Multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features

Abstract

This invention discloses a multi-view SAR target recognition method that combines scattering topological features and DCT frequency domain features. It solves the problems of insufficient target feature representation due to relying solely on single-view SAR images in existing technologies, as well as insufficient mining of scattering and frequency domain characteristics in SAR images. This invention extracts visual features, scattering topological features, and DCT frequency domain features from each image in a multi-view SAR image sequence, fully utilizing the scattering and frequency domain characteristics of SAR images. By fusing multi-modal features, it enhances the model's comprehensive representation ability of target features, resulting in more accurate classification and strong practical application value. The visual feature extraction module designed in this invention employs a more lightweight encoder, further improving target recognition performance while reducing model complexity.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically involving a multi-view SAR target recognition method that combines scattering topological features and DCT frequency domain features. Background Technology

[0002] Synthetic Aperture Radar (SAR) is a coherent imaging system operating in the microwave band. It actively transmits electromagnetic waves towards a target area and receives reflected information, thus achieving high-resolution imaging. Unlike passive sensing technologies such as optical and infrared sensors, SAR can operate stably for extended periods under various extreme weather conditions, unaffected by natural environmental factors such as clouds, fog, rain, snow, and darkness. It can also effectively distinguish camouflage and penetrate obscuring objects. With these advantages, SAR has unique application value in many fields, enabling it to perform functions that other remote sensing technologies cannot, and has become a focus of attention for countries worldwide.

[0003] Target recognition using SAR images has significant commercial value. Traditional SAR target recognition methods rely on manual interpretation of SAR images by professionals. This process not only demands high levels of expertise and theoretical knowledge but also suffers from low efficiency and strong subjectivity, making it difficult to meet the practical needs of rapid processing and real-time target recognition of large-scale SAR images. To address these issues, Automatic Target Recognition (SAR ATR) technology has emerged, primarily comprising three stages: target detection, target identification, and target recognition. The target detection stage identifies potential target areas; the target identification stage removes clutter and false alarms from the detection results; and the target recognition stage classifies and confirms the identified targets. Notably, SAR target recognition, as the final task of SAR ATR, has crucial classification performance. However, existing SAR target recognition algorithms are mainly designed for single-view SAR images. Due to the unique imaging principle of synthetic aperture radar, the same target will exhibit different visual characteristics under different observation conditions. Therefore, relying solely on a single viewpoint often fails to extract sufficient target feature information, and recognition performance is easily affected by changes in the observation angle. In contrast, multi-view SAR image sequences not only provide richer identification information for the same target from different perspectives, but also further improve identification performance by mining the inherent correlation between multi-view images.

[0004] In recent years, deep learning, with its powerful feature extraction capabilities, has gradually replaced traditional algorithms and become a research hotspot in the field of SAR target recognition. However, most existing SAR target recognition methods still primarily focus on the spatial domain of images, emphasizing the extraction of visual features, while paying insufficient attention to information such as scattering characteristics and frequency domain characteristics in SAR images. Since SAR images differ significantly from optical images in their imaging mechanisms, their imaging characteristics are closely related to the electromagnetic scattering behavior of the target. Analyzing SAR images solely from the perspective of spatial domain visual features often has limitations in characterizing target features, and there is still room for further improvement in target recognition performance. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a multi-view SAR target identification method that combines scattering topological features and DCT frequency domain features. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a multi-view SAR target identification method that combines scattering topological features and DCT frequency domain features, the method comprising: The original SAR images are sequenced to obtain a multi-view SAR image sequence; For each image in the multi-view SAR image sequence, the visual features, scattering topology features, and DCT frequency domain features corresponding to that image are extracted through the visual feature extraction module, the scattering topology feature extraction module, and the DCT frequency domain feature extraction module, respectively. Visual features, scattering topological features, and DCT frequency domain features are respectively passed through a multi-view feature learning module to model the correlation between different views under the same modality, thereby obtaining the multi-view feature representation corresponding to each modality; Features corresponding to each modality from the same perspective are concatenated according to dimensions and sent to the feature fusion module to integrate information from multiple modal features and obtain fused features. The dimensionality of the fused features is compressed using the feature dimensionality reduction module, and the target classification result is obtained using the target classification module.

[0006] In one embodiment of the present invention, the visual feature extraction module has a five-layer structure; wherein, the first four layers consist of convolutional layers, batch normalization (BN) layers, and pooling layers, used to extract visual features of the received image layer by layer and gradually compress the spatial size of the feature map; the fifth layer consists of convolutional layers and batch normalization (BN) layers; the visual feature extraction module selects the ReLU function as the activation function to enhance the nonlinear representation capability of the model.

[0007] In one embodiment of the present invention, the step of passing visual features, scattering topological features, and DCT frequency domain features through a multi-view feature learning module to model the correlation between different views under the same modality, and obtaining the multi-view feature representation corresponding to each modality, includes: The visual features corresponding to all viewpoints within the preset sequence length are concatenated to obtain the concatenated visual features. The azimuth information is embedded into the concatenated visual features using a preset position embedding formula to obtain the embedded visual features. The embedded visual features are then input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints and obtain the multi-view feature representation corresponding to the visual modality. The scattering topological features corresponding to all viewpoints within a preset sequence length are concatenated to obtain the concatenated scattering topological features. The azimuth information is embedded into the concatenated scattering topological features using a preset position embedding formula to obtain the embedded scattering topological features. The embedded scattering topological features are input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints and obtain the multi-view feature representation corresponding to the scattering topological mode. The DCT frequency domain features corresponding to all viewpoints within a preset sequence length are concatenated to obtain the concatenated DCT frequency domain features. Azimuth information is embedded into the concatenated DCT frequency domain features using a preset position embedding formula to obtain the embedded DCT frequency domain features. The embedded DCT frequency domain features are then input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints, thereby obtaining the multi-view feature representation corresponding to the frequency domain mode.

[0008] In one embodiment of the present invention, the expression of the preset position embedding formula is as follows: ; ; in, pos This indicates azimuth information, in degrees (°). d It is the dimension of the feature vector. i It is the dimension index in the feature vector.

[0009] In one embodiment of the present invention, the improved Transformer encoder based on a multi-head self-attention mechanism includes a two-layer structure; The first layer consists of a normalization layer, a multi-head self-attention layer, a random deactivation mechanism, and a first residual block arranged sequentially. The first residual block is used to implement feature interaction based on the attention mechanism. The second layer consists of a normalization layer, a multilayer perceptron, and a second residual block arranged sequentially. The multilayer perceptron is composed of a first fully connected layer, a Gelu activation function, a first random deactivation mechanism, a second fully connected layer, and a second random deactivation mechanism arranged sequentially. The second residual block is used to add the output of the first residual block to the output of the multilayer perceptron to obtain the multi-view feature representation corresponding to each modality.

[0010] In one embodiment of the present invention, the step of stitching together the viewpoint features corresponding to each modality under the same viewpoint according to dimensions and sending them to the feature fusion module to integrate information from multiple modal features to obtain fused features includes: The features corresponding to the three modalities under the same perspective are concatenated in the dimension to obtain the overall concatenated features; Nonlinear transformation and information integration are performed using fully connected layers and the PReLU activation function to obtain fused features.

[0011] In one embodiment of the present invention, the step of compressing the dimensionality of the fused features through a feature dimensionality reduction module and obtaining the target classification result through a target classification module includes: The dimensionality of the fused features is compressed by the feature dimensionality reduction module to obtain the feature representation for the classification task; The target classification module calculates the mean value of the feature representation along the sequence dimension to obtain the mean feature, and then performs a softmax operation on the mean feature to obtain the final classification result. The classification loss function used in the target classification module is... The expression is as follows:

[0012] in, Represents the cross-entropy loss function. This represents the relative weight of the regularization term. This represents the total number of regularization terms. Indicates the first i One regularization term.

[0013] In a second aspect, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features provided in the first aspect of the present invention.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features provided in the first aspect of the present invention.

[0015] The beneficial effects of this invention are: The solution provided in this invention extracts visual features, scattering topological features, and DCT frequency domain features from each image in a multi-view SAR image sequence. It fully utilizes the scattering and frequency domain characteristics of SAR images, and by fusing multi-modal features, enhances the model's comprehensive representation ability of target features, thereby making classification more accurate and possessing strong practical application value. The visual feature extraction module designed in this invention employs a more lightweight encoder, further improving target recognition performance while reducing model complexity. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the steps of a multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features provided in an embodiment of the present invention. Figure 2 This is a diagram of the overall network architecture in a multi-view SAR target recognition method that combines scattering topological features and DCT frequency domain features, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the construction of a two-view SAR image sequence according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the construction of a three-view SAR image sequence according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a visual feature extraction module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a GNN structure provided in an embodiment of the present invention; Figure 7 This is a flowchart of a frequency domain information preprocessing method provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a frequency domain channel selection module provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a frequency domain feature extraction network provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of an improved Transformer encoder based on a multi-head self-attention mechanism provided in an embodiment of the present invention. Figure 11This is a schematic diagram of 10 types of SAR image samples provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0018] This invention provides a multi-view SAR target identification method, electronic device, and storage medium that combines scattering topological features and DCT frequency domain features.

[0019] It should be noted that the execution subject of the multi-view SAR target identification method combining scattering topological features and DCT frequency domain features provided in this embodiment of the invention can be a device, which can run in an electronic device. This electronic device can be a server or a terminal device, but is not limited to these.

[0020] Below, we will first introduce a multi-view SAR target recognition method that combines scattering topological features and DCT frequency domain features, as provided in the embodiments of the present invention.

[0021] The present invention provides a multi-view SAR target recognition method that combines scattering topological features and DCT frequency domain features, such as... Figure 1 As shown, it may include the following steps: S1, Sequence construction is performed on the original SAR images to obtain a multi-view SAR image sequence; S2, for each image in the multi-view SAR image sequence, extract the corresponding visual features, scattering topology features and DCT frequency domain features of the image through the visual feature extraction module, the scattering topology feature extraction module and the DCT frequency domain feature extraction module respectively; S3 uses a multi-view feature learning module to model the correlation between different perspectives under the same modality by passing visual features, scattering topological features, and DCT frequency domain features respectively, thereby obtaining the multi-view feature representations corresponding to each modality. S4: The features corresponding to each modality under the same viewpoint are concatenated according to the dimension and sent to the feature fusion module to integrate the information of multiple modal features and obtain the fused features; S5 uses a feature dimensionality reduction module to compress the dimensionality of the fused features, and then uses a target classification module to obtain the target classification result.

[0022] The execution module of the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features provided in this embodiment of the invention mainly consists of the following six parts, such as... Figure 2As shown, the system can include: a multi-view SAR image sequence construction module, a single-view feature extraction module, a multi-view feature learning module, a feature fusion module, a feature dimensionality reduction module, and a target classification module. First, the multi-view SAR image sequence construction module constructs a sequence from the original SAR images, resulting in a multi-view SAR image sequence. Then, for each image in the multi-view SAR image sequence, the visual features, scattering topology features, and DCT frequency domain features corresponding to that image are extracted using the visual feature extraction module, scattering topology feature extraction module, and DCT frequency domain feature extraction module within the single-view feature extraction module. Based on this, the features of these three modalities are respectively processed through the multi-view feature learning module to model the correlation between different views within the same modality, thereby fully exploring the intrinsic relationships between features from different views. Subsequently, the corresponding visual features, scattering topology features, and DCT frequency domain features from the same viewpoint are concatenated dimensionally and fed into the feature fusion module to integrate information from multiple modalities and reduce feature redundancy. The fused features are further compressed in dimension by the feature dimensionality reduction module, and finally, the target classification module achieves target classification.

[0023] For step S1, the principle of sequence construction is: given an angle threshold and sequence length k Sequences are constructed using a sliding window with a step size of 1 in the original SAR image set. Specifically, a fixed-length window is set... k A sliding window of +1 slides across the original image set, and the images within the window can be combined to form... k A length of k The sequence is generated and filtered to retain only those images whose azimuth difference between any two images is less than a certain value. The sequence is used as the input sample for the network. Furthermore, it is required that the final retained sequence samples do not contain duplicate samples. Here... k Take 2 and 3 respectively. It is 45°.

[0024] To more intuitively illustrate the construction methods of multi-view SAR image sequences with different numbers of viewpoints, a schematic diagram of two-view SAR image sequence construction is shown below. Figure 3 As shown, through Figure 3 It can be clearly seen how, under a given angle threshold constraint, single-view images are combined into multi-view sequence samples through a sliding window method.

[0025] Building upon two-view sequence construction, as the sequence length increases further, the viewpoint information contained in a single sequence sample becomes even richer. A schematic diagram of three-view SAR image sequence construction is shown below. Figure 4 As shown, it can be seen from Figure 4The sequence construction method under the three perspectives can be observed.

[0026] For step S2, the single-view feature extraction module includes a visual feature extraction submodule, a scattering topology feature extraction submodule, and a DCT frequency domain feature extraction submodule, which are used to extract visual features, scattering topology features, and DCT frequency domain features from each image in the multi-view SAR image sequence, as detailed below: The visual feature extraction module extracts the visual features of each image in the multi-view SAR image sequence through a designed encoder, such as... Figure 5 As shown, the visual feature extraction module has a five-layer structure; the first four layers consist of convolutional layers, batch normalization (BN) layers, and pooling layers, which are used to extract the visual features of the received image layer by layer and gradually compress the spatial size of the feature map; the fifth layer consists of convolutional layers and batch normalization (BN) layers; the visual feature extraction module selects the ReLU function as the activation function to enhance the nonlinear representation capability of the model.

[0027] In the visual feature extraction module, the number of channels in each layer is set to 32, 64, 128, 256, and 512, respectively. The first two convolutional layers have a kernel size of 5×5 and a stride of 1; the middle two layers have a kernel size of 3×3 and a stride of 1; and the last layer has a kernel size of 4×4 and a stride of 1. Max pooling is used, with a kernel size of 2×2 and a stride of 2. Understandably, the visual feature extraction module employs a more lightweight encoder, further improving target recognition performance while reducing model complexity.

[0028] The scattering topology feature extraction module can extract the scattering topology features of each image from a multi-view SAR image sequence. The process is as follows: First, the SAR-SIFT algorithm is used to extract the scattering topology points of the SAR target. Since the SAR-SIFT algorithm extracts relatively few scattering topology points, a clustering-based strong scattering topology point detection algorithm is introduced to supplement more scattering topology points to obtain more comprehensive scattering information. The scattering points extracted by the two methods are complementary and can more comprehensively represent the geometric structure of the target. However, if the above algorithm is used directly for extraction, background areas are often misdetected as scattering topology points, thus affecting the subsequent construction of the scattering topology map. To reduce this interference, after the initial extraction, corresponding post-processing steps need to be designed based on the characteristics of the algorithm to remove invalid scattering topology points as much as possible, thereby improving the accuracy and reliability of topology modeling. Subsequently, the connection relationship is defined according to the similarity between the scattering topology point features, thereby transforming the point set into a graph structure, which is then fed into a graph neural network (GNN) to extract the scattering topology features of the SAR target.

[0029] Specifically, the steps include the following: Scattering topology point extraction based on SAR-SIFT algorithm: Scattering topology points are extracted using the SAR-SIFT algorithm, which can be divided into two main parts: keypoint detection and descriptor generation. Since the SAR-SIFT algorithm extracts relatively few keypoints, a clustering-based strong scattering point detection algorithm is introduced into the keypoint detection stage of the SAR-SIFT algorithm to expand the number of keypoints and more comprehensively and accurately describe the target's scattering topology. The scattering points extracted by these two algorithms are complementary, resulting in a better description of the target's geometry.

[0030] Since keypoints are extracted across the entire image, they inevitably contain a large number of invalid points from the background region. To ensure the accuracy of subsequent image construction, the initially obtained keypoints need to undergo two rounds of filtering to eliminate false positives as much as possible. The first round of filtering is based on pixel amplitude: First, the amplitude histogram of the entire SAR image is statistically analyzed and accumulated to obtain the cumulative amplitude distribution function. Then, keypoints with amplitudes in the last 1% of the cumulative distribution function are considered scattering topological points on the target, while keypoints in the remaining 99% are considered invalid and removed. After completing the first round of filtering, a second round of filtering based on coordinates is used: First, the centroid of the remaining keypoints is calculated. Then, keypoints with a distance greater than a threshold from the centroid are deleted. The key point.

[0031] After filtering out invalid keypoints located in the background region, the remaining steps of the SAR-SIFT algorithm are performed on the valid keypoints located on the target (i.e., scattering topology points), generating a 108-dimensional ratio descriptor for each point. This is used to characterize the gradient information in its neighborhood. Assume the total number of scattering topological points detected from the image is... m Then a target can be represented as a set of scattering topological points in a 108-dimensional space. ,in This represents the ratio descriptor for the corresponding scattering topological point.

[0032] Scattering topology graph construction: After obtaining the scattering topology point set from each image, an undirected, unweighted graph is used to model its scattering topology. The graph data mainly consists of a set of nodes. Sum of edges E Composition, commonly used adjacency matrix U To store edge information, where This indicates the existence of an edge from the i-th node to the j-th node. Assume... For a containing mA set of scattering topological points. Since the scattering topological points themselves are unordered, it is only necessary to define the connectivity between the scattering topological points to construct the scattering topological map of the target.

[0033] To describe this connectivity, the cosine similarity between nodes is used to determine whether a connection should be established, as defined below: ; in, Represents the dot product of vectors. This is a similarity threshold used to filter out redundant edges. After extracting the target's scattering topology points from the SAR image using a SAR-SIFT-based scattering topology point extraction algorithm, a graph G=( containing the target's local scattering characteristics and global topology structure is generated.) V , E The local scattering characteristics are reflected in the nodal features. v In the context of topology, the global topology is reflected in the adjacency matrix. U middle.

[0034] Scattering topological feature extraction: After obtaining the scattering topology map, a graph neural network (GNN) can extract features from the map that simultaneously encode the target's local scattering characteristics and spatial topology; these are the scattering topology features. Mathematically, for any graph G, the computation process of the graph neural network can be represented as: ; in, For the parameters of the graph neural network, This represents a nonlinear mapping process. The extracted scattering topological features.

[0035] Similar to CNNs, which construct their network structure by stacking multiple convolutional layers, GNNs are typically built by layering multiple graph attention layers. Considering that excessively deep GNNs can lead to oversmoothing of node features during training, this embodiment of the invention uses only two graph attention layers (GAT). The overall GNN, as... Figure 6As shown, the system consists of two graph attention layers, two activation layers, and one readout layer. Each of the two graph attention layers contains four attention heads. Each attention head maps the input features to F'=64 dimensions and concatenates the output features of the four attention heads to obtain a 4×64=256-dimensional node feature output. Specifically, in the first graph attention layer, each attention head maps the input node features from 108 dimensions to 64 dimensions, and then concatenates the output features of the four attention heads to obtain a 256-dimensional node feature representation. In the second graph attention layer, each attention head maps the node features from 256 dimensions to 64 dimensions, and then concatenates the output features of the four attention heads to obtain a 256-dimensional node feature representation. Both activation layers use the ELU activation function. The readout layer uses global max pooling as the readout function, performing element-wise maximum aggregation on the features of all nodes in the graph to generate a fixed-length graph-level feature representation, thus obtaining the final scattering topological features. .

[0036] The DCT frequency domain feature extraction module can extract the DCT frequency domain features of each image from a multi-view SAR image sequence. The process is as follows: First, the input image undergoes frequency domain information preprocessing to construct the DCT frequency domain representation of the image. Then, the DCT frequency domain representation is input into the frequency domain feature extraction network to extract the DCT frequency domain features of each image. Specifically, the steps include: Frequency domain information preprocessing: Frequency domain information preprocessing flowchart, as follows Figure 7 As shown in the figure, the two-dimensional discrete cosine transform (2D-DCT) is illustrated using 2×2 blocks. First, a 64×64 SAR image is subjected to an 8×8 block 2D-DCT, transforming the image from the spatial domain to the frequency domain. Specifically, the 64×64 image is divided into non-overlapping 8×8 sub-blocks (64 in total), and a 2D-DCT is performed on each sub-block, mapping the pixel intensity in the spatial domain to the corresponding frequency domain coefficients, thus obtaining the representation of each sub-block at different frequency components. Subsequently, a frequency domain reshaping operation is performed, that is, the DCT coefficients of all 8×8 sub-blocks are indexed according to the frequency domain (…). u , v The DCT coefficients with the same frequency domain position are reorganized into the same channel and rearranged into feature maps on an 8×8 spatial grid, ultimately forming a 3D DCT cube representation of size 8×8×64, where 64 represents the number of frequency domain channels. Next, frequency domain channels are selected, retaining those that significantly contribute to target recognition to reduce redundant information. Finally, the selected frequency domain channels are normalized to obtain a stable DCT frequency domain representation, which serves as the input to the subsequent frequency domain feature extraction network. A schematic diagram of the frequency domain channel selection module is shown below. Figure 8As shown, after frequency domain reshaping, the input frequency domain feature tensor 1 can be expressed as: ,in C This refers to the number of frequency domain channels. To achieve adaptive selection of frequency domain channels, this embodiment of the invention designs a channel-level gating module. First, for... Global Average Pooling (GAP) is performed to compress the spatial dimensions, resulting in a size of... Tensor 2. Then, a 1×1 convolution is used to map tensor 2 to obtain a tensor of size . Tensor 3 is used to enhance the expressive power of gating decisions. Based on this, a channel-level binary classification gating head is introduced to generate two-class scores ("off" / "on") for each frequency domain channel, mapping Tensor 3 to a tensor of size 3. Tensor 4 is used. Then, the Gumbel-Softmax mechanism is set to hard mode, and the binary classification scores are sampled in a differentiable manner to obtain an approximate one-hot selection vector of length 2 for each channel. The "on" component is then used to construct the channel selection mask. In forward propagation, m This is a strict 0 / 1 channel switch. Then, the channel selection mask is multiplied channel-by-channel by the original frequency domain feature tensor to obtain the output features. This refers to tensor 5 in the figure. Finally, batch normalization is performed on tensor 5 to obtain a stable DCT frequency domain representation. This serves as the input for the subsequent frequency domain feature extraction network. Furthermore, since frequency domain channel selection is a discrete binary decision, direct binary sampling is not differentiable. Therefore, during backpropagation, gradients are passed through a continuous relaxation form of Gumbel-Softmax, thus achieving end-to-end training.

[0037] Frequency domain feature extraction: A stable DCT frequency domain representation is obtained. Then, the DCT frequency domain features of each image in the multi-view SAR image sequence are extracted using a lightweight encoder as shown in the schematic diagram of the frequency domain feature extraction network. Figure 9 As shown, the designed encoder has three layers, with the number of channels in each layer set to 64, 128, and 256 respectively. The spatial size of the feature map is progressively compressed through convolutional operations with a stride of 2. Specifically, the first convolutional layer uses a 5×5 kernel with a stride of 2; the second layer uses a 3×3 kernel with a stride of 2; and the third layer uses a 2×2 kernel with a stride of 2. The ReLU activation function is chosen to enhance the model's non-linear representation capability.

[0038] For step S3, it may include: The visual features corresponding to all viewpoints within the preset sequence length are concatenated to obtain the concatenated visual features. The azimuth information is embedded into the concatenated visual features using a preset position embedding formula to obtain the embedded visual features. The embedded visual features are then input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints and obtain the multi-view feature representation corresponding to the visual modality. The scattering topological features corresponding to all viewpoints within a preset sequence length are concatenated to obtain the concatenated scattering topological features. The azimuth information is embedded into the concatenated scattering topological features using a preset position embedding formula to obtain the embedded scattering topological features. The embedded scattering topological features are input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints and obtain the multi-view feature representation corresponding to the scattering topological mode. The DCT frequency domain features corresponding to all viewpoints within a preset sequence length are concatenated to obtain the concatenated DCT frequency domain features. Azimuth information is embedded into the concatenated DCT frequency domain features using a preset position embedding formula to obtain the embedded DCT frequency domain features. The embedded DCT frequency domain features are then input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints, thereby obtaining the multi-view feature representation corresponding to the frequency domain mode.

[0039] The multi-view feature learning module is used to model the correlation between features from different perspectives within the same modality. For three modalities—visual features, scattering topological features, and frequency domain features—multi-view feature learning modules with identical structures but independent parameters are introduced to model the feature correlations of each modality under different perspectives, thereby fully exploring the intrinsic relationships between multi-view features.

[0040] After feature extraction from each image in a multi-view SAR image sequence, a multi-view feature learning module is needed to jointly model the features from different viewpoints, thereby learning the intrinsic relationships between features from different viewpoints. Let the th... i The feature vectors corresponding to each viewpoint are ,in d Let be the feature dimension, and let the sequence length be denoted as . k By concatenating all viewpoint features within the sequence, we can obtain... Next, location embedding is required. This is necessary because the self-attention mechanism itself does not contain sequence information, while in practical tasks, the order of input features often affects the model's judgment results. For multi-view SAR images, the images from each view constitute a sequence according to their azimuth angles, and the azimuth angle information between different views also needs to be perceived by the model. Therefore, azimuth angle information needs to be embedded here, constructed using sine and cosine functions. The preset expression for the location embedding formula is as follows: ; ; in, pos This indicates azimuth information, in degrees (°). d It is the dimension of the feature vector. i It is the dimension index in the feature vector.

[0041] After introducing positional embedding, the multi-view feature sequences are fed into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationships between features from different viewpoints. The improved Transformer encoder based on a multi-head self-attention mechanism is as follows: Figure 10 As shown, it can include a two-layer structure; The first layer consists of a normalization layer, a multi-head self-attention layer, a random deactivation mechanism, and a first residual block arranged sequentially. The first residual block is used to implement feature interaction based on the attention mechanism. The second layer consists of a normalization layer, a multilayer perceptron, and a second residual block arranged sequentially. The multilayer perceptron is composed of a first fully connected layer, a Gelu activation function, a first random deactivation mechanism, a second fully connected layer, and a second random deactivation mechanism arranged sequentially. The second residual block is used to add the output of the first residual block to the output of the multilayer perceptron to obtain the multi-view feature representation corresponding to each modality.

[0042] Understandably, the improved Transformer encoder consists of multiple layers, each containing two residual sub-modules, primarily including a Multi-Head Self-Attention (MSA) unit and a Multi-Layer Perceptron (MLP) unit. In the improved Transformer encoder, the first residual block is mainly used to implement feature interaction based on the attention mechanism. This structure, based on the input features, generates output features through residual connections after layer normalization (LN) and multi-head self-attention (MSA) processing.

[0043] The second residual block employs a feedforward neural network structure to further nonlinearly transform and enhance the features. This residual block is obtained by adding the input features to the result of processing by a Multi-Layer Perceptron (MLP). In the specific computation process, this residual block uses the output of the first residual block as its input, sequentially passing through a Layer Normalization (LN), a fully connected layer with a nonlinear activation function, and another fully connected layer. The first fully connected layer maps the features to a higher-dimensional space to enhance the model's nonlinear expressive power; the second fully connected layer remaps the feature dimension back to the original space, thus ensuring the consistency of the network structure. To further improve the model's representation performance, a nonlinear activation function is introduced into the feedforward network; here, GELU is used as the activation function. After each fully connected layer in the feedforward neural network, a dropout mechanism is introduced to alleviate overfitting problems that may occur when the number of training samples is limited. Finally, the second residual block is obtained through residual connections. The output is used as the multi-view feature representation for each modality.

[0044] For step S4, it may include: The features corresponding to the three modalities under the same perspective are concatenated in the dimension to obtain the overall concatenated features; Nonlinear transformation and information integration are performed using fully connected layers and the PReLU activation function to obtain fused features.

[0045] After performing multi-view feature learning based on a self-attention mechanism on visual features, scattering topological features, and DCT frequency domain features, a feature representation containing multi-view correlation information can be obtained. Let the output visual features be... The scattering topological characteristics are The frequency domain feature dimension is .in, k 512 represents the number of viewpoints, and 256 represents the feature dimensions. To achieve multimodal feature fusion, features from three modalities under the same viewpoint are concatenated along their feature dimensions to obtain... It can be described as: ; Subsequently, nonlinear transformation and information integration of multimodal features are performed using fully connected layers and the PReLU activation function to enhance the comprehensive representation capability of multimodal features, resulting in: ; in, and These represent the parameters and the nonlinear mapping process, respectively.

[0046] For step S5, it may include: The dimensionality of the fused features is compressed by the feature dimensionality reduction module to obtain the feature representation for the classification task; The target classification module calculates the mean value of the feature representation along the sequence dimension to obtain the mean feature, and then performs a softmax operation on the mean feature to obtain the final classification result. The classification loss function used in the target classification module is... The expression is as follows: ; in, Represents the cross-entropy loss function. This represents the relative weight of the regularization term. This represents the total number of regularization terms. Indicates the first i One regularization term.

[0047] The fused features still have high dimensionality. To adapt to subsequent classification tasks and reduce model complexity, a feature dimensionality reduction module is needed to compress the dimensionality of the fused features. This embodiment of the invention selects to use... Convolution is used for dimensionality reduction. This convolution is achieved through... The number of convolution kernels is used to adjust the dimensionality of the output features. Specifically, before performing the convolution operation, the fused features are first processed... Perform dimensional transformation to obtain Subsequently, through Convolutional layers output features with dimensions of 1. ,in C This represents the number of categories. Let the output of the convolutional layer be... , yes The convolution kernel of the convolutional layer, For bias, the dimensionality reduction process is as follows: ; After dimensionality reduction, for Perform dimensional transformation for subsequent classification: .

[0048] After completing multi-view feature learning and dimensionality reduction, the model obtains feature representations for classification tasks. ,in k Represents the number of viewpoints. C This represents the number of target categories. To map the sequential features into fixed-length classification vectors, it is necessary to aggregate the features from multiple perspectives.

[0049] In multi-view target recognition tasks, features from different perspectives characterize the discriminative information of the same target from multiple angles, and these features play a complementary role in classification decisions. To synthesize the contributions of features from various perspectives to the classification results, [the following is a separate, unrelated section:] Calculate the mean along the sequence dimension to obtain .

[0050] Finally, for The final classification result is obtained by performing a softmax operation. .

[0051] To measure the difference between the model's predicted results and the true labels, this embodiment of the invention uses the cross-entropy loss function as the classification loss, defined as follows: ; in, n Indicates the number of samples. Represents the true category label (using one-hot encoding). This represents the class probability vector predicted by the model. This represents the vector dot product operation.

[0052] To constrain channel selection in the frequency domain feature extraction module, a regularization term is introduced to balance the number of selected channels. Let... The channel before channel selection. This indicates that the proposed gating module's output satisfies .when When this occurs, it indicates that the corresponding channel is reserved, and the calculation process can be expressed as follows: ; in, This indicates element-wise multiplication.

[0053] Building upon this, a regularization term is introduced into the classification loss function to constrain the number of channels selected. This regularization term is minimized along with the classification loss. The final overall classification loss function... Defined as: ; in, It is a hyperparameter that represents the relative weight of the regularization term.

[0054] Experiments and Analysis: Original dataset: This invention embodiment was tested on 10 classes of MSTAR datasets under Standard Operating Condition (SOC) conditions, with targets including 2S1, BRDM2, BTR60, D7, T72, BMP2, BTR70, T62, ZIL131, and ZSU234. The radar operated in the X-band, using HH polarization, and the size of a single image was [missing information]. The data has a resolution of 0.3 × 0.3 m. The azimuth coverage ranges from 0° to 360°, with angular intervals of approximately 5° to 6°. The elevation angles for the training and test samples are 17° and 15°, respectively. A schematic diagram of 10 types of SAR image samples is shown below. Figure 11 As shown.

[0055] Two different training sample sampling ratios were designed in the experiment, denoted as Experiment 1 and Experiment 2. While keeping the test set unchanged, Experiment 1 randomly sampled 20% of each class from the original training set to construct the training set, while Experiment 2 randomly sampled 30% of each class from the original training set. After random sampling, the samples of each class were sorted according to their azimuth information to construct multi-view sequence data.

[0056] Under the conditions of Experiment 1, the constructed MSTAR ten-class dataset is shown in Table 1. The training samples consist of 551 images, and the test samples consist of 2425 images.

[0057] Table 1. Statistical analysis of samples from the MSTAR ten-class dataset under Experimental Setup 1

[0058] Under experimental setup 2, the constructed MSTAR ten-class dataset is shown in Table 2. The training samples consist of 826 images, and the test samples consist of 2425 images.

[0059] Table 2. Statistical analysis of samples from the MSTAR ten-class dataset under Experimental Setup 2

[0060] Multi-view dataset: Given an angle threshold and sequence length k Sequences are constructed using a sliding window with a step size of 1 on the two previously constructed datasets with different sampling ratios. Specifically, a fixed-length window is set... k A sliding window of +1 slides across the original image set, and the images within the window can be combined to form... k A length of kThe sequence is generated and filtered to retain only those images whose azimuth difference between any two images is less than a certain value. The sequence is used as the input sample for the network. Furthermore, it is required that the final retained sequence samples do not contain duplicate samples. An angle threshold is used in this embodiment of the invention. Set to 45°, sequence length k We selected 2 and 3 respectively. Based on the above settings, we constructed a 2-view sequence dataset and a 3-view sequence dataset for subsequent experiments.

[0061] Under Experiment 1, based on the constructed MSTAR ten-class dataset, a multi-view sequence dataset was further generated, and the statistical results are shown in Table 3. Among them, there are 1047 two-view training sequences and 4598 test sequences; there are 1511 three-view training sequences and 6521 test sequences.

[0062] Table 3. Statistics on the number of samples in the multi-view sequence dataset under Experimental Setup 1

[0063] Under Experiment 2, based on the constructed MSTAR ten-class dataset, a multi-view sequence dataset was further generated, and the statistical results are shown in Table 4. Among them, there are 1602 two-view training sequences and 4598 test sequences; there are 2360 three-view training sequences and 6521 test sequences.

[0064] Table 4. Statistics on the number of samples in the multi-view sequence dataset under Experimental Setup 2

[0065] Identification Results and Comparative Analysis: The following table shows the comparison results of the recognition accuracy of the method proposed in the embodiments of the present invention and several comparative methods under two experimental settings. The comparative methods include: "Multi-Aspect Convolutional-TransformerNetwork for SAR Automatic Target Recognition (hereinafter referred to as MACTN for ease of comparison)"; and "Multi-View Automatic Target Recognition using Joint Sparse Representation (JSRC)". Tables 5 and 6 show the recognition accuracy of the method provided in the embodiments of the present invention and existing methods under experimental settings 1 and 2. As shown in Tables 5 and 6, the method proposed in the present invention is significantly superior to existing methods, demonstrating the effectiveness of the method proposed in the embodiments of the present invention.

[0066] Table 5 shows the recognition accuracy of each method under Experimental Setup 1.

[0067] Table 6 shows the recognition accuracy of each method under Experimental Setup 2.

[0068] Understandably, this invention proposes a multi-view SAR target recognition method that combines scattering topological features and DCT frequency domain features. This addresses the problems in existing technologies where relying solely on single-view SAR images leads to insufficient target feature representation and inadequate mining of scattering and frequency domain characteristics in SAR images. This multi-view SAR target recognition method extracts visual features, scattering topological features, and DCT frequency domain features from each image in a multi-view SAR image sequence, fully utilizing the scattering and frequency domain characteristics of SAR images. By fusing multi-modal features, it enhances the model's comprehensive representation ability of target features, resulting in more accurate classification and strong practical application value. The visual feature extraction module designed in this invention employs a more lightweight encoder, further improving target recognition performance while reducing model complexity.

[0069] Secondly, embodiments of the present invention also provide an electronic device, such as... Figure 12 As shown, it includes a processor 001, a communication interface 002, a memory 003, and a communication bus 004, wherein the processor 001, the communication interface 002, and the memory 003 communicate with each other through the communication bus 004. The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of any of the multi-view SAR target recognition methods combining scattering topological features and DCT frequency domain features provided in the first aspect of the present invention.

[0070] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0071] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0072] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0073] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0074] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0075] Thirdly, corresponding to the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features provided in the first aspect, the present invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements any of the steps of the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features provided in the first aspect of the present invention.

[0076] For the embodiments of the device / electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0077] It should be noted that the electronic device and storage medium in the embodiments of the present invention are respectively electronic devices and storage media that apply the above-mentioned multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features. Therefore, all embodiments of the above method are applicable to the electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0078] It should be noted that, in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features, characterized in that, include: The original SAR images are sequenced to obtain a multi-view SAR image sequence; For each image in the multi-view SAR image sequence, the visual features, scattering topology features, and DCT frequency domain features corresponding to that image are extracted through the visual feature extraction module, the scattering topology feature extraction module, and the DCT frequency domain feature extraction module, respectively. Visual features, scattering topological features, and DCT frequency domain features are respectively passed through a multi-view feature learning module to model the correlation between different views under the same modality, thereby obtaining the multi-view feature representation corresponding to each modality; Features corresponding to each modality from the same perspective are concatenated according to dimensions and sent to the feature fusion module to integrate information from multiple modal features and obtain fused features. The dimensionality of the fused features is compressed using the feature dimensionality reduction module, and the target classification result is obtained using the target classification module.

2. The multi-aspect SAR target recognition method combining scattering topological features and DCT frequency domain features according to claim 1, characterized in that, The visual feature extraction module has a five-layer structure; the first four layers consist of convolutional layers, batch normalization (BN) layers, and pooling layers, which are used to extract visual features from the received image layer by layer and gradually compress the spatial size of the feature map; the fifth layer consists of convolutional layers and BN layers; the visual feature extraction module selects the ReLU function as the activation function to enhance the nonlinear representation capability of the model.

3. The multi-aspect SAR target recognition method combining scattering topological features and DCT frequency domain features according to claim 2, characterized in that, The number of channels in each layer of the visual feature extraction module is set to 32, 64, 128, 256 and 512 respectively.

4. The multi-aspect SAR target recognition method combining scattering topological features and DCT frequency domain features according to claim 1, characterized in that, The step involves using a multi-view feature learning module to model the correlation between different viewpoints within the same modality, passing visual features, scattering topological features, and DCT frequency domain features respectively, to obtain multi-view feature representations corresponding to each modality. This includes: The visual features corresponding to all viewpoints within the preset sequence length are concatenated to obtain the concatenated visual features. The azimuth information is embedded into the concatenated visual features using a preset position embedding formula to obtain the embedded visual features. The embedded visual features are then input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints and obtain the multi-view feature representation corresponding to the visual modality. The scattering topological features corresponding to all viewpoints within a preset sequence length are concatenated to obtain the concatenated scattering topological features. The azimuth information is embedded into the concatenated scattering topological features using a preset position embedding formula to obtain the embedded scattering topological features. The embedded scattering topological features are input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints and obtain the multi-view feature representation corresponding to the scattering topological mode. The DCT frequency domain features corresponding to all viewpoints within a preset sequence length are concatenated to obtain the concatenated DCT frequency domain features. Azimuth information is embedded into the concatenated DCT frequency domain features using a preset position embedding formula to obtain the embedded DCT frequency domain features. The embedded DCT frequency domain features are then input into an improved Transformer encoder based on a multi-head self-attention mechanism to learn the intrinsic relationship between features from different viewpoints, thereby obtaining the multi-view feature representation corresponding to the frequency domain mode.

5. The multi-view SAR target identification method combining scattering topological features and frequency domain features according to claim 4, characterized in that, The expression for the preset position embedding formula is as follows: ； ； in, pos This indicates azimuth information, in degrees (°). d It is the dimension of the feature vector. i It is the dimension index in the feature vector.

6. The multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features according to claim 4, characterized in that, The improved Transformer encoder based on a multi-head self-attention mechanism includes a two-layer structure; The first layer consists of a normalization layer, a multi-head self-attention layer, a random deactivation mechanism, and a first residual block arranged sequentially. The first residual block is used to implement feature interaction based on the attention mechanism. The second layer consists of a normalization layer, a multilayer perceptron, and a second residual block arranged sequentially. The multilayer perceptron is composed of a first fully connected layer, a Gelu activation function, a first random deactivation mechanism, a second fully connected layer, and a second random deactivation mechanism arranged sequentially. The second residual block is used to add the output of the first residual block to the output of the multilayer perceptron to obtain the multi-view feature representation corresponding to each modality.

7. The multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features according to claim 1, characterized in that, The process involves concatenating features corresponding to different modalities from the same perspective according to their dimensions and sending them to a feature fusion module. This module integrates information from multiple modal features to obtain fused features, including: The features corresponding to the three modalities under the same perspective are concatenated in the dimension to obtain the overall concatenated features; Nonlinear transformation and information integration are performed using fully connected layers and the PReLU activation function to obtain fused features.

8. The multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features according to claim 1, characterized in that, The step of compressing the dimensionality of the fused features through a feature dimensionality reduction module and obtaining the target classification result through a target classification module includes: The dimensionality of the fused features is compressed by the feature dimensionality reduction module to obtain the feature representation for the classification task; The target classification module calculates the mean value of the feature representation along the sequence dimension to obtain the mean feature, and then performs a softmax operation on the mean feature to obtain the final classification result. The classification loss function used in the target classification module is... The expression is as follows: ； in, Represents the cross-entropy loss function. This represents the relative weight of the regularization term. This represents the total number of regularization terms. Indicates the first i One regularization term.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the multi-view SAR target recognition method combining scattering topological features and DCT frequency domain features as described in any one of claims 1-8.

Images