High-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement

Abstract

The application discloses a high-resolution one-dimensional range profile target recognition method based on hierarchical fusion and dynamic feature enhancement, first, the HRRP data obtained by the original radar is pretreated to obtain an amplitude spectrum; secondly, aiming at the pretreated HRRP amplitude information, 13 types of target features are extracted from three key dimensions of statistical characteristics, geometric structure and target size, so that the target is multi-dimensionally represented; subsequently, a multiple feature enhancement backbone network is constructed, and the feature weight is adjusted through a dynamic gating mechanism guided by a primary classification and a multi-head attention mechanism; then, the enhanced features are input into a Transformer encoder for deep sequence modeling, and are spliced with original features; finally, a double-head decision layer containing binary classification and multi-classification is constructed to output the final recognition result. Through coarse and fine granularity joint decision and feature dynamic enhancement, the application effectively suppresses environmental interference, and significantly improves the accuracy and robustness of radar target recognition.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement. Background Technology

[0002] High-resolution range profile (HRRP) is a technique that uses broadband radar signals to obtain the vector sum of the projections of target scattering point echoes onto the radar's line-of-sight. By combining the time and position information of the reflected signals, a range profile is formed, reflecting the target's size, structure, and the distribution of scattering points along the range direction. It is a key technology for radar target detection. The peaks in HRRP correspond to the strong scattering centers of the target on the radar's line of sight; accurately detecting and locating these peaks is crucial for subsequent target feature extraction and identification.

[0003] Traditional HRRP feature extraction techniques are mostly based on a single dimension. Features based on statistical moments, such as variance and global entropy distribution, can reflect signal fluctuations and energy distribution. However, these techniques treat HRRP as a one-dimensional signal sequence, completely ignoring the spatial relationships between scattering points and the geometric structure of the target. Features based on shape and energy can describe the contour characteristics and energy concentration of the target, and have a certain ability to distinguish targets with different structures. However, they are extremely sensitive to noise clutter. In low signal-to-noise ratio environments, weak scattering points may be submerged by noise, leading to feature extraction failure and poor robustness. In summary, single-dimensional features can only capture local characteristics of the target and cannot form a complete and comprehensive characterization of the target's electromagnetic scattering properties, severely limiting their application in complex scenarios.

[0004] In recent years, neural networks have been increasingly used for feature extraction tasks. Currently, most HRRP target recognition methods based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs) typically convert one-dimensional HRRP data into time-domain, frequency-domain, or time-frequency-domain image / sequence formats for direct processing. Against this backdrop, their practical application still faces two major core challenges: First, the high dependence on labeled data and the problem of feature consistency. Neural network training heavily relies on a large amount of accurately labeled data. However, in HRRP data, target regions are difficult to define clearly, and the HRRP data of the same target at different azimuth and elevation angles will undergo translation, scaling, and amplitude changes, resulting in highly attitude-dependent features. This makes constructing a large-scale dataset covering all attitudes with uniform labels extremely costly, or even infeasible. Second, the robustness problem in noisy and cluttered environments. In real radar environments, HRRP data typically contains abundant background clutter and other interference components. Usually, the effective echo energy of the target accounts for a low proportion of the overall signal, leading to an unbalanced data distribution. This imbalance can cause the network to focus excessively on noise and clutter features rather than target features during training, leading to ineffective learning or overfitting and a decrease in the model's generalization ability and robustness in real-world scenarios. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a high-resolution one-dimensional range profile target recognition method based on hierarchical fusion and dynamic feature enhancement. First, the original HRRP (High Resolution Range Profile) data acquired by radar is preprocessed to obtain the amplitude spectrum. Second, based on the preprocessed HRRP amplitude information, 13 target features are extracted from three key dimensions: statistical characteristics, geometric structure, and target size, achieving multi-dimensional target representation. Subsequently, a multi-feature enhancement backbone network is constructed, utilizing a deep cross-network (DCN) to mine feature combinations, and employing a dynamic gating mechanism guided by primary classification and a multi-head attention mechanism to adaptively adjust feature weights according to sample semantics. Next, the enhanced features are input into a Transformer encoder for deep sequence modeling and concatenated with the original features. Finally, a dual-head decision layer containing binary and multi-class classification is constructed, using the interference / target probability output from binary classification to perform hierarchical masking correction on the multi-class classification results, outputting the final recognition result. This invention effectively suppresses environmental interference and significantly improves the accuracy and robustness of radar target recognition through coarse-grained joint decision-making and dynamic feature enhancement.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Acquire the raw radar HRRP signal and preprocess it to extract a multi-dimensional physical feature vector containing statistical characteristics, geometric structure and target size; Step 2: Construct a target recognition network that includes a feature enhancement backbone module, a Transformer encoding module, and a dual-head hierarchical decision module; Step 3: Input the multidimensional feature vectors extracted in Step 1 into the Deep Cross Network (DCN) for high-order feature combination, and input them into the coarse guiding branch to obtain the coarse guiding probability distribution of the current sample belonging to each category; Step 4: The coarse guided probability distribution obtained in Step 3 is used to adaptively weight the feature vector using a dynamic gating mechanism, and combined with subsequent interaction and attention mechanisms to generate an enhanced feature vector. Step 5: Map the enhanced features obtained in Step 4 into a sequence input to the Transformer backbone for deep context modeling, and concatenate the deep semantic features output by Transformer with the high-order enhanced features output in Step 4 along the channel dimension to obtain the wide-depth fusion features. Step 6: Input the width and depth fusion features obtained in Step 5 into the binary classification decision head and the multi-class classification decision head respectively; use the interference / target probability distribution output by the binary classification decision head as a soft mask to perform hierarchical weighted correction on the output probability of the multi-class classification decision head, and calculate the final target recognition result.

[0007] Preferably, step 1 specifically comprises: Step 1.1: Obtain the original radar HRRP complex signal, convert it into an amplitude spectrum using formula (1), and perform amplitude normalization using formula (2); then, use the cell average constant false alarm rate detection algorithm to locate and capture the target area signal and remove background noise interference; the formula for converting complex data into an amplitude spectrum is: (1) in, For amplitude spectrum, The real part of the original HRRP signal. This represents the imaginary part of the original HRRP signal. The formula for normalizing the amplitude spectrum is: (2) in, The amplitude spectrum of HRRP The normalized amplitude spectrum; Step 1.2: Based on the preprocessed HRRP, extract target features from three aspects: statistical properties, geometric structure, and target size; The statistical characteristics are parameters of the statistical regularity and inherent fluctuation characteristics of the signal, including seven features: mean, variance, entropy, radial energy, third moment, fourth central moment, and super-average amplitude. The average value characteristic is used to describe the average intensity of scattered points in the HRRP signal, and the calculation formula is as follows: (3) Where N is the number of HRRP points after preprocessing. The HRRP amplitude spectrum signal is the target. The formula for calculating variance characteristics is: (4) Entropy features reflect the amplitude distribution of target scattering points in HRRP, and the calculation formula is: (5) in, The HRRP amplitude spectrum signal after normalization of the target amplitude; Radial energy characteristics characterize the electromagnetic reflection capability of a target on radar, and the calculation formula is as follows: (6) The third-order moment feature measures the symmetry of the target HRRP data about the mean, and is calculated using the following formula: (7) in, This indicates the location of the target HRRP data distribution center; The fourth-order central moment characteristic characterizes the tail thickness of HRRP data, and the calculation formula is as follows: (8) The formula for calculating the super-average amplitude characteristic is: (9) The geometric structure refers to parameters that describe the target's shape, structural complexity, or spatial distribution characteristics, including three features: irregularity, descaled structure, and equivalent scattering center dimension. The formula for calculating the irregularity feature is: (10) in, , , These represent the first [number] in the HRRP amplitude spectrum data. -1 element, the first The element, the first +1 element; The formula for calculating descaled structural features is: (11) The formula for calculating the equivalent scattering center number characteristic is: (12) in, Represents the unit step function; The target size is a parameter that directly or indirectly describes the size of the target, including three features: relative fullness of the time-domain signal, equivalent target size, and radial size. The formula for calculating the relative fullness characteristic of a time-domain signal is: (13) in, Indicates the first In the nth sample The amplitude value of each sampling point Indicates the first The maximum value of the amplitude of each sample. Indicates the number of sampling points. Indicates the total number of samples; The formula for calculating the equivalent target size feature is: (14) in, This indicates that the echo intensity is greater than the threshold. The distance cell indices are arranged in ascending order; This represents the first index of the sequence. Indicates the last index of the sequence; The formula for calculating radial dimension features is: (15) in, Represents the set of target scattering points; Indicates the first The slant distance of each scattering point ; Step 1.3: Standardize the features obtained for the same target and in the same scene. The calculation formula is as follows: (16) in, , representing the One feature; Indicates the first The first sample The standardized values ​​of each feature; After standardizing each feature, a multidimensional feature vector is formed. , as the original input of the network.

[0008] Preferably, step 2 specifically comprises: The feature enhancement backbone module is composed of a deep cross network (DCN), a coarse guiding branch, a dynamic feature gating unit, a cross-feature interaction unit, and a multi-head attention unit cascaded together. The Transformer encoding module consists of a feature embedding layer, a position encoding layer, and a multi-layer Transformer encoder. The dual-head hierarchical decision module includes a binary classification decision head for distinguishing between "interference / target" and a multi-classification decision head for fine-grained identification.

[0009] Preferably, step 3 specifically comprises: First, the feature vector The DCN module performs feature crossover operations, outputting features that include higher-order combination information. ; Then Input to the coarse-guided branch, and output the coarse-guided class probability distribution for the current sample according to formula (17). : (17) in, and These are learnable weights and biases.

[0010] Preferably, step 4 specifically comprises: Step 4.1: Add features The input dynamic feature gating unit, after multiple linear mappings, yields a dimension of... Feature importance map for each category ; Use the class probabilities obtained in step 3 right Perform weighted fusion to generate the importance vector of the current sample. : (18) Generating feature-gated masks using the Sigmoid function And through a learnable scaling factor Intensity adjustment is performed by multiplying the original features element-wise to enhance the features, resulting in the gated features. : (19) in, This represents the Sigmoid activation function. Presentation layer normalization processing, This indicates element-wise multiplication. This represents the learnable scaling factor; Step 4.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input cross-feature interaction units, capture the nonlinear topological relationships between features through a multi-layer fully connected network (MLP), and output interactive features. Finally, residual connections and layer normalization are performed: (20) Step 4.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Projecting Q, K, and V, the enhanced features are calculated. Finally, residual connections and layer normalization are performed. (twenty one) in, The dimension of the key vector is represented by Q, where Q represents the query matrix, K represents the key matrix, and V represents the value matrix. This represents the feature output after attention weighting. This represents the final feature after residual connection and layer normalization.

[0011] Preferably, step 5 specifically comprises: Step 5.1: First, [the text abruptly ends here, likely due to an incomplete sentence or a formatting error. Mapped to a high-dimensional space and with added positional encoding, a token sequence is formed and input into the Transformer encoder to extract deep features. ; Step 5.2: Employ a width-depth fusion strategy to... Compared with the output of step 4 The features are concatenated along the channel dimension to obtain the fused features. : (twenty two) Preferably, step 6 specifically comprises: Step 6.1: The binary classification decision head outputs the probabilities of the two coarse categories for the current sample. and ; Step 6.2: The multi-class decision head outputs the probability distribution of the current sample's subcategories. ; Step 6.3: Perform hierarchical weighted correction, and perform weight fusion for categories that belong to the coarse category within the fine category: (twenty three) Finally, the corrected probability distribution Normalization is performed, and the final recognition result is output.

[0012] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described high-resolution one-dimensional distance image target recognition method.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described high-resolution one-dimensional distance image target recognition method.

[0014] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the above-described high-resolution one-dimensional distance image target recognition method.

[0015] The beneficial effects of this invention are as follows: This invention differs from directly using the original HRRP complex data. Instead, it utilizes the multidimensional feature space of the HRRP data and combines a "dynamic feature gating enhancement" mechanism with a hierarchical fusion decision strategy. By using coarse classification prior information to adaptively filter, weight, and combine features in a high-order manner, it achieves effective recognition of multiple target classes. Experimental results show that under the same simulation conditions, the average recognition accuracy of this method reaches 83.74%, which is approximately 4.54%, 1.92%, and 17.80% higher than the classic BiLSTM model (78.35%), the advanced MLP Mixer model (80.97%), and the TCNN model (70.80%), respectively. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2-1 This is the result of feature extraction for the average value; Figure 2-2 The results are the variance feature extraction results; Figure 2-3 This is the result of entropy feature extraction; Figure 2-4 The result is the third-order moment feature extraction result; Figure 2-5 The result is the fourth-order central moment feature extraction result; Figure 2-6 The result is the radial energy feature extraction result; Figure 2-7 The result is the super-average amplitude feature extraction result; Figure 2-8 The result is the extraction result of irregularity features; Figure 2-9 The results are from the descaled structural feature extraction. Figure 2-10 The result is the extraction of the dimensional features of the equivalent scattering center; Figure 2-11 The result is the extraction of the relative fullness feature of the time-domain signal; Figure 2-12 The result is the equivalent target size feature extraction result; Figure 2-13 The result is the radial dimension feature extraction result; Figure 3 To visualize the influence of features on the map; Figure 4 Training process loss and accuracy changes; Figure 5 The confusion matrix is ​​used to refine the identification results.

[0017] Figures 2-1 to 2-13 In the middle, all of (a) Target 1, (b) Target 2, (c) Target 3, (d) Target 4, (e) Target 5, and (f) are presented in a comprehensive manner. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] The purpose of this invention is to provide a high-resolution one-dimensional distance profile (HRRP) target recognition method based on hierarchical fusion and dynamic feature enhancement. Addressing the problems of one-sided feature utilization, susceptibility to environmental interference, and poor interpretability of physical features in deep networks, this invention constructs a multi-dimensional feature space and introduces a dynamic gating mechanism to achieve adaptive feature selection. Simultaneously, it innovatively designs a dual-head hierarchical architecture of "binary classification-fine classification," using coarse-grained interference / target discrimination information to logically constrain and probabilistically correct fine-grained category predictions, thereby significantly reducing the false alarm rate of interference samples and improving the model's recognition accuracy and robustness in complex electromagnetic environments.

[0020] The method described in this invention specifically includes the following steps: Step 1: Acquire the raw radar HRRP signal and preprocess it to extract a multi-dimensional physical feature vector containing statistical characteristics, geometric structure, and target size. Specifically, it is represented as follows: Step 1.1: Obtain the original radar HRRP complex signal, convert it into an amplitude spectrum using formula (1), and perform amplitude normalization using formula (2). Then, use the Cell Average Constant False Alarm Rate (CA-CFAR) detection algorithm to locate and capture the target area signal, removing background noise interference. The formula for converting the complex data into an amplitude spectrum is: (1) in, For amplitude spectrum, The real part of the original HRRP signal. This is the imaginary part of the original HRRP signal.

[0021] The formula for normalizing the amplitude spectrum is: (2) in, The amplitude spectrum of HRRP This is the normalized amplitude spectrum.

[0022] Step 1.2: Based on the preprocessed HRRP, extract target features from three aspects: statistical properties, geometric structure, and target size.

[0023] Statistical characteristics are parameters that define the statistical regularity and inherent fluctuation characteristics of a signal, and are directly related to the target's physical properties such as material, surface roughness, and electromagnetic properties. Specifically, they manifest as seven characteristics: mean, variance, entropy, radial energy, third moment, fourth central moment, and super-average amplitude.

[0024] The average value characteristic is used to describe the average intensity of scattered points in the HRRP signal, and the calculation formula is as follows: (3) Where N is the number of HRRP points after preprocessing. The target is the HRRP amplitude spectrum signal.

[0025] The formula for calculating variance characteristics is: (4) Entropy features can reflect the amplitude distribution of target scattering points in HRRP, and the calculation formula is as follows: (5) in, The HRRP amplitude spectrum signal is the target amplitude normalized.

[0026] Radial energy characteristics characterize the electromagnetic reflection capability of a target on radar, and the calculation formula is as follows: (6) The third-order moment feature measures the symmetry of the target HRRP data about the mean, and is calculated using the following formula: (7) The fourth-order central moment characteristic characterizes the tail thickness of HRRP data, and the calculation formula is as follows: (8) The super-average amplitude characteristic reflects the distribution deviation of the target scattering point amplitude. The larger the value, the more concentrated the amplitude distribution of the target scattering point. The calculation formula is: (9) Geometric structural features are parameters that describe the shape, structural complexity, or spatial distribution characteristics of a target, and are directly related to the target's geometric shape or the spatial layout of its scattering centers. These features include three types: irregularity, descaled structure, and the dimension of the equivalent scattering center.

[0027] The irregularity feature reflects the changes in the scattering points of nearby targets in HRRP, and embodies the structural changes of the target on the radar line of sight. The calculation formula is: (10) in, Indicates the first in the HRRP amplitude spectrum data Each element.

[0028] Descaled structural features are only related to the internal variations of the HRRP data, and are independent of its scaling transformation. The calculation formula is: (11) The equivalent scattering center number characteristic refers to the number of range cells exceeding the mean in the observed HRRP data. It is closely related to the number of target scattering centers, and the calculation formula is: (12) Target size characteristics are parameters that directly or indirectly describe the size of a target, reflecting its size range in a radial or equivalent sense. Specifically, they are manifested in three characteristics: relative fullness of the time-domain signal, equivalent target size, and radial size.

[0029] The relative fullness characteristic of a time-domain signal reflects the width of the target signal in the time domain, thus revealing the target's size information. The calculation formula is: (13) The equivalent target size characteristic reflects the width of the target signal in the time domain, embodying the target's size information. The calculation formula is: (14) The radial size feature describes the signal length occupied by the convergent scattering point in HRRP, and is directly related to the target size. The calculation formula is as follows: (15) in, Represents the set of target scattering points. This represents the slant distance of the scattering point.

[0030] Step 1.3: To avoid model misjudgment due to differences in units and orders of magnitude, and to optimize gradient flow, features obtained under the same target and scene are standardized. The calculation formula is as follows: (16) in, , representing the One characteristic.

[0031] After standardizing each feature, a multidimensional feature vector is formed. , as the original input of the network.

[0032] Step 2: Construct a target recognition network that includes a feature enhancement backbone module, a Transformer encoding module, and a dual-head hierarchical decision module. Specifically, this is manifested as follows: The feature enhancement backbone module consists of a deep cross-network (DCN), coarse guiding branches, dynamic feature gating units, cross-feature interaction units, and multi-head attention units cascaded together.

[0033] The Transformer encoding module consists of a feature embedding layer, a position encoding layer, and a multi-layer Transformer encoder.

[0034] The dual-head hierarchical decision module includes a binary decision head for distinguishing between "interference / target" and a multi-class decision head for fine-grained identification.

[0035] Step 3: Extract the multiple feature vectors obtained in Step 1. The input is fed into a deep cross-network (DCN) for high-order feature combination, and then fed into a coarse-guided branch to obtain the coarse-guided probability distribution of the current sample belonging to each category. Specifically, this is manifested as follows: First, the feature vector The DCN module performs feature crossover operations, outputting features that include higher-order combination information. Subsequently, Input to the coarse-guided branch, and output the coarse-guided class probability distribution for the current sample according to formula (17).

[0036] (17) in, and These are learnable weights and biases.

[0037] Step 4: The coarse-guided probability distribution obtained in Step 3 is adaptively weighted using a dynamic gating mechanism on the feature vector. Combined with subsequent interaction and attention mechanisms, this generates an enhanced feature vector. Specifically: Step 4.1: Add features The input dynamic feature gating unit, after multiple linear mappings, yields a dimension of... Feature importance map for each category ; Use the class probabilities obtained in step 3 right Perform weighted fusion to generate the importance vector of the current sample. .

[0038] (18) Next, the Sigmoid function is used to generate a feature-gated mask. And through a learnable scaling factor Intensity adjustment is performed by multiplying the original features element-wise to enhance the features, resulting in the gated features. : (19) Step 4.2, Input cross-feature interaction units, capture the nonlinear topological relationships between features through a multi-layer fully connected network (MLP), and output interactive features. Finally, residual connections and layer normalization are performed: (20) Step 4.3, Projecting to Q, K, V, the enhanced features are calculated. Finally, residual connections and layer normalization are performed. (twenty one) Step 5: Map the enhanced features obtained in Step 4 to a sequence input into the Transformer backbone for deep context modeling, and concatenate the deep semantic features output by Transformer with the high-order enhanced features output in Step 4 along the channel dimension to obtain the wide-depth fusion features.

[0039] Step 5.1: First, [the text abruptly ends here, likely due to an incomplete sentence or a formatting error. Mapped to a high-dimensional space and with added positional encoding, a token sequence is formed and input into the Transformer encoder to extract deep features. .

[0040] Step 5.2: Adopt a width-depth fusion strategy to... Compared with the output of step 4 The features are concatenated along the channel dimension to obtain the fused features. .

[0041] (twenty two) Step 6: Input the width and depth fusion features obtained in Step 5 into the binary classification decision head and the multi-class classification decision head respectively; use the interference / target probability distribution output by the binary classification decision head as a soft mask to perform hierarchical weighted correction on the output probability of the multi-class classification decision head, and calculate the final target recognition result.

[0042] Step 6.1: The binary classification decision head outputs the probabilities of the two coarse categories for the current sample. and .

[0043] Step 6.2: The multi-class decision head outputs the probability distribution of the current sample's subcategories.

[0044] Step 6.3: Perform hierarchical weighted correction. For categories belonging to the coarse category within the fine category, perform weight fusion: (twenty three) Finally, the corrected probability distribution Normalization is performed, and the final recognition result is output.

[0045] Example: This embodiment verifies the method based on simulated high-resolution range image data of static targets. The experimental dataset covers five classes (class 1 – class 5) of typical HRRP data, where class 1 and class 2 belong to the "interference" category, and classes 3, 4, and 5 belong to the "target" category. Combining the radar platform's airborne motion parameters and spatiotemporal observation conditions, HRRP observation data for different scenarios are constructed, with 16,968 data sets in each class, totaling 84,840 data sets. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Feature extraction is performed using Matlab 2025a, and the recognition environment is Python 3.9.19, PyTorch 1.12.2, with a batch size of 32. The optimizer is Adam, and the system uses an NVIDIA GeForce RTX 4060 8G processor.

[0046] Step 1: Acquire raw radar HRRP data, perform data preprocessing, and extract multi-dimensional feature vectors containing statistical features, geometric structure, and target size. Specifically, the implementation is as follows: Step 1.1: Convert the acquired raw radar HRRP complex data signal into an amplitude spectrum according to formula (1), then normalize the amplitude spectrum according to formula (2), and finally use CA-CFAR detection to obtain the target area.

[0047] (1) The formula for normalizing the amplitude spectrum is: (2) in, The amplitude spectrum of HRRP This is the normalized amplitude spectrum.

[0048] Step 1.2: Based on the preprocessed HRRP amplitude signal, 13 target features with clear physical meanings are extracted from three aspects: statistical characteristics, geometric structure, and target size. Examples of feature extraction results are shown below. Figures 2-1 to 2-13 .

[0049] The statistical characteristics are as follows: the mean, variance, entropy, radial energy, third moment, fourth central moment and super-average amplitude characteristics are calculated according to (3)-(9).

[0050] (3) (4) (5) (6) (7) (8) (9) The geometric structure features are specifically calculated based on (10)-(12) to determine the irregularity, descaled structure, and equivalent scattering center dimension features.

[0051] (10) (11) (12) Specifically, the target size features are calculated based on (13)-(15) to determine the relative fullness of the time-domain signal, the equivalent target size, and the radial size features.

[0052] (13) (14) (15) Step 1.3: Perform feature standardization, outputting a 13-dimensional feature set. This is to avoid model misjudgment caused by differences in units and orders of magnitude, and to optimize gradient flow. Taking one feature as an example, the features obtained under the same target and the same scene are standardized according to formula (16).

[0053] (16) Step 2: Construct a target recognition network that includes a feature enhancement backbone module, a Transformer encoding module, and a dual-head hierarchical decision module. Specifically, this is manifested as follows: The aforementioned meta- and cross-feature interaction units and multi-head attention units are cascaded together. The DCN consists of 5 layers of cross-network, the coarse guiding branch consists of two fully connected layers, the cross-feature interaction consists of 3 layers of MLP, and the number of multi-head attention units is 8.

[0054] The Transformer encoding module consists of a feature embedding layer, a positional encoding layer, and a multi-layer Transformer encoder. It contains a total of 4 encoding layers, with a feature embedding dimension of 126 and a hidden layer dimension of 256.

[0055] The dual-head hierarchical decision module includes a binary decision head for distinguishing between "interference / target" and a five-class decision head for fine-grained identification.

[0056] In all the above structures, a random inactivation probability of 0.3 is set to prevent overfitting.

[0057] Step 3: Extract the multiple feature vectors obtained in Step 1. The input is fed into a deep cross-network (DCN) for high-order feature combination, and then fed into a coarse-guided branch to obtain the coarse-guided probability distribution of the current sample belonging to each category. Specifically, this is manifested as follows: First, the feature vector The DCN module performs feature crossover operations, outputting features that include higher-order combination information. Specifically, the formulas for each DCN layer are as follows: (16) After five iterations, the output features include higher-order combination information. .

[0058] Then Input to the coarse-guided branch, and output the coarse-guided class probability distribution for the current sample according to formula (17).

[0059] (17) in, and These are learnable weights and biases. This represents the preliminary predicted probability that the current sample belongs to one of the five subcategories. This probability is not used as the final result, but rather as prior information to guide feature enhancement.

[0060] Step 4: The coarse guided probability distribution obtained in Step 3 is adaptively weighted using a dynamic gating mechanism on the feature vector. Combined with subsequent cross-feature interaction and attention mechanisms, an enhanced feature vector is generated. Specifically: Step 4.1: Add features The input dynamic feature gating unit, after multiple linear mappings, yields a dimension of... Feature importance map for each category Feature maps are attached. Figure 3 As shown; using the class probabilities obtained in step 3 right Perform weighted fusion to generate the importance vector of the current sample. .

[0061] (18) Next, the Sigmoid function is used to generate a feature-gated mask. And through a learnable scaling factor Intensity adjustment is performed, with the scaling factor initially set to 1.0. Feature enhancement is achieved by element-wise multiplication of the original features, resulting in the gated features. : (19) Step 4.2, Input cross-feature interaction units, capture the nonlinear topological relationships between features through a 3-layer fully connected network (MLP), and output interactive features. Finally, residual connections and layer normalization are performed: (20) Step 4.3, Projecting to Q, K, V, the enhanced features are calculated. Finally, residual connections and layer normalization are performed. (twenty one) Step 5: Map the enhanced features obtained in Step 4 to a sequence input into the Transformer backbone for deep context modeling, and concatenate the deep semantic features output by Transformer with the high-order enhanced features output in Step 4 along the channel dimension to obtain the wide-depth fusion features.

[0062] Step 5.1: First, [the text abruptly ends here, likely due to an incomplete sentence or a formatting error. Mapped to a high-dimensional space and with added positional encoding, a token sequence is formed and input into the Transformer encoder to extract deep features. .

[0063] Step 5.2: Adopt a width-depth fusion strategy to... Compared with the output of step 4 The features are concatenated along the channel dimension to obtain the fused features. .

[0064] (twenty two) Step 6: Input the width and depth fusion features obtained in Step 5 into the binary classification decision head and the multi-class classification decision head respectively; use the interference / target probability distribution output by the binary classification decision head as a soft mask to perform hierarchical weighted correction on the output probability of the multi-class classification decision head, and calculate the final target recognition result.

[0065] Step 6.1: The binary classification decision head outputs the probability that the current sample belongs to the "target" or "interference" category.

[0066] (twenty three) in To identify the probability as "interference", The probability of identifying it as a "target".

[0067] Step 6.2: The five-class decision head outputs the probability distribution of the current sample's subcategories.

[0068] Step 6.3: Perform hierarchical weighted adjustment. For categories belonging to the coarse category within the fine category, perform weight adjustment. The adjustment logic is as follows: (twenty four) (25) Finally, the corrected probability distribution is normalized to output the final recognition result.

[0069] (26) The loss versus accuracy curves during network training are as follows: Figure 4 As shown. The confusion matrix of the test set recognition results is as follows. Figure 5 As shown in Table 1, the recognition performance of the proposed method TF_automask is compared with that of the baseline method Transformer, the classic model BiLSTM, the advanced model MLP Mixer, and TCNN, demonstrating the effectiveness of the proposed method.

Claims

1. A high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement, characterized in that, Includes the following steps: Step 1: Acquire the raw radar HRRP signal and preprocess it to extract a multi-dimensional physical feature vector containing statistical characteristics, geometric structure and target size; Step 2: Construct a target recognition network that includes a feature enhancement backbone module, a Transformer encoding module, and a dual-head hierarchical decision module; Step 3: Input the multidimensional feature vectors extracted in Step 1 into the Deep Cross Network (DCN) for high-order feature combination, and input them into the coarse guiding branch to obtain the coarse guiding probability distribution of the current sample belonging to each category; Step 4: The coarse guided probability distribution obtained in Step 3 is used to adaptively weight the feature vector using a dynamic gating mechanism, and combined with subsequent interaction and attention mechanisms to generate an enhanced feature vector. Step 5: Map the enhanced features obtained in Step 4 into a sequence input to the Transformer backbone for deep context modeling, and concatenate the deep semantic features output by Transformer with the high-order enhanced features output in Step 4 along the channel dimension to obtain the wide-depth fusion features. Step 6: Input the width and depth fusion features obtained in Step 5 into the binary classification decision head and the multi-class classification decision head, respectively; The interference / target probability distribution output by the binary classification decision head is used as a soft mask to perform hierarchical weighted correction on the output probability of the multi-class decision head, and the final target recognition result is calculated.

2. The high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1: Obtain the original radar HRRP complex signal, convert it into an amplitude spectrum using formula (1), and perform amplitude normalization using formula (2); then, use the cell average constant false alarm rate detection algorithm to locate and capture the target area signal and remove background noise interference; the formula for converting complex data into an amplitude spectrum is: (1) in, For amplitude spectrum, The real part of the original HRRP signal. This represents the imaginary part of the original HRRP signal. The formula for normalizing the amplitude spectrum is: (2) in, The amplitude spectrum of HRRP The normalized amplitude spectrum; Step 1.2: Based on the preprocessed HRRP, extract target features from three aspects: statistical properties, geometric structure, and target size; The statistical characteristics are parameters of the statistical regularity and inherent fluctuation characteristics of the signal, including seven features: mean, variance, entropy, radial energy, third moment, fourth central moment, and super-average amplitude. The average value characteristic is used to describe the average intensity of scattered points in the HRRP signal, and the calculation formula is as follows: (3) Where N is the number of HRRP points after preprocessing. The HRRP amplitude spectrum signal is the target. The formula for calculating variance characteristics is: (4) Entropy features reflect the amplitude distribution of target scattering points in HRRP, and the calculation formula is: (5) in, The HRRP amplitude spectrum signal after normalization of the target amplitude; Radial energy characteristics characterize the electromagnetic reflection capability of a target on radar, and the calculation formula is as follows: (6) The third-order moment feature measures the symmetry of the target HRRP data about the mean, and is calculated using the following formula: (7) in, This indicates the location of the target HRRP data distribution center; The fourth-order central moment characteristic characterizes the tail thickness of HRRP data, and the calculation formula is as follows: (8) The formula for calculating the super-average amplitude characteristic is: (9) The geometric structure refers to parameters that describe the target's shape, structural complexity, or spatial distribution characteristics, including three features: irregularity, descaled structure, and equivalent scattering center dimension. The formula for calculating the irregularity feature is: (10) in, , , These represent the first [number] in the HRRP amplitude spectrum data. -1 element, the first The element, the first +1 element; The formula for calculating descaled structural features is: (11) The formula for calculating the equivalent scattering center number characteristic is: (12) in, Represents the unit step function; The target size is a parameter that directly or indirectly describes the size of the target, including three features: relative fullness of the time-domain signal, equivalent target size, and radial size. The formula for calculating the relative fullness characteristic of a time-domain signal is: (13) in, Indicates the first In the nth sample The amplitude value of each sampling point Indicates the first The maximum value of the amplitude of each sample. Indicates the number of sampling points. Indicates the total number of samples; The formula for calculating the equivalent target size feature is: (14) in, This indicates that the echo intensity is greater than the threshold. The distance cell indices are arranged in ascending order; This represents the first index of the sequence. Indicates the last index of the sequence; The formula for calculating radial dimension features is: (15) in, Represents the set of target scattering points; Indicates the first The slant distance of each scattering point ; Step 1.3: Standardize the features obtained for the same target and in the same scene. The calculation formula is as follows: (16) in, , representing the One feature; Indicates the first The first sample The standardized values ​​of each feature; After standardizing each feature, a multidimensional feature vector is formed. , as the original input of the network.

3. The high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement according to claim 2, characterized in that, Step 2 specifically involves: The feature enhancement backbone module is composed of a deep cross network (DCN), a coarse guiding branch, a dynamic feature gating unit, a cross-feature interaction unit, and a multi-head attention unit cascaded together. The Transformer encoding module consists of a feature embedding layer, a position encoding layer, and a multi-layer Transformer encoder. The dual-head hierarchical decision module includes a binary classification decision head for distinguishing between "interference / target" and a multi-classification decision head for fine-grained identification.

4. The high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement according to claim 3, characterized in that, Step 3 specifically involves: First, the feature vector The DCN module performs feature crossover operations, outputting features that include higher-order combination information. ; Then Input to the coarse-guided branch, and output the coarse-guided class probability distribution for the current sample according to formula (17). : (17) in, and These are learnable weights and biases.

5. The high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement according to claim 4, characterized in that, Step 4 specifically involves: Step 4.1: Add features The input dynamic feature gating unit, after multiple linear mappings, yields a dimension of... Feature importance map for each category ; Using the class probabilities obtained in step 3 right Perform weighted fusion to generate the importance vector of the current sample. : (18) Generating feature-gated masks using the Sigmoid function And through a learnable scaling factor Intensity adjustment is performed by multiplying the original features element-wise to enhance the features, resulting in the gated features. : (19) in, This represents the Sigmoid activation function. Presentation layer normalization processing, This indicates element-wise multiplication. This represents the learnable scaling factor; Step 4.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input cross-feature interaction units, capture the nonlinear topological relationships between features through a multi-layer fully connected network (MLP), and output interactive features. Finally, residual connections and layer normalization are performed: (20) Step 4.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Projecting Q, K, and V, the enhanced features are calculated. Finally, residual connections and layer normalization are performed. (21) in, The dimension of the key vector is represented by Q, where Q represents the query matrix, K represents the key matrix, and V represents the value matrix. This represents the feature output after attention weighting. This represents the final feature after residual connection and layer normalization.

6. The high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1: First, [the text abruptly ends here, likely due to an incomplete sentence or a formatting error. Mapped to a high-dimensional space and with added positional encoding, a token sequence is formed and input into the Transformer encoder to extract deep features. ; Step 5.2: Employ a width-depth fusion strategy to... Compared with the output of step 4 The features are concatenated along the channel dimension to obtain the fused features. : (22)。 7. The high-resolution one-dimensional range image target recognition method based on hierarchical fusion and dynamic feature enhancement according to claim 6, characterized in that, Step 6 specifically involves: Step 6.1: The binary classification decision head outputs the probabilities of the two coarse categories for the current sample. and ; Step 6.2: The multi-class decision head outputs the probability distribution of the current sample's subcategories. ; Step 6.3: Perform hierarchical weighted correction, and perform weight fusion for categories that belong to the coarse category within the fine category: (23) Finally, the corrected probability distribution Normalization is performed, and the final recognition result is output.

8. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

10. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 7.

Images