Target material classification method based on time-frequency dual-path attention fusion network
By using a time-frequency dual-path attention fusion network, the problem of insufficient accuracy of traditional underwater target classification methods in complex underwater environments is solved, achieving high-precision classification of underwater target materials and improving the robustness and real-time performance of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NAVAL UNIV OF ENG PLA
- Filing Date
- 2026-04-21
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional underwater target classification methods struggle to achieve high-precision classification in complex underwater environments. They are affected by water scattering, background noise, and interference from multiple targets. Furthermore, pseudo-color mapping alters the grayscale distribution characteristics of the original echo signal, making it difficult to distinguish targets with similar shapes but different materials.
A target material classification method based on the time-frequency dual-path attention fusion network (TF-CTAM-InceptionTime) is adopted. Through complementary extraction of time-frequency dual-path features, adaptive channel-time dual attention mechanism, and residual enhancement multi-scale Inception, high-precision classification of raw echo data is achieved.
It significantly reduced the misclassification rate, improved the robustness and real-time performance of the system, and effectively distinguished targets with similar shapes but different materials, thus improving classification accuracy.
Smart Images
Figure CN122157699A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of sonar identification of underwater targets, specifically involving a target material classification method based on a time-frequency dual-path attention fusion network. Background Technology
[0002] Underwater target classification is one of the core technologies in marine resource exploration, underwater security monitoring, and anti-submarine early warning systems, and its accuracy directly determines the efficiency and safety of underwater operations. For a long time, traditional underwater target classification methods have been based on sonar detection technology, relying on two-dimensional images generated by sonar equipment to complete target identification and classification tasks. This approach has been widely used in the field of underwater detection due to its simple implementation logic and strong hardware adaptability.
[0003] In traditional technical processes, sonar equipment transmits sound waves and receives reflected echoes from targets, converting the acoustic signals into two-dimensional grayscale or pseudo-color images. Image processing algorithms are then used to segment, extract features, and classify the targets. This image-level processing approach essentially simplifies complex underwater acoustic signals into visually interpretable two-dimensional planar information, attempting to differentiate underwater targets by simulating the logic of visual recognition.
[0004] However, such methods have significant technical limitations and are difficult to meet the classification accuracy requirements of complex underwater environments. First, underwater sonar echo signals are easily affected by water scattering, background noise, and multi-target interference. To improve image readability, traditional methods need to suppress interference signals through noise filtering algorithms. However, the filtering process often results in the loss of weak features in the original echo signal, which are crucial for distinguishing target materials and structural details. Second, pseudo-color mapping, as a common method for image visualization, can enhance the visual contrast of images, but it alters the grayscale distribution characteristics of the original echo signal, disrupting the correlation between the signal and the target material and surface roughness. This makes it difficult for the algorithm to distinguish targets with similar shapes but different materials, such as metal targets and ceramic targets, or solid targets and hollow targets. Summary of the Invention
[0005] The target material classification method based on the time-frequency dual-path attention fusion network (TF-CTAM-InceptionTime) provided by this invention achieves high-precision classification of raw echo data through time-frequency dual-path feature complementary extraction (TF), adaptive channel-time dual attention mechanism (CTAM) and residual-enhanced multi-scale Inception, significantly reducing the misclassification rate and improving the robustness and real-time performance of the system.
[0006] The target material classification method based on a time-frequency dual-path attention fusion network provided by this invention includes:
[0007] The data acquisition module acquires raw echo data from the image sonar system.
[0008] The preprocessing module preprocesses the raw callback data;
[0009] And the TF-CTAM-InceptionTime model, which processes the preprocessed input data and outputs the results, including:
[0010] The sampling and dataset partitioning module divides the dataset into training, validation, and test sets based on the sampling strategy.
[0011] The time-frequency dual-path feature extraction module operates in parallel with the time-domain path and the frequency path, and then fuses the outputs of the time-domain path and the frequency path.
[0012] An adaptive channel-temporal dual attention module is provided, which includes channel attention branches and temporal attention branches, and fuses the outputs of the channel attention branches and temporal attention branches.
[0013] The residual enhancement Inception module introduces the Bottleneck layer and residual connections, stacking multiple layers to form hierarchical features;
[0014] The classification head module outputs the class probability distribution through global average pooling, Dropout, and fully connected layers.
[0015] As a further optimization of the present invention, the data acquisition module employs raw echo data acquired through synthetic aperture sonar, side-scan sonar, and forward-looking multibeam sonar systems.
[0016] As a further optimization of the present invention, the preprocessing module includes the following steps:
[0017] Calculate the signal mean and remove it;
[0018] Suppress direct wave interference from the transmitted pulse while preserving the target echo signal;
[0019] Filter out environmental noise, system noise and high-frequency interference outside the signal frequency band, and retain the effective echo signal within the sonar operating frequency band;
[0020] Generate a linear frequency modulation reference signal that matches the transmitted signal, compress the pulse, and improve the signal-to-noise ratio;
[0021] Normalization is performed based on the characteristics of the data.
[0022] As a further optimization of the present invention, the time-frequency dual-path feature extraction module includes:
[0023] Temporal path extraction includes four parallel convolutional branches with kernel sizes of k=1, k=3, k=5, and k=7. The k=1 point convolution is used for inter-channel information integration, k=3 captures local impulse features, k=5 extracts mesoscale temporal patterns, and k=7 obtains macroscopic temporal trends. The outputs of the four branches are concatenated in the channel dimension and then batch normalized and activated by the ReLU function.
[0024] Frequency domain path extraction involves performing a Fast Fourier Transform (FFT) on the input signal to obtain the amplitude and phase spectra. Then, in the frequency domain, multi-scale one-dimensional convolution branches are applied to extract spectral features at different frequency scales in parallel. This includes four parallel convolution branches with kernel sizes of k'=1, k'=3, k'=5, and k'=7, all with a stride of 1 and a "same" padding method. The k'=1 point convolution is used for information integration between frequency channels, k'=3 captures local frequency band features, k'=5 extracts mid-frequency band patterns, and k'=7 obtains wideband response trends. The outputs of the four branches are concatenated in the channel dimension, followed by batch normalization and a ReLU activation function to capture frequency response characteristics related to the target material.
[0025] Time-frequency fusion outputs the time-domain path. and frequency domain path output The features are concatenated along the feature dimension to form time-frequency features. .
[0026] As a further optimization of the present invention, the adaptive channel-temporal dual attention module includes:
[0027] The channel attention branch employs a Squeeze-and-Excitation structure. Features are compressed into channel descriptors through global average pooling, then passed through two fully connected layers. The first layer compresses the number of channels from C to C / r, and after ReLU activation, the second layer restores it to C channels. Finally, the channel attention weights are generated using the Sigmoid function. ;
[0028] The temporal attention branch uses a 1×1 convolutional layer to map multi-channel features into a single-channel attention map, and the temporal attention weights are obtained after Softmax normalization. ;
[0029] The channel attention branch and the temporal attention branch are fused by element-wise multiplication to generate joint attention weights. Output .
[0030] As a further optimization of the present invention, the residual enhancement Inception module includes:
[0031] Bottleneck layer: Add a convolutional bottleneck layer at the entrance of each Inception module to compress the number of input channels;
[0032] Multi-scale Inception branch: Parallel branches with different convolutional kernel sizes are used to capture multi-granularity temporal features;
[0033] Residual connections: Skip connections are introduced in each Inception module to directly add the input to the output. When the number of input and output channels is inconsistent, dimension matching is performed through convolution.
[0034] Regularization strategy: Add a Dropout layer after each Inception module and use batch normalization to stabilize the training process.
[0035] As a further optimization of the present invention, a strategy optimization module is also included to optimize the category probability distribution output by the classification head module, including the following steps:
[0036] We employ the cross-entropy loss function to introduce class weighting for the class imbalance problem;
[0037] The AdamW optimizer was used, and the initial learning rate and weight decay coefficient were set.
[0038] The learning rate is dynamically adjusted using a cosine annealing strategy.
[0039] An early stopping mechanism is adopted to monitor the loss of the validation set, and training is terminated early if there is no improvement after several consecutive rounds.
[0040] The target material classification method based on a time-frequency dual-path attention fusion network provided by this invention combines time-domain multi-scale convolution with frequency-domain FFT feature extraction for sonar target classification, making full use of the complementary information of time-domain waveform details and frequency-domain spectral structure; it adopts a dual attention mechanism to adaptively focus on key frequency components and important moments of sonar echoes, enhancing the ability to distinguish target materials; it introduces residual connections and Bottleneck structures on the basis of InceptionTime to improve the training stability of deep networks and accelerate convergence; it avoids information loss caused by image-level processing, retains complete amplitude and phase information, and achieves efficient recognition and classification of target shape and material. Attached Figure Description
[0041] Figure 1 This is an architecture diagram of this embodiment;
[0042] Figure 2 It is a training curve graph, showing the convergence process of the loss function and accuracy as the number of training epochs increases;
[0043] Figure 3 It is a confusion matrix diagram that shows the classification performance and misclassification of each category. Detailed Implementation
[0044] like Figure 1 As shown, this embodiment includes a data acquisition module, a preprocessing module, and a TF-CTAM-InceptionTime model.
[0045] The data acquisition module acquires raw echo data from the image-based sonar system as input.
[0046] The acquisition of raw echo data includes acquisition through synthetic aperture sonar, side-scan sonar or forward-looking multibeam sonar systems. Multi-channel acquisition configuration can be used. The acquired raw echo data retains complete amplitude and phase information to ensure the integrity of information in subsequent processing.
[0047] Specifically, the dataset in this embodiment contains five types of targets: solid polyurethane spheres, hollow aluminum spheres, fiberboard in the shape of the letter O, fiberboard in the shape of the letter Q, and background negative samples.
[0048] The dataset is sampled using an image sonar system to obtain raw echo data samples. Each sample contains approximately 1000 echo sequences with a sequence length of L=3000 and a sampling rate of fs=100kHz.
[0049] The preprocessing module preprocesses the acquired raw echo data to extract effective features and eliminate interference. Specifically, it includes steps such as removing DC bias, suppressing direct waves, bandpass filtering, filter matching, and normalization.
[0050] The step of removing DC bias is as follows: Removing DC bias eliminates the influence of DC components on subsequent processing. The specific calculation formula is s(t)=x(t)-mean(x(t)), where x(t) is the value of the original echo data, mean(x(t)) is the mean of the original echo data, and s(t) is the value after removing DC bias.
[0051] One step involved suppressing direct wave interference. Specifically, a combination of Tukey windowing and time-domain masking was used to suppress direct wave interference from the transmitted pulse while preserving the target echo signal.
[0052] Direct waves reach the receiver directly without encountering any target, thus exhibiting strong signal strength and short latency, while echoes are characterized by weak signals and long latency.
[0053] Based on this characteristic, the time-domain mask distinguishes signals according to time, marking the direct wave interference zone (e.g., 0-0.1 seconds) and the target echo retention zone (e.g., 0.2-1 seconds) according to time, and only locks the interference zone for processing.
[0054] The Tukey window works by multiplying the signal by a coefficient. A small coefficient (0) suppresses the signal strength; a coefficient of 1 keeps the signal unchanged. The coefficient gradually decreases from 1 to 0 and then gradually increases back to 1. In other words, during the direct wave period of 0-0.1 seconds, the Tukey window's smoothing coefficient gradually suppresses the strong signal of the direct wave. After 0.1 seconds, the coefficient returns to 1. The target echo region of 0.2-1 seconds is left unprocessed, and the signal is preserved as is.
[0055] The bandpass filtering step involves designing a Butterworth bandpass filter with upper and lower cutoff frequencies fl and fh set according to the sonar operating frequency. This filter removes environmental noise, system noise, and high-frequency interference outside the signal band, while retaining the effective echo signal within the sonar operating frequency band.
[0056] The filtering and matching step involves generating a linear frequency modulated (LFM) reference signal that matches the transmitted signal. This LFM reference signal is then compared point-by-point with the received echo signal. This comparison produces two results: signal amplification and noise suppression. This amplification and suppression improve the signal-to-noise ratio.
[0057] The output includes three modes: envelope, real part, and complete complex form.
[0058] The envelope extraction method retains only the intensity profile of the signal, simplifying the complex complex signal and phase changes, ultimately resulting in an intuitive intensity curve. The peak of the curve represents the target's position / distance. The real part extraction method breaks down the complete complex signal into real and imaginary parts, extracting only the real part for output and removing the imaginary part. The complete complex form method does not simplify in any way, retaining all the information of the pulse-compressed signal, including intensity, phase, real part, and imaginary part.
[0059] The normalization process involves selecting a normalization method based on the data characteristics. This includes Z-Score normalization, Min-Max normalization, and sample-level normalization, as well as independent normalization for each individual sample, which is suitable for data acquired with different gains.
[0060] The TF-CTAM-InceptionTime model processes the preprocessed input data and outputs the results. The TF-CTAM-InceptionTime model includes a sampling and dataset partitioning module, a time-frequency dual-path feature extraction module, an adaptive channel-temporal dual attention module, a residual enhancement Inception module, and a classification head module.
[0061] The sampling and dataset partitioning module divides the dataset into training, validation, and test sets based on a sampling strategy. The sampling strategy, such as random sampling, intermediate sampling, or full sampling, is selected according to task requirements.
[0062] In this embodiment, the LFM signal parameters are: bandwidth B = 20kHz, pulse width T = 0.5ms. Data preprocessing in this embodiment employs sample-level normalization, dividing the training, validation, and test sets in an 8:1:1 ratio. Introducing negative samples is supported to improve the model's robustness against interference.
[0063] The time-frequency dual-path feature extraction module operates in parallel with the time-domain path and the frequency path, and then fuses the outputs of the time-domain path and the frequency path.
[0064] The process extracts temporal path features by employing multi-scale one-dimensional convolutional branches to extract echo features at different time scales in parallel. Specifically, it includes four parallel convolutional branches with kernel sizes of k=1, k=3, k=5, and k=7, all with a stride of 1 and the same padding method. The k=1 pointwise convolution is used for inter-channel information integration, k=3 captures local impulse features, k=5 extracts mid-scale temporal patterns, and k=7 obtains macro-temporal trends. The outputs of the four branches are concatenated along the channel dimension, followed by batch normalization (BN) and the ReLU activation function.
[0065] To extract frequency path features, the input signal is first subjected to a Fast Fourier Transform (FFT) to transform the time-domain signal into the frequency domain, obtaining the amplitude spectrum |X(f)| and phase spectrum ∠X(f). Then, multi-scale one-dimensional convolution branches are applied in the frequency domain to extract spectral features at different frequency scales in parallel. Specifically, this includes four parallel convolution branches with kernel sizes k'=1, k'=3, k'=5, and k'=7, all with a stride of 1 and the same padding method. The k'=1 point convolution is used for information integration between frequency channels, k'=3 captures local frequency band features, k'=5 extracts mid-frequency patterns, and k'=7 obtains wideband response trends. The outputs of the four branches are concatenated along the channel dimension, followed by batch normalization (BN) and ReLU activation functions to capture frequency response characteristics related to the target material. The frequency domain features are then fused with the time-domain features after an inverse transform.
[0066] The time-domain path features and frequency path features are fused together, and the time-domain path output is obtained. and frequency domain path output The features are spliced and fused along the feature dimension to form a comprehensive time-frequency feature. This fully utilizes the waveform details in the time domain and the spectral structure information in the frequency domain.
[0067] The adaptive channel-temporal dual attention module includes channel attention branches and temporal attention branches, and fuses the outputs of the channel attention branches and temporal attention branches.
[0068] Key features are enhanced through channel attention and temporal attention, and channel attention and temporal attention are fused together.
[0069] Channel attention branch. A Squeeze-and-Excitation structure is adopted. First, global average pooling compresses the features into channel descriptors, then two fully connected layers learn the dependencies between channels. Let the shape of the input feature map be (B, C, L), where B is the batch size, C is the number of channels (the total number of feature channels after time-frequency dual-path fusion), and L is the number of sampling points. The first fully connected layer compresses the number of channels from C to C / r (compression ratio r=16). After ReLU activation, the second layer restores it to C channels. Finally, the channel attention weights are generated using the Sigmoid function. ,in This represents a C-dimensional vector, where each element has a value between 0 and 1, representing the importance weight of the corresponding channel.
[0070] Temporal attention branch. Designed specifically for the temporal characteristics of sonar echoes, attention weights are learned along the time dimension. Specifically, a 1×1 convolutional layer maps multi-channel features to a single-channel attention map, and the temporal attention weights are obtained after Softmax normalization. Where L is the number of sampling points, This represents an L-dimensional vector, where each element has a value between 0 and 1, representing the importance weight of the corresponding moment. This weight highlights key moments in the echo signal (such as the target reflection peak point) and suppresses background noise and irrelevant time periods.
[0071] (3) Dual Attention Fusion. Channel attention and temporal attention are fused through a broadcast mechanism and element-wise multiplication. First, the channel attention weights are... (Shape C×1) and temporal attention weights Perform an outer product operation (ⓧ represents the outer product) on the matrix of shape 1×L to generate the joint attention weight matrix. Its shape is C×L, representing the combined attention weight of each channel at each time step. The output is finally calculated using element-wise multiplication (⊙ represents element-wise multiplication, i.e., the Hadamard product). ,in To achieve dual refinement and enhancement of features for time-frequency fusion, the network can simultaneously focus on important frequency channels and key time points.
[0072] The Residual Augmentation Inception module introduces Bottleneck layers and residual connections, stacking multiple layers to form hierarchical features, thereby improving the network's feature representation ability and training stability. Specifically, the Residual Augmentation Inception module includes:
[0073] Bottleneck layer: Add a 1×1 convolutional bottleneck layer at the entrance of each Inception module to compress the number of input channels to 1 / 4 of the original, reducing computational complexity while enhancing non-linear expressive power.
[0074] Multi-scale Inception branch: Retains the multi-scale convolution design of the original InceptionTime network model, including parallel branches with different convolution kernel sizes (10, 20, 40) to capture multi-granular temporal features.
[0075] Residual connections: Skip connections are introduced in each Inception module to directly add the input to the output, alleviating the vanishing gradient problem in deep networks and accelerating convergence. When the number of input and output channels is inconsistent, 1×1 convolutions are used for dimension matching.
[0076] Regularization strategy: Add a Dropout layer after each Inception module to prevent overfitting, and use batch normalization to stabilize the training process.
[0077] In this embodiment, the network stacks N=6 Inception modules, with the number of convolutional kernels increasing layer by layer. The first two Inception modules use 32 kernels, the middle two Inception modules use 64 kernels, and the last two Inception modules use 128 kernels, thus forming a hierarchical feature representation.
[0078] The classification head module outputs the class probability distribution through global average pooling, Dropout, and fully connected layers.
[0079] After the output of the final residual enhancement Inception module, the feature sequence is compressed into a fixed-length vector through global average pooling. Then, it is regularized through a Dropout layer, and finally mapped to the number of classes through a fully connected layer. The Softmax function outputs the probability distribution of each class.
[0080] Furthermore, it also includes a strategy optimization module, which optimizes the category probability distribution output by the classification head module, specifically including the following steps:
[0081] Cross-entropy loss function is used to introduce class weighting for class imbalance; AdamW optimizer is used with an initial learning rate and weight decay coefficient set, specifically the initial learning rate is set to 1×10^(-3) and the weight decay coefficient is set to 1×10^(-4); cosine annealing strategy is used to dynamically adjust the learning rate; an early stopping mechanism is used to monitor the validation set loss and terminate training early if there is no improvement after several consecutive rounds, specifically if there is no improvement after 10 consecutive rounds. The training batch size is set to 64 and the maximum number of training rounds is 100.
[0082] The trained model is used to perform inference and classification on the preprocessed echo data.
[0083] The trained TF-CTAM-InceptionTime model is used to perform inference on preprocessed echo data to achieve joint classification of underwater target types (such as letter shapes O / Q, spheres, etc.) and materials (such as polyurethane, aluminum, steel, etc.). The output includes class labels and corresponding confidence probabilities.
[0084] like Figure 2 As shown in the training curve, the loss decreases rapidly in the initial iteration phase and drops below 0.05 after about 2500 iterations. The accuracy stabilizes above 98% after 25 rounds, verifying the effective convergence of the training process.
[0085] like Figure 3 As shown in the confusion matrix results, the classification accuracy for target type and material reached 99.2%, while the misclassification rate of negative background samples was less than 6%. In particular, for solid spheres (polyurethane) and hollow spheres (aluminum), which are the same shape but different materials, the classification accuracy improved from 92.8% in the original InceptionTime to 100%, proving that the CTAM attention mechanism effectively enhances material features.
[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A target material classification method based on a time-frequency dual-path attention fusion network, characterized in that, include: The data acquisition module acquires raw echo data from the image sonar system. The preprocessing module preprocesses the raw callback data; And the TF-CTAM-InceptionTime model, which processes the preprocessed input data and outputs the results, including: The sampling and dataset partitioning module divides the dataset into training, validation, and test sets based on the sampling strategy. The time-frequency dual-path feature extraction module operates in parallel with the time-domain path and the frequency path, and then fuses the outputs of the time-domain path and the frequency path. An adaptive channel-temporal dual attention module is provided, which includes channel attention branches and temporal attention branches, and fuses the outputs of the channel attention branches and temporal attention branches. The residual enhancement Inception module introduces the Bottleneck layer and residual connections, stacking multiple layers to form hierarchical features; The classification head module outputs the class probability distribution through global average pooling, Dropout, and fully connected layers.
2. The target material classification method based on a time-frequency dual-path attention fusion network according to claim 1, characterized in that, The data acquisition module uses synthetic aperture sonar, side-scan sonar and forward-looking multibeam sonar systems to collect raw echo data.
3. The target material classification method based on a time-frequency dual-path attention fusion network according to claim 1, characterized in that, The preprocessing module includes the following steps: Calculate the signal mean and remove it; Suppress direct wave interference from the transmitted pulse while preserving the target echo signal; Filter out environmental noise, system noise and high-frequency interference outside the signal frequency band, and retain the effective echo signal within the sonar operating frequency band; Generate a linear frequency modulation reference signal that matches the transmitted signal, compress the pulse, and improve the signal-to-noise ratio; Normalization is performed based on the characteristics of the data.
4. The target material classification method based on a time-frequency dual-path attention fusion network according to claim 1, characterized in that, The time-frequency dual-path feature extraction module includes: Temporal path extraction includes four parallel convolutional branches with kernel sizes of k=1, k=3, k=5, and k=7. The k=1 point convolution is used for inter-channel information integration, k=3 captures local impulse features, k=5 extracts mesoscale temporal patterns, and k=7 obtains macroscopic temporal trends. The outputs of the four branches are concatenated in the channel dimension and then batch normalized and activated by the ReLU function. Frequency domain path extraction involves performing a Fast Fourier Transform (FFT) on the input signal to obtain the amplitude and phase spectra. Then, in the frequency domain, multi-scale one-dimensional convolution branches are applied to extract spectral features at different frequency scales in parallel. This includes four parallel convolution branches with kernel sizes of k'=1, k'=3, k'=5, and k'=7, all with a stride of 1 and a "same" padding method. The k'=1 point convolution is used for information integration between frequency channels, k'=3 captures local frequency band features, k'=5 extracts mid-frequency band patterns, and k'=7 obtains wideband response trends. The outputs of the four branches are concatenated in the channel dimension, followed by batch normalization and a ReLU activation function to capture frequency response characteristics related to the target material. Time-frequency fusion outputs the time-domain path. and frequency domain path output The features are concatenated along the feature dimension to form time-frequency features. .
5. The target material classification method based on a time-frequency dual-path attention fusion network according to claim 1, characterized in that, The adaptive channel-temporal dual attention module includes: The channel attention branch employs a Squeeze-and-Excitation structure. Features are compressed into channel descriptors through global average pooling, then passed through two fully connected layers. The first layer compresses the number of channels from C to C / r, and after ReLU activation, the second layer restores it to C channels. Finally, the channel attention weights are generated using the Sigmoid function. ; The temporal attention branch uses a 1×1 convolutional layer to map multi-channel features into a single-channel attention map, and the temporal attention weights are obtained after Softmax normalization. ; The channel attention branch and the temporal attention branch are fused by element-wise multiplication to generate joint attention weights. Output .
6. The target material classification method based on a time-frequency dual-path attention fusion network according to claim 1, characterized in that, The residual enhancement Inception module includes: Bottleneck layer: Add a convolutional bottleneck layer at the entrance of each Inception module to compress the number of input channels; Multi-scale Inception branch: Parallel branches with different convolutional kernel sizes are used to capture multi-granularity temporal features; Residual connections: Skip connections are introduced in each Inception module to directly add the input to the output. When the number of input and output channels is inconsistent, dimension matching is performed through convolution. Regularization strategy: Add a Dropout layer after each Inception module and use batch normalization to stabilize the training process.
7. The target material classification method based on a time-frequency dual-path attention fusion network according to claim 1, characterized in that, It also includes a strategy optimization module, which optimizes the category probability distribution output by the classification head module, including the following steps: We employ the cross-entropy loss function to introduce class weighting for the class imbalance problem; The AdamW optimizer was used, and the initial learning rate and weight decay coefficient were set. The learning rate is dynamically adjusted using a cosine annealing strategy. An early stopping mechanism is adopted to monitor the loss of the validation set, and training is terminated early if there is no improvement after several consecutive rounds.