A SAR image target classification system based on a complex domain feature map network

Abstract

The present invention discloses a SAR image target classification system based on a complex domain feature map network, comprising the following steps: dividing the SAR dataset into a training set, a validation set, and a test set according to a division ratio, and performing data augmentation on the training set; dividing the SAR images into a support set and a query set; respectively obtaining the complex domain features of the SAR through a complex domain transformation feature extraction module for the support set and the query set; using the graph network in the system to update and obtain the support set image most similar to the query set image; printing out the final classification accuracy result in a display window. The present invention relates to the field of image processing. This kind of SAR image target classification system based on a complex domain feature map network solves the problem that most existing methods focus on feature extraction and analysis in the spatial domain or time-frequency domain, while ignoring the potential advantages of the complex domain in SAR image classification, and improves the classification accuracy of SAR images.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and specifically to a SAR image target classification system based on a complex domain feature map network. Background Art

[0002] Synthetic Aperture Radar (SAR) is a system for ground imaging based on radar technology. Compared with traditional optical imaging technology, SAR has many advantages that are difficult to quantify. It has the unique ability to observe the Earth all-weather and all-day long, and is widely used in many fields such as military, ocean, communication, geological exploration, and disaster monitoring. Studying the complex domain features of SAR images helps to better cope with interference, enrich the feature dimensions, accurately capture changes, and master the details of changes.

[0003] Most traditional image classification methods focus on feature extraction and analysis based on the spatial domain or time-frequency domain, which often destroys the internal connection between the real part and the imaginary part of SAR images and ignores the potential advantages of the complex domain in SAR image classification. SAR images are essentially complex-valued, which means that each pixel can be represented by a complex number. A complex number consists of two different parts, the real part and the imaginary part, which provide different information about the target object in different ways. The complex domain method can improve the classification accuracy and robustness by comprehensively considering the intensity and phase information, and provide a more refined feature representation than the spatial domain or time-frequency domain. Therefore, studying the complex domain of SAR images is crucial for the analysis of SAR images. Summary of the Invention

[0004] (1) Technical Problems to be Solved

[0005] Aiming at the deficiencies of the prior art, the present invention provides a SAR image target classification system based on a complex domain feature map network, which solves the limitation that the currently studied SAR images only target the real domain and improves the classification accuracy of SAR images.

[0006] (2) Technical Solutions

[0007] To achieve the above object, the present invention is realized through the following technical solutions: A SAR image target classification system based on a complex domain feature map network, including the following steps:

[0008] S1: Divide the SAR dataset into a training set, a validation set, and a test set according to the division ratio, and perform data augmentation on the training set;

[0009] S2: Divide the SAR images into a support set and a query set;

[0010] S3: Respectively obtain the complex domain features of SAR through the complex domain transformation feature extraction module for the support set and the query set;

[0011] S4: Use the graph network in the system to train the network model to obtain the support set image that is most similar to the query set image;

[0012] S5: Print the final classification accuracy result in the display window;

[0013] Among them, the optimal model training described in step S3 includes the following steps:

[0014] The first step, complex domain transformation: Respectively pass the support set and the query set through the Hilbert transform. The goal of the Hilbert transform is to find the adjoint function y(t) for the real function x(t), so that the complex function z(t)=x(t)+iy(t) can be analytic on the real axis t∈R and extended to the upper half of the complex plane. By convolving the signal, through this transformation, a complex-valued signal can be generated from the real-valued signal I(x). Therefore, through the Hilbert transform, the complex domain images of the support set and the query set can be obtained;

[0015] The second step, feature extraction: Extract the complex domain features from the complex domain image through the ResNet12 backbone network. It is divided into 4 stages, where the first stage is the processing of the input, and the input is It successively passes through the convolutional layer, BN layer, LeakyReLU activation function, and max pooling layer to obtain The output features of the last stage That is the output feature of the backbone network, where B is the size of each batch.

[0016] As an optimization, the specific steps of training the network model by the graph network described in step S4 are as follows:

[0017] (1). Construct a complex domain node initialization and update module: Extract the complex domain node features through the backbone network. A total of 5 nodes are composed of 4 support set samples and 1 query set sample for node initialization. Let v i be the node feature of V i , V i ={v i |i = 1, 2,..., N×K + T}, which is 1 when the support set true label is the same as the query set predicted label, and 0 otherwise. Given the node and edge of the (l - 1)-th layer, update the node features through the process of neighborhood aggregation. The features of the nodes in the l-th layer are updated by aggregating the features of other nodes together. The aggregation method is proportional to the edge features of the nodes, and then the aggregated features are transformed to update the features of the nodes in the l-th layer The node features of the (l - 1)-th layer Update the node features of the l-th layer The edge features of the (l-1)-th layer are used to quantify the contribution degree of the corresponding neighboring nodes

[0018] (2) Construct a hybrid edge initialization and update module: Nodes effectively capture the similarity information between each other through similarity measurement edges, thereby generating preliminary edge features. The similarity measurement edges and relative measurement edges are concatenated to form the hybrid edges of the nodes, thus constituting the hybrid edges of the graph. The similarity measurement edges can only reflect the similarity on node features, while the relative measurement edges encode the relative positions or topological relationships between nodes. Nodes perform feature aggregation with their adjacent nodes. This dynamic feature fusion process enables the hybrid edges to more comprehensively reflect the multi-dimensional relationships between nodes and can effectively distinguish nodes with similar features but different categories. Denote the relative measurement edge as e ij1 represents the edge feature of the first dimension between node i and node j. Denote the similarity measurement edge as e ij2 , and denote the dissimilarity measurement edge as e ij3 , initialize the relative measurement edge and combine it with the similar structure and dissimilar structure between nodes to initialize the edges of the graph After the node features are updated, new node features will be obtained, thereby realizing the update of the hybrid edge features

[0019] As an optimization, the update of the hybrid edge features constructs the hybrid edges of the graph in two ways: similarity measurement edges and relative measurement edges. The relative measurement edges are measured and encoded by calculating the distance between misaligned nodes, while the similarity edges are encoded by directly measuring the distance between nodes. After the hybrid edge update is completed, the nodes will be classified according to the features of the points.

[0020] As an optimization, in step S1, the SAR dataset includes four categories: ships, airplanes, oil tanks, and bridges. The proportion of the number of target pictures varies greatly. Therefore, data augmentation is performed on the training set to balance the number of samples of different categories in the dataset.

[0021] As an optimization, the methods of data augmentation in step S1 include rotating the image, horizontally flipping the image, vertically flipping the image, and random cropping.

[0022] As an optimization, the angle of rotating the image is randomly selected between -45 and 45 degrees.

[0023] As an optimization, in step S5, the classification accuracy result passes the label result of jointly predicting the query set with the support set true label and the hybrid edges

[0024] (III) Advantageous Effects

[0025] The present invention provides a SAR image target classification system based on a complex domain feature map network, which has the following advantageous effects:

[0026] By studying the complex domain features of SAR images, the present invention can extract more refined feature representations than the spatial domain or the time-frequency domain. Using deep learning methods, the SAR image recognition work is intelligentized. Good recognition results can be obtained through low-cost complex domain transformation feature extraction and hybrid edge distance similarity calculation, solving the problem that traditional recognition methods are based on manually designed features and it is difficult to fully exploit the complex target information in SAR images. For some subtle and representative features, they may not be effectively extracted, resulting in limited recognition accuracy. To a certain extent, this invention can detect and identify military targets and monitor the battlefield environment, making certain contributions to the military field. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flowchart of the system of the present invention;

[0028] Figure 2 is a framework diagram of the model training of the present invention;

[0029] Figure 3 is a framework diagram of the complex domain feature extraction module and the graph network module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0031] Please refer to Figures 1-3 , the present invention provides a technical solution:

[0032] As Figure 2 shown, the model training and retrieval are completed in the following steps:

[0033] The first step, dataset preprocessing: Divide the SAR images into a training set, a validation set, and a test set according to a ratio. Perform data augmentation on the training set. The main methods include rotating the image, horizontally flipping the image, vertically flipping the image, and random cropping. Among them, the rotation angle of the image is randomly selected between -45 and 45 degrees. Each dataset is further divided into a support set and a query set.

[0034] The second step, model training: Construct a complex domain transformation feature extraction module, and use the graph network to train the network model to finally achieve classification.

[0035] The complex domain transformation feature extraction module includes the following steps:

[0036] (1) Complex domain transformation: Apply the Hilbert transform to the support set and the query set respectively. The goal of the Hilbert transform is to find the adjoint function y(t) for the real function x(t) such that the complex function z(t) = x(t) + iy(t) can be analytic on the real axis t ∈ R and extended to the upper half of the complex plane. By convolving the signal, this transformation can generate a complex-valued signal from a real-valued signal I(x).

[0037] (2) Feature extraction: Extract complex domain features from the support set and the query set after the Hilbert transform through the ResNet12 backbone network. It is divided into 4 stages. The first 3 stages are preprocessing of the input, and the input is It successively passes through the convolutional layer, BN layer, LeakyReLU activation function, and max pooling layer to obtain The output features of the last stage are the output features of the backbone network, where B is the number of samples in each batch, C is the number of channels, and H and W represent the height and width of the features respectively.

[0038] The training of the graph network training network model is as follows:

[0039] (1) Complex domain node initialization and update module: Extract complex domain node features through the backbone network. A total of 5 nodes are composed of 4 support set samples and 1 query set sample for node initialization. Let v i be the node feature of V i , V i = {v i |i = 1, 2,..., N×K + T}. When the true label of the support set is the same as the predicted label of the query set, it is 1, otherwise it is 0. Given the nodes and edges of the (l - 1)-th layer, update the node features through the process of neighborhood aggregation. The features of the nodes in the l-th layer are updated by aggregating the features of other nodes together. The aggregation method is proportional to the edge features of the nodes, and then the aggregated features are transformed to update the features of the nodes in the l-th layer. The node features of the (l - 1)-th layer Update the node features of the l-th layer The edge features of the (l - 1)-th layer are used to quantify the contribution degree of the corresponding neighboring nodes.

[0040] (2) Construct a hybrid edge initialization and update module: Nodes effectively capture similarity information between each other through similarity metric edges, thereby generating preliminary edge features. The similarity metric edges and relative metric edges are concatenated to form the hybrid edges of the nodes, thus constituting the hybrid edges of the graph. Similarity metric edges can only reflect the similarity in node features, while relative metric edges can effectively distinguish nodes with similar features but different categories by encoding the relative positions or topological relationships between nodes. Denote the relative metric edge as e ij1 Represents the edge feature of the first dimension between node i and node j. Denote the similarity metric edge as e ij2 , and the dissimilarity metric edge as e ij3 . Initialize the relative metric edge And combine it with the similar and dissimilar structures between nodes to initialize the edges of the graph After the node features are updated, new node features will be obtained, thereby realizing the update of the hybrid edge features

[0041] Third step, image classification function: After l rounds of node and hybrid edge feature updates, use the final node features to generate the prediction results of the hybrid edge labels, that is where y ij ∈[0,1] is regarded as the probability that two nodes v i and v j come from the same class. Therefore, the classification result can be obtained by jointly predicting the label results of the query set with the true labels of the support set and the hybrid edges

[0042] Based on the above steps, as Figure 1 shown, a SAR image target classification system based on a complex domain feature map network, the specific operation steps are as follows:

[0043] 1. A SAR image target classification system based on a complex domain feature map network, characterized in that it includes the following steps:

[0044] S1: Divide the SAR dataset into a training set, a validation set, and a test set according to the division ratio, and perform data augmentation on the training set;

[0045] S2: Divide the SAR images into a support set and a query set;

[0046] S3: Pass the support set and the query set through the complex domain transformation feature extraction module respectively to obtain the complex domain features of the SAR;

[0047] S4: Use the graph network in the system to update to obtain the support set image most similar to the query set image;

[0048] S5: Print the final classification accuracy result in the display window.

[0049] In the optimization process of the entire network of the present invention, the network gradually learns the differences between intra-class nodes and inter-class nodes, as well as the different characteristics between similarity measurement edges and relative measurement edges. This process can be regarded as a process for an expert in the relevant field to learn corresponding knowledge. When the model finds the optimal network model through continuous optimization, the entire network is equivalent to an expert with rich relevant knowledge. At this time, the model has the ability to quickly identify SAR images.

[0050] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A SAR image target classification system based on complex domain feature graph network, characterized by: The following steps are involved: S1: Divide the SAR dataset into training set, validation set, and test set according to the division ratio, and perform data augmentation on the training set; S2: Divide the SAR images into support set and query set; S3: The support set and the query set are respectively subjected to the complex domain transformation feature extraction module to obtain the complex domain features of SAR; S4: Use the graph network in the system to train the network model to obtain the support set image that is most similar to the query set image; S5: Print the final classification accuracy result in the display window; The complex domain transformation feature extraction module described in step S3 includes the following steps: The first step is complex domain transformation: the support set and query set images are transformed by Hilbert transform respectively. The goal of Hilbert transform is to transform the real function Find the adjoint function , so that the complex function , can be found on the real axis The upper analysis is performed and extended to the upper half of the complex plane by convolving the signal; The second step is feature extraction: the complex domain image is extracted through the ResNet12 backbone network. It is divided into four stages, of which the first stage is the processing of the input. The input is , which has been obtained through the convolution layer, BN layer, LeakyReLU activation function, and maximum pooling layer , the output features of the last stage That is, the output feature of the backbone network, where B is the size of each batch; The specific steps of training the graph network training network model described in step S4 are as follows: (1) Complex domain node initialization and update module: The complex domain node features are extracted through the backbone network. Five nodes are composed of four support set samples and one query set sample. The nodes are initialized. yes The node characteristics of , when the true label of the support set is consistent with the predicted label of the query set, it is 1, otherwise it is 0, ; Given Layer Node and edge ,Update node features through the process of domain aggregation, The features of the layer nodes are updated by aggregating the features of other nodes in a way that is proportional to the edge features of the nodes, and then transforming the aggregated features to update the first Characteristics of layer nodes , No. Node characteristics of the layer Update Node characteristics of the layer , No. Layer edge features Used to quantify the contribution of the corresponding neighboring nodes ; (2) Constructing the hybrid edge initialization and update module: Nodes effectively capture the similarity information between each other through similarity measurement edges, thereby generating preliminary edge features. The similarity measurement edges and relative measurement edges are spliced ​​to form the hybrid edges of the nodes, thereby forming the hybrid edges of the graph. The relative measurement edges are recorded as , indicating a node and nodes The first dimension edge feature between them, the similarity measure edge is recorded as , the dissimilarity measure edge is recorded as , initialize the relative metric edge , and combined with the similar and dissimilar structures between nodes, initialize the edges of the graph After the node feature is updated, a new node feature will be obtained, thereby realizing the update of the mixed edge feature .

2. The SAR image target classification system based on complex domain feature graph network according to claim 1, characterized in that: The update of mixed edge features uses similarity metric edges and relative metric edges to construct the mixed edges of the graph. After the mixed edge update is completed, the nodes will be classified according to the characteristics of the points.

3. The SAR image target classification system based on complex domain feature graph network according to claim 1, characterized in that: In step S1, the SAR dataset includes four categories: ships, aircraft, oil tanks, and bridges. Data enhancement is performed on the training set to balance the number of samples of different categories in the dataset.

4. The SAR image target classification system based on complex domain feature graph network according to claim 1, characterized in that: The data augmentation method in step S1 includes rotating the image, horizontally flipping the image, vertically flipping the image, and randomly cropping.

5. The SAR image target classification system based on complex domain feature graph network according to claim 4, characterized in that: The angle of the rotated image is randomly chosen between -45 and 45 degrees.

6. The SAR image target classification system based on complex domain feature graph network according to claim 1, characterized in that: In step S5, the classification accuracy result is obtained by predicting the label result of the query set with the true label of the support set and the mixed edge. .

Images