Radar signal classification method and system based on adaptive fusion of multimodal features

Abstract

This application provides a radar signal classification method and system based on multimodal feature adaptive fusion. The method includes: performing compressed sensing processing on a time-frequency image using a pre-constructed sparse sampling matrix to generate observation data; inputting the observation data into a multi-branch feature extraction network to extract local texture features and global semantic features; calculating a global information-theoretic feature tensor based on the time-frequency image, and generating a gating weight matrix based on the global information-theoretic feature tensor; performing adaptive fusion of the local texture features and global semantic features using the gating weight matrix to generate adaptive fused features; and generating a radar signal classification result based on the adaptive fused features. The technical solution provided in this application reduces data dimensionality through compressed sensing, extracts complementary features using a multi-branch network, and combines an information-theoretic-guided adaptive fusion mechanism, effectively improving the accuracy and robustness of radar signal classification.

Description

Technical Field

[0001] This application relates to the field of radar signal processing technology, and in particular to a radar signal classification method and system based on adaptive fusion of multimodal features. Background Technology

[0002] In the field of radar signal processing, especially in target recognition and classification tasks under complex electromagnetic environments, efficient and accurate automatic classification of radar echo signals is crucial. These applications typically face the following core requirements: first, the complete time-frequency characteristics of the signal must be captured; second, the algorithm must meet real-time processing or bandwidth constraints; and finally, the classification model needs to maintain robustness and generalization ability in complex and ever-changing environments.

[0003] The existing approach uses a deep learning-based convolutional neural network. This approach first performs a time-frequency transformation on the original radar signal to obtain a time-frequency image, then applies compressed sensing technology to obtain low-dimensional observation data. Subsequently, a deep learning-based convolutional neural network model is designed, focusing on extracting fine local texture features and macroscopic global semantic features from the input observation data. Finally, the extracted features are subjected to a fixed fusion strategy, such as simple concatenation or weighted averaging, and then input into a subsequent classifier for target recognition.

[0004] However, the core flaw of existing solutions lies in the difficulty of adaptively optimizing the fusion ratio and method of features at different levels using fixed or simple fusion strategies. Specifically, the contributions of local texture features and global semantic features to the final classification of different categories of radar signals, or the same signal in different environments, may dynamically change. Fixed fusion strategies cannot dynamically adjust the fusion weights of the two features according to the specific content and characteristics of the signal, which can easily lead to insufficient information fusion or the introduction of redundant or even conflicting information, limiting the model's discriminative ability and generalization performance in complex scenarios. Summary of the Invention

[0005] This application provides a radar signal classification method and system based on multimodal feature adaptive fusion, which solves the problem of low radar signal classification accuracy caused by insufficient dynamic adaptive fusion capability in the prior art.

[0006] Firstly, this application provides a radar signal classification method based on adaptive fusion of multimodal features, including:

[0007] The time-domain waveform of the radar signal is acquired, and a short-time Fourier transform is performed on the time-domain waveform to generate a time-frequency image;

[0008] Compressed sensing processing is performed on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data;

[0009] The observation data is input into a multi-branch feature extraction network. Local texture features are extracted through the first branch of the multi-branch feature extraction network, and global semantic features are extracted through the second branch.

[0010] Calculate a global information theory feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and generate a gated weight matrix based on the global information theory feature tensor;

[0011] The local texture features and global semantic features are adaptively fused using the gated weight matrix to generate adaptive fused features, and the classification result of the radar signal is generated based on the adaptive fused features.

[0012] Optionally, the step of adaptively fusing the local texture features and global semantic features using the gating weight matrix to generate adaptive fused features includes:

[0013] The gate weight matrix is ​​split into a first sub-weight matrix and a second sub-weight matrix along the channel dimension;

[0014] The local texture features and the first sub-weight matrix are subjected to element-wise multiplication at the channel dimension to generate weighted local texture features. The global semantic features and the second sub-weight matrix are subjected to element-wise multiplication at the channel dimension to generate weighted global semantic features.

[0015] The weighted local texture features and the weighted global semantic features are concatenated along the channel dimension to generate a fused feature tensor;

[0016] Perform a convolution operation on the fused feature tensor to generate adaptive fused features.

[0017] Optionally, the step of performing a convolution operation on the fused feature tensor to generate adaptive fused features includes:

[0018] The sum of the number of channels of the local texture features and the number of channels of the global semantic features is used as the number of convolution kernels to construct a set of convolution kernels;

[0019] A sliding window convolution operation is performed on the fused feature tensor using a set of convolution kernels to generate a set of feature maps;

[0020] All feature maps in the feature map set are concatenated along the channel dimension to generate adaptive fusion features.

[0021] Optionally, the step of concatenating all feature maps in the feature map set along the channel dimension to generate adaptive fusion features includes:

[0022] Calculate the channel statistics for each feature map in the feature map set to generate a channel statistics set;

[0023] Based on the set of channel statistics, the distribution differences of different feature maps are compared to determine the channel selection priority parameters, which are a set of priority weight values ​​for all channels.

[0024] The channel sequences corresponding to the feature mapping set are reordered according to the channel selection priority parameter to generate an ordered channel sequence.

[0025] The channels corresponding to different priority weight values ​​in the ordered channel sequence are alternately recombined to generate a set of recombined feature maps;

[0026] All recombined feature maps in the recombined feature map set are concatenated along the channel dimension to generate adaptive fusion features.

[0027] Optionally, the step of performing compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data includes:

[0028] The pre-constructed basis function vectors are arranged and combined according to a predefined dimension sequence to generate a sparse sampling matrix;

[0029] The pixel data of the time-frequency image is converted into a data vector according to the raster scan order;

[0030] Calculate the dot product of each measurement vector in the sparse sampling matrix with the data vector to generate a set of scalar elements;

[0031] The elements in the scalar element set are combined in row index order to generate observation data.

[0032] Optionally, the step of calculating a global information-theoretic feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and generating a gated weight matrix based on the global information-theoretic feature tensor, includes:

[0033] The spatial dimensions of the time-frequency image are divided to generate multiple time-frequency sub-blocks;

[0034] Calculate the Shannon entropy features of all time-frequency sub-blocks in each time-frequency region, and select the time-frequency sub-block at the center of each time-frequency region as the reference time-frequency sub-block;

[0035] The relative entropy characteristics of the remaining time-frequency sub-blocks within the time-frequency region are calculated based on the reference time-frequency sub-block;

[0036] By concatenating the Shannon entropy features and relative entropy features of all time-frequency sub-blocks within all time-frequency regions, a global information-theoretic feature tensor is generated.

[0037] The channel dimension of the global information theory feature tensor is compressed to generate a gated weight mapping;

[0038] The gate weight mapping is added to the preset gate basis to generate a gate weight matrix.

[0039] Optionally, the step of splicing the Shannon entropy features and relative entropy features of all time-frequency sub-blocks in all time-frequency regions to generate a global information-theoretic feature tensor includes:

[0040] By concatenating the Shannon entropy features and relative entropy features of all time-frequency sub-blocks within each time-frequency region, a local information-theoretic feature vector is generated;

[0041] The local information-theoretic feature vectors corresponding to all time-frequency regions are aggregated to generate a global information-theoretic feature tensor.

[0042] Secondly, this application provides a radar signal classification system based on adaptive fusion of multimodal features, comprising:

[0043] The acquisition module is used to acquire the time-domain waveform of the radar signal, perform a short-time Fourier transform on the time-domain waveform, and generate a time-frequency image.

[0044] The compression module is used to perform compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data.

[0045] The extraction module is used to input the observation data into a multi-branch feature extraction network, extract local texture features through the first branch of the multi-branch feature extraction network, and extract global semantic features through the second branch;

[0046] The calculation module is used to calculate a global information theory feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and to generate a gated weight matrix based on the global information theory feature tensor;

[0047] The fusion module is used to perform adaptive fusion of the local texture features and global semantic features using the gating weight matrix to generate adaptive fused features, and to generate classification results of radar signals based on the adaptive fused features.

[0048] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a radar signal classification method based on multimodal feature adaptive fusion as described in any of the first aspects.

[0049] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a radar signal classification method based on multimodal feature adaptive fusion as described in any of the first aspects.

[0050] This application provides a radar signal classification method based on adaptive fusion of multimodal features. The method includes: acquiring the time-domain waveform of the radar signal; performing a short-time Fourier transform on the time-domain waveform to generate a time-frequency image; performing compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data; inputting the observation data into a multi-branch feature extraction network, extracting local texture features through the first branch of the network and extracting global semantic features through the second branch; calculating a global information-theoretic feature tensor containing Shannon entropy and relative entropy features based on the time-frequency image, and generating a gating weight matrix based on the global information-theoretic feature tensor; performing adaptive fusion of the local texture features and global semantic features using the gating weight matrix to generate adaptive fused features; and generating a radar signal classification result based on the adaptive fused features.

[0051] This application converts radar time-domain signals into time-frequency images using short-time Fourier transform and effectively reduces data dimensionality by utilizing compressed sensing to generate observation data. Furthermore, a multi-branch feature extraction network is employed to capture local texture details and global semantic structure from the observation data. Simultaneously, a global information-theoretic feature tensor is innovatively constructed based on the Shannon entropy and relative entropy of the time-frequency images, and a gating weight matrix is ​​generated accordingly. Finally, this gating weight matrix is ​​used to adaptively fuse local texture and global semantic features, generating more discriminative fused features, thereby improving the classification accuracy, robustness, and generalization ability of radar signals in complex scenarios.

[0052] Furthermore, the gated weight matrix is ​​split into two sub-weight matrices along the channel dimension. Then, element-wise multiplication along the channel dimension is performed on the local texture features and the first sub-weight matrix to generate weighted local texture features; simultaneously, element-wise multiplication along the channel dimension is performed on the global semantic features and the second sub-weight matrix to generate weighted global semantic features. Next, these two weighted features are concatenated along the channel dimension to form a fused feature tensor. Finally, a specific convolution operation is performed on this fused feature tensor: a set of convolution kernels with the number of kernels equal to the sum of the number of channels for the local texture features and the global semantic features is constructed. This set is used to perform a sliding window convolution on the fused feature tensor, and all generated feature maps are concatenated along the channel dimension to finally output the adaptive fused feature. By splitting the gated weight matrix along the channel dimension and performing element-wise multiplication, fine-grained and differentiated weighting of local texture features and global semantic features along the channel dimension is achieved, ensuring that the key information of each feature is enhanced in a targeted manner. Furthermore, the concatenation operation preserves the complete information of the weighted features. Finally, by using a specific convolution operation that uses the sum of the feature channels as the number of convolution kernels, we effectively promote the deep interaction and information integration of weighted local features and weighted global features in spatial and channel dimensions, thereby generating more discriminative and robust adaptive fusion features, laying a solid foundation for subsequent high-precision classification.

[0053] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating a radar signal classification method based on adaptive fusion of multimodal features, provided in this application embodiment;

[0056] Figure 2 A schematic diagram of the structure of a radar signal classification system with adaptive fusion of multimodal features provided in this application embodiment;

[0057] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0058] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0059] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 11, 12, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0060] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] To address the problem of low radar signal classification accuracy caused by insufficient dynamic adaptive fusion capabilities in existing technologies, this application provides a radar signal classification method based on multimodal feature adaptive fusion. This method employs the following concept: first, the original signal is converted into a time-frequency image and simplified sampling is performed to reduce computational burden; then, through dual-channel parallel processing, signal detail texture and overall structural features are extracted separately; most importantly, an intelligent adjustment mechanism is introduced, dynamically generating weight switches by analyzing the image's own information characteristics to automatically adjust the fusion ratio of local and global features; finally, more adaptive fusion features are generated, improving the accuracy and stability of signal classification under complex interference environments.

[0062] Figure 1 A flowchart of a radar signal classification method based on adaptive fusion of multimodal features provided in this application embodiment is shown below. Figure 1 As shown, the method includes:

[0063] S11. Acquire the time-domain waveform of the radar signal, perform a short-time Fourier transform on the time-domain waveform, and generate a time-frequency image.

[0064] Radar signals refer to electromagnetic wave signals emitted or received by a radar system, including pulse or continuous wave forms, used to detect the distance, velocity, or azimuth of target objects. Time-domain waveforms represent the amplitude variations of radar signals along the time axis, reflecting a continuous sequence of signal energy evolution over time. Short-time Fourier transform (SFT) is a time-frequency analysis technique that segments the signal and applies Fourier transform to each segment to generate a joint time-frequency distribution, used to capture the frequency characteristics of non-stationary signals. A time-frequency image is a two-dimensional matrix obtained through SFT, with the horizontal axis representing time, the vertical axis representing frequency, and pixel values ​​representing signal energy intensity, used to visualize the time-frequency characteristics of signals.

[0065] In this embodiment, firstly, a time-domain waveform of a radar signal is acquired by a radar receiving device. This waveform reflects a continuous sequence of signal amplitude changes over time. Secondly, a short-time Fourier transform is performed on the time-domain waveform. This transform involves dividing the signal into multiple short time windows and applying a Fourier transform within each window to calculate the frequency components, thereby generating a time-frequency image that displays the energy distribution characteristics of the signal in both the time and frequency dimensions.

[0066] S12. Perform compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data.

[0067] In this context, a sparse sampling matrix refers to a matrix used in compressed sensing. It can be designed with a highly sparse structure, meaning most elements are zero, to perform dimensionality reduction sampling on high-dimensional data. Compressed sensing is a signal processing technique that, based on the assumption of signal sparsity, uses a sparse sampling matrix to obtain a small number of linear measurements for efficient signal compression and reconstruction. Observational data refers to the low-dimensional data vector output by compressed sensing, serving as a compressed representation of the original time-frequency image.

[0068] In this embodiment, firstly, a pre-constructed sparse sampling matrix is ​​used. This matrix is ​​designed based on compressed sensing theory and has a low-density non-zero element structure. Secondly, compressed sensing processing is performed on the generated time-frequency image. Specifically, this includes performing a linear projection operation between the sparse sampling matrix and the time-frequency image, acquiring a small number of key data points through the measurement process, thereby generating observation data. This data serves as a compressed representation of the time-frequency image for subsequent processing.

[0069] S13. Input the observation data into the multi-branch feature extraction network, extract local texture features through the first branch of the multi-branch feature extraction network, and extract global semantic features through the second branch.

[0070] Multi-branch feature extraction networks (MBINs) are neural network architectures that incorporate multiple parallel processing branches to extract features of different types and scales from input data. The first branch is dedicated to extracting local details, typically using convolutional layers to focus on local patterns in the signal. Local texture features are feature vectors extracted by the feature extraction network, reflecting the local structure, edges, or texture variations of the time-frequency image. The second branch is dedicated to extracting the global context, typically using fully connected or deep convolutional layers to capture the overall semantics of the signal. Global semantic features are feature vectors extracted by the feature extraction network, reflecting the overall category, structure, or semantic information of the time-frequency image.

[0071] In this embodiment, the first branch adopts a three-layer convolutional structure. First, a 3×3 convolutional kernel is used to extract the spatial local patterns of the observed data, and the output feature map is batch normalized. Second, a 2×2 max pooling layer with a stride of 2 is used to compress the spatial dimension and enhance texture invariance. Finally, two cascaded 3×3 convolutional layers are used to deepen the capture of local details, and the number of output channels is gradually increased to 128 dimensions. The end is connected to a global average pooling layer to generate a local texture feature vector. The second branch is designed as a multi-scale dilated convolutional path. The first layer uses a 3×3 dilated convolution with an dilation rate of 1 to capture the basic context. Then, two 3×3 dilated convolutional layers with dilation rates of 2 and 4 are connected in parallel to extract semantic information from different receptive fields simultaneously. The three outputs are concatenated and then fused with multi-scale features through a 1×1 convolution. Finally, a depthwise separable convolution with a stride of 2 is used to reduce the dimensionality, and a 128-dimensional global semantic feature vector is generated through global max pooling.

[0072] In this embodiment, the observation data is first input into a multi-branch feature extraction network. Then, through the first branch of this network, which employs a convolutional neural network structure, the local data of the observation data is analyzed to extract local texture features. These features capture subtle patterns and structural changes in the time-frequency images. Simultaneously, through the second branch, which utilizes fully connected layers or large-scale convolutional layers, the overall context of the observation data is analyzed to extract global semantic features. These features reflect the macroscopic semantics and category information of the signal.

[0073] S14. Calculate the global information theory feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and generate the gated weight matrix based on the global information theory feature tensor.

[0074] Here, Shannon entropy features refer to entropy measures in information theory, used to calculate the degree of information uncertainty or randomness in time-frequency images, obtained based on the logarithmic operation of probability distributions. Relative entropy features refer to divergence in information theory, used to measure the degree of difference between the probability distribution of a time-frequency image and a reference distribution. Both Shannon and relative entropy features calculate the spatial pixel distribution of the time-frequency image. The global information theory feature tensor is a multidimensional data tensor composed of Shannon and relative entropy features, used to describe the global statistical and uncertainty characteristics of the time-frequency image. The gated weight matrix is ​​a weight matrix generated based on the global information theory feature tensor, explicitly defined by trainable parameters and nonlinear functions. The formula for generating the gated weight matrix is: Among them, M gate G is the gated weight matrix. comp This is the compressed single-channel feature map. This is an operation to negate the exponent.

[0075] In this embodiment, firstly, a global information-theoretic feature tensor is calculated based on the time-frequency image. This tensor includes Shannon entropy features and relative entropy features. The Shannon entropy features are calculated, which evaluate the information uncertainty of the time-frequency image using the information entropy formula, i.e., the logarithmic expectation of the probability distribution. The relative entropy features are calculated, which measure the degree of difference between the time-frequency image and the reference distribution. Secondly, a gating weight matrix is ​​generated based on the global information-theoretic feature tensor.

[0076] S15. Adaptive fusion of local texture features and global semantic features is performed using a gating weight matrix to generate adaptive fused features. The classification result of radar signal is generated based on the adaptive fused features.

[0077] Adaptive fusion features refer to feature vectors generated by weighted fusion of local texture features and global semantic features using a gating weight matrix, which are used to enhance classification discriminative power. The classification result refers to the output of the radar signal after classification processing, indicating the target category or type label to which the signal belongs.

[0078] In this embodiment, firstly, an adaptive fusion of local texture features and global semantic features is performed using a gated weight matrix. Specifically, this involves performing element-wise multiplication of the weight matrix and the features, followed by weighted combination to generate an adaptive fused feature. Secondly, a classification result for the radar signal is generated based on this adaptive fused feature.

[0079] Here is a specific example: First, the radar receiving system acquires the time-domain waveform data of the radar signal reflected from the target. Second, a short-time Fourier transform is applied to this time-domain waveform to generate corresponding time-frequency image data. Next, compressed sensing measurements are performed on this time-frequency image using a pre-generated sparse sampling matrix to obtain an observation data vector. Then, the observation data is input into a multi-branch feature extraction network, where the first branch processes the data to extract local texture features, and the second branch processes the data to extract global semantic features. Simultaneously, Shannon entropy and relative entropy features are calculated based on the time-frequency image to form a global information-theoretic feature tensor, and a transformation function is applied to generate a gating weight matrix. Subsequently, the gating weight matrix is ​​used to weight and combine the local texture features and global semantic features to generate an adaptive fusion feature vector. Finally, a classifier processes this fusion feature vector to output the final classification result of the radar signal, such as the target identification category.

[0080] By executing S11 to S15, the embodiments of this application reduce data dimensionality through compressed sensing processing, extract complementary local and global features by combining multi-branch networks, and use information theory features to guide adaptive fusion, effectively improving the accuracy and robustness of radar signal classification, while reducing the computational resource requirements, making it suitable for real-time signal processing in complex environments.

[0081] In one possible embodiment, S15, adaptive fusion of local texture features and global semantic features is performed using a gating weight matrix to generate adaptive fused features, including:

[0082] Step 151: Split the gate weight matrix along the channel dimension into a first sub-weight matrix and a second sub-weight matrix.

[0083] The first sub-weight matrix refers to the block matrix obtained by segmenting the gated weight matrix along the channel dimension. Its channel number is consistent with the local texture features, and it is used to weight the local features by channel-level importance. The second sub-weight matrix refers to another block matrix after segmenting the gated weight matrix. Its channel number matches the global semantic features, and it is used to adjust the contribution strength of the global features. Segmentation along the channel dimension involves determining the total number of channels C of the gated weight matrix, simultaneously obtaining the number of local texture feature channels C1 from the first branch output and the number of global semantic feature channels C2 from the second branch output, and verifying that the dimension matching condition C = C1 + C2 is satisfied. The first C1 channels of the gated weight matrix are extracted to generate a sub-matrix with the dimension of the first sub-weight, where each channel is strictly aligned with the corresponding channel of the local texture feature. The last C2 channels of the gated weight matrix are extracted to generate a sub-matrix with the dimension of the second sub-weight, where each channel is bound one-to-one with the corresponding channel of the global semantic feature.

[0084] In this embodiment, firstly, the gated weight matrix is ​​divided along the channel dimension, splitting it into two independent parts by specifying the proportion of the number of channels. Secondly, a first sub-weight matrix matching the number of local texture feature channels is generated, and a second sub-weight matrix matching the number of global semantic feature channels is generated simultaneously, ensuring that the channel dimensions of the two sub-matrices correspond to the feature maps to be weighted.

[0085] Step 152: Perform element-wise multiplication of the local texture features and the first sub-weight matrix in the channel dimension to generate weighted local texture features. Perform element-wise multiplication of the global semantic features and the second sub-weight matrix in the channel dimension to generate weighted global semantic features.

[0086] Among them, weighted local texture features refer to the features generated by element-wise multiplication of local texture features with the first sub-weight matrix, reflecting detailed information that retains the original texture structure but has been optimized by weights. Weighted global semantic features refer to the features generated by channel-level multiplication of global semantic features with the second sub-weight matrix, containing overall semantic information enhanced by weight filtering.

[0087] In this embodiment, firstly, a channel-dimensional alignment operation is performed on the local texture features, and then a channel-wise element-wise multiplication operation is performed with the first sub-weight matrix, that is, the feature map of each channel is multiplied by the weight coefficient of the corresponding channel. Secondly, a weighted local texture feature is generated, which retains the original texture details but adjusts the importance according to the weights. At the same time, the same operation is performed on the global semantic features, and a channel-wise element-wise multiplication operation is performed with the second sub-weight matrix to generate a weighted global semantic feature, which strengthens key semantic information and suppresses minor components.

[0088] Step 153: Concatenate the weighted local texture features and the weighted global semantic features along the channel dimension to generate a fused feature tensor.

[0089] The fused feature tensor refers to multidimensional data formed by concatenating weighted local texture features and weighted global semantic features along the channel dimension. It contains both local and global features but is not deeply integrated. Concatenating along the channel dimension, rather than the spatial dimension, ensures complete spatial alignment between the weighted local texture features and the weighted global semantic features. Concatenating along the spatial dimension would disrupt the spatial topology of the feature map, leading to the loss of continuity in the target's time-frequency structure.

[0090] In this embodiment, firstly, the weighted local texture features and the weighted global semantic features are aligned along the channel dimension. Secondly, a concatenation operation is performed along the channel dimension to sequentially connect the channels of the two feature tensors, generating a fused feature tensor. This tensor simultaneously contains local details and global semantic information, and its channel dimension is expanded to the sum of the original features.

[0091] Step 154: Perform a convolution operation on the fusion feature tensor to generate adaptive fusion features.

[0092] Convolution operation refers to the process of performing spatial sliding computation on the fused feature tensor using a convolution kernel, achieving feature integration and dimensionality compression through linear weighting and nonlinear activation.

[0093] In this embodiment, firstly, a convolution operation is applied to the fused feature tensor, using a convolution kernel of a preset size to slide and compute in the spatial dimension. Secondly, the concatenated multi-channel features are integrated through a weighted summation mechanism of the convolution kernel to generate an adaptive fused feature. This feature retains complementary information while eliminating channel redundancy and enhancing classification discriminativeness.

[0094] Here's a concrete example: First, the gated weight matrix is ​​split along the channel axis into a first sub-weight matrix and a second sub-weight matrix. Second, the local texture features of the radar signal are multiplied channel-wise with the first sub-weight matrix to generate weighted local texture features, while the global semantic features are multiplied channel-wise with the second sub-weight matrix to generate weighted global semantic features. Next, the weighted local texture features and weighted global semantic features are concatenated along the channel dimension to form a fused feature tensor. Finally, a convolution kernel is applied to this tensor to perform spatial feature integration, outputting an adaptive fused feature for radar target classification.

[0095] By executing steps 151 to 154, this embodiment of the application achieves feature-specific weighting through channel-level weight segmentation, retains multi-scale information by combining splicing operations, and finally eliminates redundancy through convolutional fusion, thereby enhancing the discriminativeness and classification robustness of features while maintaining computational efficiency.

[0096] In one possible embodiment, step 154, performing a convolution operation on the fused feature tensor to generate adaptive fused features, includes:

[0097] Step a1: Use the sum of the number of channels of local texture features and the number of channels of global semantic features as the number of convolution kernels to construct a set of convolution kernels.

[0098] Here, the number of channels refers to the number of feature tensors in the channel dimension, reflecting the number of information components of different types or scales contained in the feature map. A convolution kernel is a weight matrix used for convolution operations; it extracts local spatial pattern features by sliding across the input feature map and performing dot product operations. A convolution kernel set is a collection of multiple convolution kernels, the number of which is determined by the channel requirements of the input features, used for parallel extraction of multiple types of spatial features.

[0099] In this embodiment, firstly, the sum of the number of channels for local texture features and the number of channels for global semantic features is calculated. Secondly, this sum is used as the basis for generating the number of convolutional kernels, and a set of convolutional kernels containing the corresponding number of convolutional kernels is constructed to ensure that each convolutional kernel corresponds to the channel processing requirements of the fused feature tensor.

[0100] Step a2: Perform a sliding window convolution operation on the fused feature tensor using the set of convolution kernels to generate a set of feature maps.

[0101] Among them, the sliding window convolution operation refers to the process of moving the convolution kernel pixel by pixel in the feature map space dimension and performing local weighted calculations, which is used to capture spatial local correlations.

[0102] In this embodiment, firstly, a sliding window convolution operation is performed on the fused feature tensor using a set of convolutional kernels. Specifically, this involves moving the local window of the feature map covered by each convolutional kernel in the spatial dimension. Secondly, the feature values ​​within each local window are linearly weighted and summed with the convolutional kernel weights to generate the corresponding feature map. Finally, the processing results of all convolutional kernels together constitute a feature map set, which contains multi-scale spatial response information.

[0103] Step a3: Concatenate all feature maps in the feature map set along the channel dimension to generate adaptive fusion features.

[0104] In this context, a feature map refers to the output matrix generated after a single convolutional kernel performs a sliding window convolution on the input features, reflecting the activation response of a specific spatial pattern. The feature map set is the collection of feature maps generated after all convolutional kernels process the input features, containing heterogeneous spatial features extracted by multiple kernels.

[0105] In this embodiment, firstly, all feature maps in the feature map set are aligned in channel order. Secondly, a concatenation operation is performed along the channel dimension to sequentially connect the feature maps generated by different convolutional kernels, generating an adaptive fusion feature. This feature integrates complementary spatial information extracted by multiple kernels, and the channel dimension is expanded to the sum of the number of convolutional kernels.

[0106] Here's a concrete example: First, the sum of the number of local texture feature channels and the number of global semantic feature channels in the radar signal is calculated, and a set of convolutional kernels is constructed based on this number. Second, this set of convolutional kernels is used to perform a sliding window convolution operation on the fused feature tensor, traversing each spatial location of the feature map to generate a set of feature maps. Finally, all feature maps are concatenated sequentially along the channel dimension to generate the final adaptive fused features for radar target recognition.

[0107] By executing steps a1 to a3, this embodiment of the application dynamically configures the number of convolutional kernels to adapt to the feature channels, and combines multi-kernel spatial feature extraction and channel splicing to achieve efficient deep fusion of local and global features, thereby improving feature representation capabilities and optimizing classification accuracy.

[0108] In one possible embodiment, step a3, concatenating all feature maps in the feature map set along the channel dimension to generate adaptive fusion features, includes:

[0109] Step a31: Calculate the channel statistics for each feature map in the feature map set, and generate a channel statistics set.

[0110] Channel statistics refer to statistical indicators of feature mappings along the channel dimension, including mean or variance, used to describe the central tendency or dispersion of the feature distribution. The set of channel statistics refers to the collection of channel statistics for all feature mappings, reflecting the overall distribution characteristics of the feature mapping set.

[0111] In this embodiment, firstly, the channel statistics of each feature map in the feature map set are calculated. These statistics are obtained by calculating the mean or variance of the feature map along the channel dimension. Secondly, the channel statistics of all feature maps are merged to form a channel statistics set, which reflects the distribution characteristics of each feature map.

[0112] Step a32: Based on the set of channel statistics, compare the distribution differences of different feature maps to determine the channel selection priority parameters. The channel selection priority parameters are a set of priority weight values ​​for all channels.

[0113] The distribution difference value refers to a scalar value calculated using a probability distribution difference metric. It quantifies the degree of statistical distribution difference between different feature maps, and the metric is defined as the Kullback-Leibler (KL) divergence. The distribution difference values ​​are compared by ranking. The channel selection priority parameter is a set of weight coefficients generated based on the distribution difference value comparison, used to define the importance order of each feature map channel in the fusion process. The priority weight value refers to the specific numerical component in the channel selection priority parameter; a larger value indicates a higher priority for the corresponding channel in the fusion process.

[0114] In this embodiment, firstly, the distribution difference value between different feature maps is calculated based on the channel statistics set. This value is obtained by comparing the probability distribution difference measure of the feature maps. Secondly, a channel selection priority parameter is determined based on the relative magnitude of the distribution difference value. This parameter is a set of weight values ​​used to identify the importance priority of each channel in feature fusion.

[0115] Step a33: Reorder the channel sequences corresponding to the feature mapping set according to the channel selection priority parameter to generate an ordered channel sequence.

[0116] Here, a channel sequence refers to the sequence of all channels in the feature map set arranged in their original order. An ordered channel sequence refers to a channel sequence rearranged in descending order of priority weight values, ensuring that highly important channels are at the beginning of the sequence.

[0117] In this embodiment, firstly, the original channel sequence of the feature mapping set is sorted according to the channel selection priority parameter. Secondly, channels with higher priority weight values ​​are arranged first to generate an ordered channel sequence, thus placing the key feature channels at the forefront.

[0118] Step a34: Alternately recombine the channels corresponding to different priority weight values ​​in the ordered channel sequence to generate a recombined feature mapping set.

[0119] Among them, the recombined feature map set refers to the new feature map set generated after performing alternating recombination operations on the ordered channel sequence, so as to achieve a balanced priority distribution.

[0120] In this embodiment, firstly, the ordered channel sequence is divided into a high-priority group and a low-priority group according to its priority weight value. Secondly, an alternating recombination operation is performed on the two groups of channels, that is, the high-priority channels and low-priority channels are arranged alternately at fixed intervals to generate a recombined feature map set to enhance feature diversity.

[0121] Step a35: Concatenate all recombined feature maps in the recombined feature map set along the channel dimension to generate adaptive fusion features.

[0122] In this embodiment, firstly, all feature maps in the recombined feature map set are aligned according to the recombined channel order. Secondly, a splicing operation is performed along the channel dimension to generate the final adaptive fusion feature, which retains priority information and has enhanced discriminability.

[0123] Here's a specific example: First, the channel mean of each feature map in the radar signal feature map set is calculated to form a channel statistics set. Second, based on this set, the distribution difference value is calculated to generate channel selection priority parameters. Next, the channel sequences are sorted in descending order according to their priority weight values ​​to form an ordered channel sequence. Subsequently, high-priority channels and low-priority channels are alternately recombine to generate a recombined feature map set. Finally, the recombined feature maps are concatenated along the channel dimension to generate adaptive fusion features for target recognition.

[0124] By executing steps a31 to a35, the embodiments of this application optimize feature priority allocation through channel statistics analysis and enhance feature diversity by combining an alternating recombination strategy, thereby effectively improving the discriminative ability and classification robustness of the fused features.

[0125] In one possible embodiment, S12, performing compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data, including:

[0126] Step 121: Arrange and combine the pre-constructed basis function vectors according to the predefined dimension sequence to generate a sparse sampling matrix.

[0127] Here, the dimension sequence refers to the predefined order of the dimensions of the basis function vectors, which guides the rules for recombining the basis function vectors. The basis function vectors are the fundamental vectors that constitute the sparse sampling matrix, and are usually selected from orthogonal function sets such as Fourier basis or wavelet basis.

[0128] Step 122: Convert the pixel data of the time-frequency image into a data vector according to the raster scan order.

[0129] Here, the data vector refers to a one-dimensional vector obtained by converting a time-frequency image according to the raster scan order, containing all pixel values ​​of the original image. The raster scan order refers to the standardized order in which pixels in the image are traversed row by row from left to right and from top to bottom, used to flatten two-dimensional data into a one-dimensional vector.

[0130] Step 123: Calculate the dot product of each measurement vector and data vector in the sparse sampling matrix to generate a set of scalar elements.

[0131] Here, the measurement vector refers to a single row vector of the sparse sampling matrix, representing the projection direction of a linear measurement operation. The dot product result is the scalar value generated by the sum of the element-wise products of the measurement vector and the data vector, reflecting the strength of the linear correlation between the two vectors. The set of scalar elements is the set of all dot product results of the measurement vector and the data vector, containing the original measurements from compressed sensing.

[0132] Step 124: Combine the elements in the scalar element set according to the row index order to generate observation data.

[0133] The row index order refers to the row numbering order of the sparse sampling matrix, which is used to determine the position of scalar elements in the observation data.

[0134] Here is a specific example: First, the basis function vectors are reorganized according to a predefined dimensional sequence to generate a sparse sampling matrix. Second, the radar time-frequency image is converted into a data vector according to the raster scan order. Next, the measurement vectors of the sparse sampling matrix are extracted row by row, and dot products are performed with the data vectors to generate a set of scalar elements. Finally, the scalar element sets are combined according to the row index order to generate observation data for radar signal classification processing.

[0135] By executing steps 121 to 124, this embodiment of the application constructs a sparse sampling matrix through basis function recombination, combines raster scanning to achieve image vectorization, and uses dot product operations to efficiently generate compressed observation data, reducing data dimensionality while retaining key information and improving subsequent processing efficiency.

[0136] In one possible embodiment, S14, calculating a global information-theoretic feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and generating a gated weight matrix based on the global information-theoretic feature tensor, includes:

[0137] Step 141: Divide the spatial dimensions of the time-frequency image to generate multiple time-frequency sub-blocks.

[0138] Spatial dimension refers to the height and width dimensions of the time-frequency image, reflecting the two-dimensional spatial distribution characteristics of the signal along the time and frequency axes. A time-frequency sub-block refers to a local image block generated by dividing the spatial dimension of the time-frequency image, containing signal energy distribution information for a specific time period and frequency band.

[0139] Step 142: Calculate the Shannon entropy features of all time-frequency sub-blocks in each time-frequency region, and select the time-frequency sub-block at the center of each time-frequency region as the reference time-frequency sub-block.

[0140] In this context, the time-frequency region refers to a subset of the image composed of adjacent time-frequency sub-blocks, used to define the range unit for local feature calculation. The reference time-frequency sub-block refers to the sub-block selected as the center position within the time-frequency region, whose pixel distribution serves as the reference benchmark for relative entropy calculation. The time-frequency region is divided by dividing the spatial dimension of the time-frequency image into a uniform rectangular grid, with each grid unit called a time-frequency sub-block.

[0141] Step 143: Calculate the relative entropy characteristics of the remaining time-frequency sub-blocks within the time-frequency region based on the reference time-frequency sub-block.

[0142] Step 144: Concatenate the Shannon entropy features and relative entropy features of all time-frequency sub-blocks in all time-frequency regions to generate a global information theory feature tensor.

[0143] Step 145: Compress the channel dimension of the global information theory feature tensor to generate a gated weight mapping.

[0144] Among them, gated weight mapping refers to the intermediate weights generated by channel compression of the global information theory feature tensor, which reflects the importance distribution of information statistical characteristics.

[0145] Step 146: Add the gate weight mapping to the preset gate basis to generate the gate weight matrix.

[0146] Among them, the preset gating basis refers to the pre-set weight basis matrix, which is used to adjust the numerical range and distribution characteristics of the gating weight mapping.

[0147] Here is a specific example: First, the radar time-frequency image is divided into multiple time-frequency sub-blocks. Second, for each time-frequency region, the Shannon entropy features of all sub-blocks are calculated, and a central sub-block is selected as the reference. Next, the relative entropy features of non-reference sub-blocks in each region relative to the reference sub-block are calculated. Subsequently, all entropy features are concatenated to generate a global information-theoretic feature tensor. Then, the channel dimension of the global information-theoretic feature tensor is compressed to obtain a gated weight map. Finally, the gated weight map is added to a preset gated basis to generate a gated weight matrix for radar signal feature fusion.

[0148] By executing steps 141 to 146, this embodiment of the application captures time-frequency statistical characteristics through local information entropy analysis, enhances feature discriminativeness by combining a benchmark reference mechanism, and finally generates optimized gating weights through compression and basis adjustment, thereby improving the accuracy and robustness of feature fusion.

[0149] In one possible embodiment, step 144, splicing the Shannon entropy features and relative entropy features of all time-frequency sub-blocks within all time-frequency regions to generate a global information-theoretic feature tensor, includes:

[0150] Step b1: Concatenate the Shannon entropy features and relative entropy features of all time-frequency sub-blocks in each time-frequency region to generate a local information theory feature vector.

[0151] Among them, the local information theory feature vector refers to the feature vector formed by splicing the Shannon entropy features and relative entropy features of all time-frequency sub-blocks in a single time-frequency region, which is used to describe the signal statistical distribution characteristics and information difference characteristics of the local region.

[0152] In this embodiment, firstly, for all time-frequency sub-blocks contained in each time-frequency region, their Shannon entropy features and relative entropy features are extracted respectively. Secondly, the Shannon entropy features and relative entropy features of all sub-blocks in the same time-frequency region are sequentially concatenated along the feature dimension to generate a local information-theoretic feature vector for that region. This vector comprehensively reflects the information statistical characteristics within the region.

[0153] Step b2: Aggregate the local information-theoretic feature vectors corresponding to all time-frequency regions to generate a global information-theoretic feature tensor.

[0154] In this embodiment, firstly, local information-theoretic feature vectors corresponding to all time-frequency regions are collected. Secondly, an aggregation operation is performed along the spatial dimension to arrange and combine the local information-theoretic feature vectors of different regions according to their original spatial positions, generating a global information-theoretic feature tensor. This tensor fully preserves the spatial distribution and statistical characteristics of the time-frequency image.

[0155] Here is a specific example: First, for each time-frequency region of the radar time-frequency image, the Shannon entropy features and relative entropy features of all its internal time-frequency sub-blocks are concatenated to generate a local information-theoretic feature vector. Second, the local information-theoretic feature vectors of all regions are aggregated into a global information-theoretic feature tensor in spatial topological order, which is used for the subsequent generation of the gating weight matrix.

[0156] By executing steps b1 to b2, the embodiments of this application integrate local information entropy features in a hierarchical manner, preserving spatial correlation while constructing a global statistical representation, thereby enhancing the discriminative ability and semantic integrity of the feature tensor and providing a precise information theory basis for gating fusion.

[0157] Figure 2 A schematic diagram of a radar signal classification system with adaptive fusion of multimodal features provided in this application embodiment is shown below. Figure 2 As shown, the system includes:

[0158] The acquisition module 21 is used to acquire the time-domain waveform of the radar signal, perform a short-time Fourier transform on the time-domain waveform, and generate a time-frequency image.

[0159] Compression module 22 is used to perform compressed sensing processing on time-frequency images using a pre-constructed sparse sampling matrix to generate observation data.

[0160] The extraction module 23 is used to input the observation data into the multi-branch feature extraction network, extract local texture features through the first branch of the multi-branch feature extraction network, and extract global semantic features through the second branch.

[0161] The calculation module 24 is used to calculate a global information theory feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and to generate a gated weight matrix based on the global information theory feature tensor.

[0162] The fusion module 25 is used to perform adaptive fusion of local texture features and global semantic features using a gated weight matrix to generate adaptive fused features, and generate classification results of radar signals based on the adaptive fused features.

[0163] Figure 2 The aforementioned radar signal classification system based on adaptive fusion of multimodal features can perform... Figure 1The implementation principle and technical effects of the radar signal classification method based on adaptive fusion of multimodal features described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit performs its operations in the radar signal classification system based on adaptive fusion of multimodal features described in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0164] In one possible design, Figure 2 The radar signal classification system based on multimodal feature adaptive fusion of the embodiment shown can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32.

[0165] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0166] The processing component 32 performs the following steps: acquiring the time-domain waveform of the radar signal, performing a short-time Fourier transform on the time-domain waveform to generate a time-frequency image; performing compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data; inputting the observation data into a multi-branch feature extraction network, extracting local texture features through the first branch of the network, and extracting global semantic features through the second branch; calculating a global information-theoretic feature tensor containing Shannon entropy and relative entropy features based on the time-frequency image, and generating a gating weight matrix based on the global information-theoretic feature tensor; performing adaptive fusion of local texture features and global semantic features using the gating weight matrix to generate adaptive fused features, and generating a classification result for the radar signal based on the adaptive fused features.

[0167] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0168] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Random Access Memory (RAM), Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0169] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0170] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0171] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0172] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0173] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown illustrates a radar signal classification method based on multimodal feature adaptive fusion.

[0174] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0175] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A radar signal classification method of multi-modal feature adaptive fusion, characterized in that, The method comprises the following steps: Collecting a time-domain waveform of a radar signal, performing a short-time Fourier transform on the time-domain waveform to generate a time-frequency image; Performing compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data; Inputting the observation data into a multi-branch feature extraction network, extracting local texture features through a first branch of the multi-branch feature extraction network, and extracting global semantic features through a second branch; Calculating a global information theory feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and generating a gating weight matrix based on the global information theory feature tensor; Performing adaptive fusion on the local texture features and the global semantic features using the gating weight matrix to generate adaptive fusion features, and generating a classification result of the radar signal according to the adaptive fusion features; The adaptive fusion features generated by performing adaptive fusion on the local texture features and the global semantic features using the gating weight matrix comprise: Splitting the gating weight matrix along the channel dimension into a first sub-weight matrix and a second sub-weight matrix; Performing element-level multiplication operation on the local texture features and the first sub-weight matrix along the channel dimension to generate weighted local texture features, and performing element-level multiplication operation on the global semantic features and the second sub-weight matrix along the channel dimension to generate weighted global semantic features; Concatenating the weighted local texture features and the weighted global semantic features along the channel dimension to generate a fusion feature tensor; Performing convolution operation on the fusion feature tensor to generate adaptive fusion features; The global information theory feature tensor containing Shannon entropy features and relative entropy features calculated based on the time-frequency image and the gating weight matrix generated based on the global information theory feature tensor comprise: Dividing the spatial dimension of the time-frequency image to generate a plurality of time-frequency sub-blocks; Calculating the Shannon entropy features of all time-frequency sub-blocks in each time-frequency region, and selecting a time-frequency sub-block at the center position in each time-frequency region as a reference time-frequency sub-block; Calculating the relative entropy features of the remaining time-frequency sub-blocks in the time-frequency region based on the reference time-frequency sub-block; Concatenating the Shannon entropy features and the relative entropy features of all time-frequency sub-blocks in all time-frequency regions to generate a global information theory feature tensor; Compressing the channel dimension of the global information theory feature tensor to generate a gating weight map; Adding the gating weight map to a preset gating base to generate a gating weight matrix.

2. The method of claim 1, wherein, The convolution operation performed on the fusion feature tensor to generate adaptive fusion features comprises: Taking the sum of the number of channels of the local texture features and the number of channels of the global semantic features as the number of convolution kernels, and constructing a convolution kernel set; Performing sliding window convolution operation on the fusion feature tensor using the convolution kernel set to generate a feature map set; Concatenating all feature maps in the feature map set along the channel dimension to generate adaptive fusion features.

3. The method of claim 2, wherein, The concatenation of all feature maps in the feature map set along the channel dimension to generate adaptive fusion features comprises: Calculating the channel statistics of each feature map in the feature map set to generate a channel statistics set; According to the channel statistics set, the distribution difference values of different feature maps are compared, and a channel selection priority parameter is determined, the channel selection priority parameter being a set of priority weight values for all channels; According to the channel selection priority parameter, the channel sequence corresponding to the feature map set is reordered to generate an ordered channel sequence; The channels corresponding to different priority weight values in the ordered channel sequence are alternately recombined to generate a recombined feature map set; All recombined feature maps in the recombined feature map set are spliced along the channel dimension to generate adaptive fusion features.

4. The method of claim 1, wherein, The use of a pre-constructed sparse sampling matrix to perform compressed sensing processing on the time-frequency image generates observation data, including: According to a pre-defined dimension sequence, the pre-constructed basis function vector is arranged and combined to generate a sparse sampling matrix; The pixel data of the time-frequency image is converted into a data vector in raster scan order; The dot product results of each measurement vector in the sparse sampling matrix and the data vector are calculated to generate a set of scalar elements; The elements in the set of scalar elements are combined in row index order to generate observation data.

5. The method of claim 1, wherein, The splicing of the Shannon entropy features and relative entropy features of all time-frequency subblocks in all time-frequency regions generates a global information theory feature tensor, including: The Shannon entropy features and relative entropy features of all time-frequency subblocks in each time-frequency region are spliced to generate a local information theory feature vector; The local information theory feature vectors corresponding to all time-frequency regions are aggregated to generate a global information theory feature tensor.

6. A radar signal classification system with multi-modal feature adaptive fusion, characterized in that, A radar signal classification method for performing adaptive fusion of multi-modal features according to any one of claims 1-5, comprising: A collection module for collecting a time-domain waveform of a radar signal, performing short-time Fourier transform on the time-domain waveform, and generating a time-frequency image; A compression module for performing compressed sensing processing on the time-frequency image using a pre-constructed sparse sampling matrix to generate observation data; An extraction module for inputting the observation data into a multi-branch feature extraction network, extracting local texture features through a first branch of the multi-branch feature extraction network, and extracting global semantic features through a second branch; A calculation module for calculating a global information theory feature tensor containing Shannon entropy features and relative entropy features based on the time-frequency image, and generating a gating weight matrix based on the global information theory feature tensor; A fusion module for performing adaptive fusion of the local texture features and global semantic features using the gating weight matrix to generate adaptive fusion features, and generating a classification result of the radar signal based on the adaptive fusion features.

7. A computing device, comprising: A processing component and a storage component are included; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a radar signal classification method for adaptive fusion of multi-modal features according to any one of claims 1-5.

8. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, a radar signal classification method for adaptive fusion of multi-modal features according to any one of claims 1-5 is implemented.

Images