features Sum after weighting: ,in, and As a learnable scalar, it is initialized to 0.5 and constrained. Alternatively, it can be allowed to learn and optimize independently. Each feature channel can be treated as a whole and weighted accordingly. Scalar and This reflects the model's evaluation of the discriminative contribution of the two types of features from a global perspective. The advantages of this approach are that it has very few parameters, hardly increases the model complexity, effectively avoids overfitting, and is easy to optimize and deploy.

[0101] Decision-level fusion, performed at the very end of the model, offers the highest robustness. Its operation is not on features, but rather on the preliminary decision score vectors made by the two branches based on their respective features. and The final decision score is obtained by weighted integration of the two preliminary decision scores. , ,in, For a middle Learnable scalar weights are used to balance the confidence levels of the two branch decisions. Decision-level fusion is the highest level of fusion. The two branches are relatively independent in feature extraction and classification, effectively reducing the overall performance risk caused by fluctuations in the feature quality of one branch. Learnable weights This allows the model to adaptively trust branches that perform more reliably on specific samples. For example, for time-insensitive targets, time-series branches may be more reliable. For targets with significant structural features, spatial branching may be more reliable. The size is relatively small. This fusion approach improves the overall robustness and stability of the system when facing different types of interference.

[0102] In another embodiment, it includes: 1) The raw data acquired by the broadband radar undergoes motion compensation, alignment, and normalization preprocessing to form a preliminary high-resolution range image sequence. Subsequently, it is processed through an angle fusion module. The original elevation angle corresponding to each high-resolution range image unit is then... Compared with the original azimuth angle The synthesized attitude angles are obtained by using a geometric synthesis formula. The calculation formula is: ,in, This represents the time step index. Borrowing from the positional encoding concept of the Transformer, a set of sine and cosine functions are used to transform the synthesized pose angle sequence, generating a high-dimensional angle encoding vector. This angle encoding vector is then projected through a learnable linear layer to the same dimension as the initial features obtained from the high-resolution range image sequence, and fused with the initial features through feature addition to generate enhanced features.

[0103] 2) The enhanced features are simultaneously input into two parallel feature extraction channels, specifically: ① Local Feature Extraction Channel: A residual network structure based on one-dimensional convolution is used to extract spatial features, and a dual attention module is embedded in the network. This module sequentially applies channel attention weighting and spatial attention weighting to the features, outputting the weighted local features. ② Temporal Feature Extraction Channel: A hybrid architecture combining a Transformer encoder and a bidirectional long short-term memory network is adopted. First, the self-attention mechanism of the Transformer encoder is used to capture the global dependencies of the sequence, outputting global enhanced features; then, the bidirectional long short-term memory network captures the forward and backward local temporal dynamics in the global enhanced features, and the bidirectional information is concatenated to form a temporal feature that integrates the local temporal dynamics.

[0104] 3) The weighted local features and the time-series features that integrate local temporal dynamics are fused. Three optional fusion strategies are provided, specifically: ① Feature-level adaptive weighted fusion: Define a learnable parameter matrix with the same dimension as the dual-channel output features. The fused feature is obtained by directly weighting local features through element-wise multiplication and by complementary weighting of temporal features, and then summing the results. The formula is as follows: ② Channel-level importance-weighted fusion: The fusion weights are simplified into two learnable scalars. and The temporal and spatial features are weighted separately and then summed. The formula is: ③ Decision-level fusion: Each of the two channels generates an initial decision score vector through an independent classifier. and Then through a learnable scalar weight The two are weighted and integrated to obtain the fusion decision score. The formula is: The fused features or fusion decision scores are fed into the final classifier, which outputs the target category identification result.

[0105] 4) The entire model is trained end-to-end using the cross-entropy loss function and the Adam optimizer. By setting hyperparameters such as the initial learning rate, batch size, and number of training epochs, the model parameters are iteratively optimized on the training set until convergence. A validation set can be used to monitor performance during training.

[0106] 5) Input the HRRP sequence to be classified into the trained recognition model, and output the target classification result corresponding to the HRRP sequence to be classified.

[0107] Based on the above analysis, existing recognition methods combining convolutional neural networks and unidirectional long short-term memory networks still suffer from insufficient feature extraction. To address this, this invention proposes a recognition model based on an attention mechanism and a bidirectional long short-term memory network. This model can effectively extract deep local features and long-range temporal features from high-resolution distance image sequence data. The model employs a dual-feature channel fusion architecture to enhance its comprehensive feature extraction capabilities.

[0108] The recognition model based on attention mechanism and bidirectional long short-term memory network has made two key improvements on the traditional convolutional neural network and long short-term memory network structure: First, a dual attention module is integrated into the local feature extraction module, which can adaptively enhance the weight of key local features; Second, a hybrid architecture of Transformer encoder and bidirectional long short-term memory network is constructed in the temporal feature extraction part. By fusing forward and backward context information and combining the global dependency modeling capability of Transformer encoder, the model's ability to capture and represent spatiotemporal discriminative features in high-resolution distance image sequence data is improved.

[0109] The recognition model based on attention mechanisms and bidirectional long short-term memory networks provided by this invention achieves the following beneficial effects compared to existing technologies through a series of collaboratively designed modular innovations: 1) Improved robustness of the model to target attitude changes: Through the angle fusion module, the original pitch and azimuth angles are explicitly synthesized into a synthetic attitude angle with clear physical meaning, and the Transformer position encoding concept is adopted to transform it into a high-dimensional angle encoding vector. This design enables the model to directly learn and fuse the spatial geometric information of the target, enhances its adaptability to distortion of high-resolution range image sequences caused by target attitude changes, and improves the stability of the recognition system under complex observation conditions.

[0110] 2) Achieved more refined and adaptive feature fusion: The flexible multi-feature fusion framework proposed in this invention provides three selectable fusion strategies: feature-level adaptive weighted fusion, channel-level importance-weighted fusion, and decision-level fusion. This design allows the model to adaptively select the fusion granularity according to different task requirements and data characteristics. Feature-level adaptive weighted fusion achieves dynamic calibration of spatiotemporal features at the pixel or vector level through a learnable parameter matrix, preserving discriminative information and thus improving the representation quality and discriminative power of the fused features.

[0111] 3) Enhanced extraction efficiency of local and temporal features: In local feature extraction, by integrating a dual attention module into a one-dimensional convolutional residual network, automatic focusing and noise suppression of key scattering centers in high-resolution range image sequences are achieved, enabling the model to extract more discriminative weighted local features. In temporal feature extraction, a hybrid architecture of a Transformer encoder and a bidirectional long short-term memory network is used to leverage the global dependency modeling capability of the Transformer encoder and the bidirectional local temporal dynamic capture advantage of the bidirectional long short-term memory network. This overcomes the time-shift sensitivity and long-sequence dependency problems of high-resolution range image sequence data, extracting more comprehensive temporal features that integrate local temporal dynamics.

[0112] 4) Balancing Model Performance and Implementation Efficiency: The multi-layered fusion strategy provided by this invention constitutes a complete technical system from fine-grained to coarse-grained. Users can flexibly choose between computationally efficient channel-level importance-weighted fusion or robust decision-level fusion, depending on the trade-off between recognition accuracy and computational resources in practical applications, making this invention both high-performance and practical.

[0113] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0114] like Figure 4 As shown, an embodiment of the present invention provides a target classification system 200 based on HRRP sequences, which includes a model building module 201, a model training module 202, and a model application module 203. The model building module 201 is used to: construct a recognition model based on an attention mechanism and a bidirectional long short-term memory network. The recognition model includes an angle fusion module, a local feature extraction module, a temporal feature extraction module, a feature fusion module, and a classifier. The angle fusion module is used to: synthesize the original elevation angle and original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, and obtain enhanced features based on the synthetic attitude angle and the initial features obtained from the HRRP sequence. The local feature extraction module is used to: obtain weighted local features based on the enhanced features. The temporal feature extraction module is used to: obtain temporal features that integrate local temporal dynamics based on the enhanced features. The feature fusion module is used to: fuse the weighted local features and the temporal features that integrate local temporal dynamics to obtain fused features or fused decision scores. The classifier is used to: classify the target based on the fused features or fused decision scores. The model training module 202 is used to: train the recognition model using labeled HRRP sequence samples; The model application module 203 is used to: input the HRRP sequence to be classified into the trained recognition model, and output the target classification result corresponding to the HRRP sequence to be classified.

[0115] Optionally, in the above technical solution, the angle fusion module is specifically used to: synthesize the original elevation angle and original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, encode the synthetic attitude angle to generate an angle encoding vector, and fuse the angle encoding vector with the initial features obtained from the HRRP sequence to obtain enhanced features.

[0116] Optionally, in the above technical solution, the local feature extraction module includes a one-dimensional convolutional residual network and a dual attention module. The local feature extraction module is specifically used to: extract spatial features from the enhanced features through the one-dimensional convolutional residual network, and perform channel attention and spatial attention weighting on the spatial features through the dual attention module to output the weighted local features.

[0117] Optionally, in the above technical solution, the temporal feature extraction module includes a temporal feature mapping layer, a Transformer encoder, and a bidirectional long short-term memory network. Specifically, the temporal feature extraction module is used to: extract initial temporal features from the enhanced features through the temporal feature mapping layer; model global dependencies on the initial temporal features through the Transformer encoder and output global enhanced features; and capture the forward and backward local temporal dynamics in the global enhanced features through the bidirectional long short-term memory network to obtain temporal features that integrate local temporal dynamics.

[0118] Optionally, in the above technical solution, the feature fusion module is specifically used to: obtain fused features or fused decision scores by using at least one of the following fusion methods: feature-level adaptive weighted fusion, channel-level importance weighted fusion, and decision-level fusion.

[0119] It should be noted that the beneficial effects of the target classification system 200 based on HRRP sequences provided in the above embodiments are the same as those of the target classification method based on HRRP sequences described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0120] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned target classification methods based on HRRP sequences.

[0121] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned target classification methods based on HRRP sequences.

[0122] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0123] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A target classification method based on HRRP sequences, characterized in that, include: A recognition model based on an attention mechanism and a bidirectional long short-term memory network is constructed. The recognition model includes an angle fusion module, a local feature extraction module, a temporal feature extraction module, a feature fusion module, and a classifier. The angle fusion module is used to synthesize the original elevation angle and the original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, and to obtain enhanced features based on the synthetic attitude angle and the initial features obtained from the HRRP sequence. The local feature extraction module is used to: obtain weighted local features based on the enhanced features; the temporal feature extraction module is used to: obtain temporal features integrating local temporal dynamics based on the enhanced features; the feature fusion module is used to: fuse the weighted local features and the temporal features integrating local temporal dynamics to obtain fused features or fused decision scores; the classifier is used to: classify targets based on the fused features or the fused decision scores. The recognition model was trained using labeled HRRP sequence samples; The HRRP sequence to be classified is input into the trained recognition model, and the target classification result corresponding to the HRRP sequence to be classified is output.

2. The target classification method based on HRRP sequences according to claim 1, characterized in that, The angle fusion module is specifically used to: synthesize the original elevation angle and original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, encode the synthetic attitude angle to generate an angle encoding vector, and fuse the angle encoding vector with the initial features obtained from the HRRP sequence to obtain the enhanced features.

3. The target classification method based on HRRP sequences according to claim 2, characterized in that, The local feature extraction module includes a one-dimensional convolutional residual network and a dual attention module. Specifically, the local feature extraction module is used to: extract spatial features from the enhanced features through the one-dimensional convolutional residual network, and perform channel attention and spatial attention weighting on the spatial features through the dual attention module to output the weighted local features.

4. The target classification method based on HRRP sequences according to claim 3, characterized in that, The temporal feature extraction module includes a temporal feature mapping layer, a Transformer encoder, and a bidirectional long short-term memory network. Specifically, the temporal feature extraction module is used to: extract initial temporal features from the enhanced features through the temporal feature mapping layer; and model global dependencies on the initial temporal features through the Transformer encoder to output global enhanced features. By capturing the forward and backward local temporal dynamics in the global enhancement features through the bidirectional long short-term memory network, a temporal feature integrating the local temporal dynamics is obtained.

5. The target classification method based on HRRP sequences according to claim 4, characterized in that, The feature fusion module is specifically used to: obtain fused features or fused decision scores by employing at least one of the following fusion methods: feature-level adaptive weighted fusion, channel-level importance weighted fusion, and decision-level fusion.

6. A target classification system based on HRRP sequences, characterized in that, It includes a model building module, a model training module, and a model application module; The model building module is used to: construct a recognition model based on an attention mechanism and a bidirectional long short-term memory network. The recognition model includes an angle fusion module, a local feature extraction module, a temporal feature extraction module, a feature fusion module, and a classifier. The angle fusion module is used to: synthesize the original elevation angle and original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, and obtain enhanced features based on the synthetic attitude angle and the initial features obtained from the HRRP sequence. The local feature extraction module is used to: obtain weighted local features based on the enhanced features; the temporal feature extraction module is used to: obtain temporal features integrating local temporal dynamics based on the enhanced features; the feature fusion module is used to: fuse the weighted local features and the temporal features integrating local temporal dynamics to obtain fused features or fused decision scores; the classifier is used to: classify targets based on the fused features or the fused decision scores. The model training module is used to: train the recognition model using labeled HRRP sequence samples; The model application module is used to: input the HRRP sequence to be classified into the trained recognition model, and output the target classification result corresponding to the HRRP sequence to be classified.

7. A target classification system based on HRRP sequences according to claim 6, characterized in that, The angle fusion module is specifically used to: synthesize the original elevation angle and original azimuth angle of the target relative to the radar line of sight in the HRRP sequence into a synthetic attitude angle, encode the synthetic attitude angle to generate an angle encoding vector, and fuse the angle encoding vector with the initial features obtained from the HRRP sequence to obtain the enhanced features.

8. A target classification system based on HRRP sequences according to claim 7, characterized in that, The local feature extraction module includes a one-dimensional convolutional residual network and a dual attention module. Specifically, the local feature extraction module is used to: extract spatial features from the enhanced features through the one-dimensional convolutional residual network, and perform channel attention and spatial attention weighting on the spatial features through the dual attention module to output the weighted local features.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a target classification method based on HRRP sequences as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a target classification method based on HRRP sequences as described in any one of claims 1 to 5.

Images