A SAR target recognition method, system, device and medium based on complex domain statistical characteristic disentanglement and dynamic recombination network

Abstract

The application relates to the technical field of image processing, and provides a SAR target recognition method, system, device and medium based on complex domain statistical characteristics unwinding and a dynamic recombination network, the method comprising the following steps: constructing a complex convolutional neural network as an initial network model based on ResNet34, and designing a real-imaginary cross-channel feature unwinding and dynamic recombination module integrated into a network architecture; taking an enhanced image as input, and jointly training the initial network model by using an amplitude consistency constraint loss, a phase separability constraint loss and a cross-entropy loss to obtain an optimized network model; and using the optimized network model to recognize a test image set to obtain a recognition result; in the recognition process, the real-imaginary cross-channel feature unwinding and dynamic recombination module unwinds coupled complex features in a to-be-tested image into real part exclusive features, imaginary part exclusive features and shared features based on a statistical threshold, and combines SAR scattering physical priors to adaptively assign feature weights to different regions, and weighted fusion is performed to obtain recombination features. Compared with the prior art, the application has the advantages of high recognition accuracy, and can be widely used in the technical field of image processing.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a SAR target recognition method, system, device, and medium based on unwrapping and dynamic reassembly of networks using complex domain statistical properties. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an all-weather, all-day active microwave imaging system, plays an irreplaceable role in military reconnaissance, topographic mapping, and disaster monitoring. Research on target recognition technology in SAR images has become a hot topic in remote sensing. The raw data acquired by SAR imaging is in complex form, with each pixel consisting of a real part and an imaginary part, corresponding to the in-phase and quadrature components of the scattered echo, respectively. The real part mainly reflects the distribution of target scattering intensity and is closely related to the geometry of the target's strong scattering center; the imaginary part carries the phase changes caused by the propagation path and the target's fine structure, and is uniquely valuable for characterizing target contours, distinguishing local structures, and suppressing speckle noise.

[0003] Existing SAR target recognition methods mostly focus on amplitude information, neglecting the statistical regularities and physical correlations hidden in complex data. Some studies directly separate the real and imaginary parts of complex data into two channels and input them into traditional real-valued convolutional neural networks, allowing the network to autonomously learn feature correlations; another type of research designs complex-valued convolutional neural networks, extending convolution, activation, and other operations from the real domain to the complex domain, achieving end-to-end processing of complex information. For example, CN113420659B discloses a SAR image target recognition method based on a complex convolutional network, including: inputting images: inputting multiple complex images and their corresponding category labels into a training set, and inputting multiple complex images different from those in the training set into a test set; obtaining training samples and test samples: separately saving the real and imaginary parts of the complex images in the training set to obtain real part training samples and imaginary part training samples; separately saving the real and imaginary parts of the complex images in the test set to obtain real part test samples and imaginary part test samples; constructing a complex convolutional network: employing a two-layer complex convolutional layer, a complex batch normalization layer, and a fully connected layer. The two-layer complex convolutional layer contains two complex convolutional layers, each containing complex convolution, complex batch normalization, and a CReLU activation function. The batch normalization layer includes complex batch normalization and the CReLU activation function; Training the complex convolutional network: Initialize the complex convolutional network, setting all complex parameters appearing in the complex convolutional network to W, inputting real and imaginary training samples into two complex convolutional layers, passing through complex batch normalization and the CReLU activation function, using a 300-dimensional fully connected layer, outputting initial class labels, comparing the class labels of complex images in the training set with the initial class labels, using batch stochastic gradient descent to iteratively update the parameters of the complex convolutional network 3000 times, obtaining the updated complex convolutional network; Testing the complex convolutional network: Input real and imaginary test samples into the updated complex convolutional network, using a 300-dimensional fully connected layer, obtaining the class labels of complex images in the test set.

[0004] However, these methods lack explicit guidance and decoupling design for the differentiated physical meanings carried by the real and imaginary parts. The network still learns redundant coupling features, which easily leads to overfitting when the training samples are limited, and does not fully utilize the statistical properties of the complex domain as constraints. Therefore, this invention discloses a SAR target recognition method, system, device, and medium based on the unwrapping and dynamic reassembly of networks using the statistical properties of the complex domain. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a SAR target recognition method, system, device, and medium based on complex domain statistical characteristics unwrapping and dynamic recombination networks, in order to solve the problems of redundant coupling of real and imaginary part features and insufficient utilization of complex domain statistical characteristics in synthetic aperture radar image target recognition.

[0006] This invention is achieved through the following technical solution: A SAR target recognition method based on complex domain statistical properties unwrapping and dynamic reassembly of networks includes the following steps: A complex convolutional neural network based on ResNet34 was constructed as the initial network model, and a module for unwrapping and dynamically recombining cross-channel features of real and imaginary parts was designed and integrated into the network architecture. Using the enhanced image as input, the initial network model is jointly trained with amplitude consistency constraint loss, phase separability constraint loss, and cross-entropy loss to obtain the optimized network model; The optimized network model is used to identify the test image set and obtain the identification results. During the identification process, the real and imaginary part cross-channel feature unwrapping and dynamic recombination module unwrappes the coupled complex features in the image under test into real part exclusive features, imaginary part exclusive features and shared features based on statistical thresholds. It also adaptively assigns feature weights to different regions based on SAR scattering physics priors and obtains recombined features through weighted fusion.

[0007] Preferably, the initial network model includes a pre-reinforced ResNet34 backbone network, a module for unwrapping and dynamically reassembling cross-channel features of real and imaginary parts, and a post-reinforced ResNet34 backbone network; The front-end ResNet34 backbone network includes a series of convolutional layers, pooling layers, batch normalization layers, and ReLU activation functions; The real and imaginary part cross-channel feature unwrapping and dynamic recombination module includes a real part-specific feature extraction unit, an imaginary part-specific feature extraction unit, a shared feature extraction unit, and a physical prior dynamic feature recombination unit; The ResNet34 backbone network consists of a series of convolutional layers, batch normalization layers, pooling layers, and ReLU activation functions.

[0008] Preferred image acquisition process: Based on the statistical property that the real and imaginary parts of SAR complex images follow a joint Gaussian distribution, a joint Gaussian enhancement-structure preservation strategy for real and imaginary parts is designed. Adaptive perturbations are generated through local structure correlation analysis, and the enhanced samples are screened for structural integrity to obtain enhanced images.

[0009] Preferably, in the joint Gaussian enhancement-structure preservation strategy of real and imaginary parts, the local structure correlation of the local structure adaptive perturbation strategy is... The calculation formula is:

[0010] in,( , () represents the pixel coordinates of the window center; and These are the original complex images in The real and imaginary parts of the value, The global variance of the real and imaginary parts of this sample; x-coordinate relative to the center pixel of the window The lateral offset, The range of values ​​[ , ]; The vertical coordinate relative to the center pixel of the window The vertical offset, The range of values ​​[ , ]; It is the horizontal offset of the center pixel. Longitudinal offset The real grayscale value of the subsequent pixel; It is the horizontal offset of the center pixel. Longitudinal offset The imaginary grayscale value of the subsequent pixel; This represents the total number of pixels within the local window. This is the window half-width parameter, and the window size is... ; The closer the value is to 1, the smaller the difference in gray levels between the real and imaginary parts of each pixel in the window and the center pixel, the higher the consistency of pixel distribution, indicating that the structure in the window is more regular and is a high-structure region; the closer the value is to 0, the larger the difference in gray levels between the real and imaginary parts of each pixel in the window and the center pixel and the more disordered the distribution, the weaker the correlation between pixels, indicating that the structure is more random and is a low-structure region. The local structure adaptive perturbation strategy determines the perturbation coefficient based on the local structure correlation value. The perturbation term is further adjusted and superimposed with the original sample to obtain the initial enhanced sample. The enhancement process is as follows:

[0011]

[0012] in, The structural correlation threshold, For weak disturbance coefficients, This is the strong disturbance coefficient. , These are the real Gaussian perturbation term and the imaginary Gaussian perturbation term, respectively, generated based on the global statistical parameters of the real and imaginary parts. , These are the real and imaginary parts of the sample, respectively. The initial augmented samples are screened for structural integrity using the following formula:

[0013] in, To enhance the local structural correlation index of the initial sample, The threshold for structural change; compared with the original sample ( If the comparison does not exceed the threshold, If the structure is deemed acceptable, then the perturbation term is regenerated and the verification is repeated until an acceptable enhanced sample is obtained. ).

[0014] Preferably, during the initial training of the network model, the Adam optimization algorithm is used, and the data batch, training epochs, learning rate, and weight parameter values ​​are set. During training, forward propagation is used to calculate the loss function, and backpropagation is used to update the network weights until the loss function converges, thus obtaining the parameters of the optimized network model.

[0015] Preferably, the process of constructing the amplitude consistency constraint loss is as follows:

[0016] in, , These are the global mean and variance of the amplitude characteristics within the target region of similar samples. , They represent the first The mean and variance of the amplitude of each sample target region. The total number of samples; Let the mean square error function be used. The process of constructing the phase separability constraint loss is as follows:

[0017] in, The phase distance between any two out-of-class sample vectors. For the total number of out-of-class sample pairs, It is the phase separation threshold; The expression for the total loss function is:

[0018] in, Indicates the total loss. For cross-entropy loss, For amplitude consistency constraint loss, For phase separability constraint loss, parameters , These are used to adjust the weights of amplitude consistency and phase separability constraint losses in the total loss, respectively.

[0019] Preferably, the cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts unwrappes the coupled complex features in the image under test into real-part specific features, imaginary-part specific features, and shared features based on statistical thresholds. It then adaptively assigns feature weights to different regions based on SAR scattering physics priors, and weightedly fuses these weighted features to obtain the recombined features, including: Step S100, Feature unwrapping: Extract real part specific features from the image to be tested. Virtual Part Specific Features and shared features The details are as follows: The real-dominant region was screened by calculating the local variance ratio of the real and imaginary parts, and real-part specific features were extracted. The specific process is as follows:

[0020]

[0021] in, , The input consists of the original real part features and the original imaginary part features. Indicates in a local window Internal variance calculation For the Sigmoid function, This is the real part-specific mask for the generated data; Represents matrix dot product. For statistically constrained convolution; The edge response ratio is used to screen the imaginary part-dominant region and extract imaginary part-specific features. The specific process is as follows:

[0022]

[0023] in, This indicates the calculation of gradient magnitude. This is the mask for the generated imaginary part. Shared features The extraction process is as follows:

[0024] in, It is achieved by calculating the cross-channel Pearson correlation coefficient matrix of the real and imaginary parts of the characteristic channels. ( (Based on the number of feature channels) a set of highly correlated feature channels whose absolute correlation coefficients are greater than a threshold are selected. , They are Features of the real and imaginary parts within, It is a two-dimensional Gaussian smoothing operator; Step S200, Feature Weighted Recombination: Assign differential weights based on the scattering type of the region where the pixel is located. Give the real part exclusive features Virtual Part Specific Features and shared features Acquire recombination characteristics The reorganization process is as follows:

[0025] in, , , These are the recombination weights of the extracted real-part specific features, imaginary-part specific features, and shared features. This is the final reorganization characteristic; The feature weight combination is based on the scattering contribution and real / imaginary part characteristics of different regions of the SAR target, classifying them into different types of image regions. Predefined.

[0026] A SAR image target recognition system based on complex domain statistical property unwrapping and dynamic reassembly network, used to implement the SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network, includes: The network building module is used to build a complex convolutional neural network based on ResNet34 as the initial network model. The module for unwrapping and dynamically recombining cross-channel features of real and imaginary parts is designed and integrated into the network architecture. The network training module is used to train the initial network model by taking the enhanced image as input and using amplitude consistency constraint loss, phase separability constraint loss and cross-entropy loss together to obtain the optimized network model. The image recognition module is used to identify the test image set using an optimized network model and obtain the recognition results.

[0027] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the SAR target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network.

[0028] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the SAR target recognition method based on complex domain statistical properties unwrapping and dynamic reassembly networks.

[0029] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a SAR target recognition method based on complex domain statistical property unwrapping and dynamic recombining networks, which includes a cross-channel feature unwrapping and dynamic recombining module for real and imaginary parts. The feature unwrapping part of this module separates real and imaginary part-specific features from shared features based on the differentiated physical characteristics of the real and imaginary parts, thereby reducing feature redundancy. The dynamic recombining part assigns differentiated weights to the three types of features according to the scattering type of the pixel's region. Through weighted recombining of features, it achieves adaptive fusion of real and imaginary part-specific features and shared features, thereby simulating the regional differences in real electromagnetic scattering and improving the feature's ability to represent the target's scattering characteristics. The recombined features are directly input into the subsequent classification module, providing high-quality support for accurate classification.

[0030] This invention constructs a complex-domain statistical contrastive constraint mechanism during the initial network model training, fusing amplitude consistency constraint loss and phase separability constraint loss. This mechanism introduces statistical characteristics into the contrastive loss, guiding the network to learn complex feature representations that maintain intra-class stability while amplifying inter-class discriminativeness, thus statistically strengthening the model's ability to model the essential scattering characteristics of SAR targets. Simultaneously, the total loss function, combined with cross-entropy loss, guides the network to learn discriminative features that conform to the statistical laws of SAR scattering, further enhancing the robustness and discriminative power of the model's features from a statistical perspective.

[0031] Based on the statistical characteristic that the real and imaginary parts of SAR complex images follow a joint Gaussian distribution, this invention designs a joint Gaussian enhancement-structure preservation strategy for the real and imaginary parts. Adaptive perturbations are generated through local structural correlation analysis, and the enhanced samples are screened for structural integrity. This increases data diversity while maintaining the statistical rationality and structural integrity of the enhanced samples.

[0032] This invention delves into the relationship between the statistical properties of the real and imaginary parts of SAR and the scattering physics mechanism, providing quantifiable screening criteria and statistical basis for feature unwrapping and recombination. This addresses the problems of redundant coupling of real and imaginary features, insufficient utilization of complex domain statistical properties leading to weak model generalization ability, and limited recognition performance in complex scenes during SAR image target recognition. Compared with existing technologies, the method proposed in this invention differs from traditional complex domain processing methods. It mines and utilizes the complementary scattering information carried by the real and imaginary parts of SAR complex data to achieve efficient feature decoupling and physical constraints. It boasts advantages such as high recognition accuracy and strong robustness, and can be widely applied in the field of image processing technology. Attached Figure Description

[0033] Figure 1 This is a flowchart of the invention; Figure 2 This is a schematic diagram of the complex convolutional neural network structure constructed in this invention; Figure 3 yes Figure 2 A schematic diagram of the structure of the module for untangling and dynamically recombining features across channels of real and imaginary parts. Detailed Implementation

[0034] The present invention will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and are not intended to limit the scope of the invention.

[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] This invention provides a SAR target recognition method based on complex domain statistical property unwrapping and dynamic reassembly networks, such as... Figure 1 As shown, it includes the following steps: S1 uses a complex convolutional neural network built on ResNet34 as the initial network model, and integrates a module for unwrapping and dynamically recombining real and imaginary cross-channel features into the network architecture. Specifically, based on statistical thresholds, the coupled complex features are unwrapped into real-part specific features, imaginary-part specific features, and shared features. In addition, feature weights are adaptively assigned to different regions based on SAR scattering physics priors, thereby achieving differentiated dynamic recombination of features.

[0038] like Figure 2As shown, the complex convolutional neural network consists of a pre-ResNet34 backbone network, a cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts, and a post-ResNet34 backbone network. The pre-ResNet34 backbone network includes a series of convolutional layers, pooling layers, batch normalization layers, and ReLU activation functions, extracting basic shallow and mid-level features of the real and imaginary parts layer by layer. The cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts includes a real-part-specific feature extraction unit, an imaginary-part-specific feature extraction unit, a shared feature extraction unit, and a physically prior dynamic feature recombination unit. The post-ResNet34 backbone network includes a series of convolutional layers, batch normalization layers, pooling layers, and ReLU activation functions, used to extract high-level semantic features.

[0039] S2 takes the enhanced image as input and trains the initial network model jointly with amplitude consistency constraint loss, phase separability constraint loss and cross-entropy loss to obtain the optimized network model.

[0040] (1) Data Augmentation: Based on the statistical characteristics that the real and imaginary parts of SAR complex images follow a joint Gaussian distribution, a joint Gaussian augmentation-structure preservation strategy for the real and imaginary parts is designed. Adaptive perturbations are generated through local structure correlation analysis, and the augmented samples are screened for structural integrity to obtain augmented images. The complex images acquired by SAR imaging are from the ten classes of Moving and Stationary Target Acquisition and Recognition (MSTAR) datasets, including training image sets and test image sets.

[0041] Among them, in the joint Gaussian enhancement-structure preservation strategy of real and imaginary parts, the local structure correlation of the local structure adaptive perturbation strategy. The calculation formula is:

[0042] ( , () represents the pixel coordinates of the window center; and These are the original complex images in The real and imaginary parts of the value, The global variance of the real and imaginary parts of this sample; x-coordinate relative to the center pixel of the window The lateral offset, The range of values ​​[ , ]; The vertical coordinate relative to the center pixel of the window The vertical offset, The range of values ​​[ , ]; It is the horizontal offset of the center pixel. Longitudinal offset The real grayscale value of the subsequent pixel; It is the horizontal offset of the center pixel. Longitudinal offset The imaginary grayscale value of the subsequent pixel; This represents the total number of pixels within the local window. This is the window half-width parameter, and the window size is... ; The closer the value is to 1, the smaller the difference in grayscale between the real and imaginary parts of each pixel within the window and the center pixel, indicating a higher consistency in pixel distribution and a more regular structure within the window, thus representing a high-structure region. Conversely, the closer the value is to 0, the larger the difference in grayscale between the real and imaginary parts of each pixel within the window and the center pixel, indicating a chaotic distribution and weak correlation between pixels, thus representing a more random structure and a low-structure region. For example, suppose the local window is 3... 3, It is 1; , The values ​​are all -1, 0, 1. If the center pixel coordinates are (2, 2), then the coordinates of the surrounding pixels are (2-1, 2-1), (2-1, 2), (2-1, 2+1), (2, 2-1), (2, 2+1), (2+1, 2-1), (2+1, 2), (2+1, 2+1).

[0043] The local structure adaptive perturbation strategy determines the perturbation coefficient based on the local structure correlation value. The perturbation term is further adjusted and superimposed with the original sample to obtain the initial enhanced sample. The enhancement process is as follows:

[0044]

[0045] in, The structural correlation threshold, For weak disturbance coefficients, This is the strong disturbance coefficient. , These are the real Gaussian perturbation term and the imaginary Gaussian perturbation term, respectively, generated based on the global statistical parameters of the real and imaginary parts. , These are respectively enhancing the real and imaginary parts of the samples; adaptive perturbation not only improves the sample diversity in the background region but also avoids damage to the key structures of the target.

[0046] In the joint Gaussian augmentation-structure preservation strategy involving real and imaginary parts, the initial augmented samples are screened for structural integrity. The screening formula is as follows:

[0047] in, To enhance the local structural correlation index of the initial sample, This is the threshold for structural change. Compared with the original sample ( If the comparison does not exceed the threshold, the structure is considered acceptable. Otherwise, the perturbation term is regenerated and the verification is repeated until an acceptable enhanced sample is obtained. ).

[0048] (2) Loss function The process of constructing the magnitude consistency constraint loss by combining the loss function with the complex domain statistical comparison constraint mechanism and cross-entropy loss is as follows:

[0049] in, , These are the global mean and variance of the amplitude characteristics within the target region of similar samples. , They represent the first The mean and variance of the amplitude of each sample target region. The total number of samples; The mean squared error function is used to force the amplitude distribution of individual samples to closely approximate the overall distribution of their respective categories. The process of constructing the phase separability constraint loss is as follows:

[0050] in, The phase distance between any two out-of-class sample vectors. For the total number of out-of-class sample pairs, This is the phase separation threshold, used to control the degree of separation between outlier samples in the phase feature space. This loss penalizes outlier samples whose phase distance is less than a certain threshold. This situation forces the network to mine subtle differences in phase information that are more discriminative in distinguishing targets.

[0051] The process of the total loss function is as follows:

[0052] in, Indicates the total loss. For cross-entropy loss, For amplitude consistency constraint loss, For phase separability constraint loss, parameters , These are used to adjust the weights of the amplitude consistency and phase separability constraint losses in the total loss, respectively. The total loss function, on the one hand, through... and This ensures the model can learn complex features that conform to the true scattering laws; on the other hand, it utilizes... This makes the model's predictions closer to the true labels.

[0053] This invention guides the network to learn discriminative features that conform to the statistical laws of SAR scattering by combining the total loss function of complex domain statistical comparison constraint loss and cross-entropy loss, thereby enhancing the robustness and discriminative ability of the model features from a statistical perspective. During training, the enhanced image dataset is input into the network, and the complex features of the training set images are input into the network and passed through all modules sequentially. The Adam optimization algorithm is used during training, with settings for data batch size, training epochs, learning rate, and weight parameters. Specifically, in this embodiment, the data batch size can be set to 32, the training epochs to 100, the learning rate to 0.0001, and the loss weights... , It can be set to 0.3 or 0.2.

[0054] During training, the loss function is calculated through forward propagation, and the weights of the network are updated through backpropagation until the loss function converges, thus obtaining the parameters of the trained network model.

[0055] S3, using an optimized network model to identify the test image set, obtain the recognition results; such as Figure 3 As shown, during the recognition process, the real and imaginary part cross-channel feature unwrapping and dynamic recombination module unwrappes the coupled complex features in the image under test into real-part specific features, imaginary-part specific features, and shared features based on statistical thresholds. It also adaptively assigns feature weights to different regions based on SAR scattering physics priors, and obtains the recombined features through weighted fusion. Specifically: Step S100, Feature unwrapping: Extract real part specific features from the image to be tested. Virtual Part Specific Features and shared features The details are as follows: The real-dominant region was screened by calculating the local variance ratio of the real and imaginary parts, and real-part specific features were extracted. The specific process is as follows:

[0056]

[0057] in, , The input consists of the original real part features and the original imaginary part features. Indicates in a local window Internal variance calculation For the Sigmoid function, The generated real part-specific mask is used to quantify the importance of real part features at different locations. The closer the value is to 1, the more significant the real part-specific information in that region is; the closer the value is to 0, the more the region is dominated by shared information or imaginary part information. Represents matrix dot product. This is a statistically constrained convolution. It incorporates a priori constraint kernel learning based on the variance statistics of SAR real-part features, retaining only features that conform to the statistical laws of real-part scattering, thus achieving the extraction of real-part specific features. .

[0058] The edge response ratio (characterized by the ratio of the gradient magnitudes of the imaginary and real parts in this invention) is used to screen regions dominated by the imaginary part, and imaginary part-specific features are extracted. The specific process is as follows:

[0059]

[0060] in, This indicates the calculation of gradient magnitude. The generated imaginary part-specific mask is used to quantify the importance of imaginary part features at different locations. A larger value indicates a stronger imaginary part gradient response and more significant edge information, while a smaller value indicates a weaker contribution of the imaginary part gradient and a bias towards the real part or the dominance of shared information. After fusing the mask with the original imaginary part features, statistically constrained convolution is also performed to obtain the imaginary part-specific features that carry the detailed edge information of the target. .

[0061] Shared features The extraction focuses on integrating redundant information in the real and imaginary parts. The extraction process is as follows:

[0062] in, It is achieved by calculating the cross-channel Pearson correlation coefficient matrix of the real and imaginary parts of the characteristic channels. ( (Based on the number of feature channels) a set of highly correlated feature channels whose absolute correlation coefficients are greater than a threshold are selected. , They are Features of the real and imaginary parts within, This is a two-dimensional Gaussian smoothing operator used to suppress speckle noise on shared features. Interference.

[0063] Step S200, Feature Weighted Recombination: Assign differential weights based on the scattering type of the region where the pixel is located. Give the real part exclusive features Virtual Part Specific Features and shared features Acquire recombination characteristics The reorganization process is as follows:

[0064] in, , , These are the recombination weights of the extracted real-part specific features, imaginary-part specific features, and shared features. This is the final reconstructed feature; based on the scattering contribution and real / imaginary part characteristics of different regions of the SAR target, it is used to reconstruct different types of image regions. (e.g., strong scattering regions, target edges, background) Predefined feature weight combinations Differential weights are assigned based on the scattering type of the pixel's region to achieve adaptive fusion of the three types of features.

[0065] This invention addresses the problems of redundant coupling of real and imaginary features and insufficient utilization of complex domain statistical properties by constructing a cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts. This module includes statistically driven feature unwrapping and physically prior dynamic recombination, offering advantages such as high recognition accuracy and widespread application in the field of image processing technology. Specifically, the feature unwrapping module decomposes the hybrid coupled real and imaginary joint features into real-part-specific features, imaginary-part-specific features, and shared features, thereby removing redundant information and eliminating feature overlap and waste. After unwrapping, the dynamic recombination module assigns adaptive differentiated weights to the three types of unwrapped features based on the differences in electromagnetic scattering statistical laws in different pixel regions of the image, according to the pixel scattering type. The three types of features are then weighted and fused, with dynamic weights matching the differences in complex domain statistical distribution in different regions, simulating the characteristics of electromagnetic scattering regions in real scenes and fully utilizing the complex statistical laws of real and imaginary parts. The recombined features can accurately characterize the target scattering characteristics, thereby improving the SAR target scattering feature characterization capability and enhancing recognition performance.

[0066] Example 1 The SAR target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network in this embodiment consists of the following steps: S1 constructs a complex convolutional neural network for augmenting data, consisting of a pre-ResNet34 backbone network, a module for unwrapping and dynamically reassembling cross-channel features of the real and imaginary parts, and a post-ResNet34 backbone network. The pre-ResNet34 backbone network includes a series of convolutional layers, pooling layers, batch normalization layers, and ReLU activation functions, extracting basic shallow and mid-level features of the real and imaginary parts layer by layer. The module for unwrapping and dynamically reassembling cross-channel features of the real and imaginary parts includes a real-part-specific feature extraction unit, an imaginary-part-specific feature extraction unit, a shared feature extraction unit, and a physically prior dynamic feature reassembly unit. The post-ResNet34 backbone network includes a series of convolutional layers, batch normalization layers, pooling layers, and ReLU activation functions, used to extract high-level semantic features.

[0067] In this embodiment, the front-end ResNet34 backbone network, the real and virtual part cross-channel feature unwrapping and dynamic recombination module, and the rear-end ResNet34 backbone network are connected in series, and the target recognition category result is output at the end.

[0068] S2, Training the Network Data augmentation: The MSTAR ten-class dataset was used as the training and testing data. The image size was 128×128, with the original training set having an observation pitch angle of 17° and the original testing set having an observation pitch angle of 15°.

[0069] Adaptive perturbations are generated through local structural correlation analysis, and the enhanced samples are then screened for structural integrity. In this embodiment, we take... We obtain the enhanced image dataset after processing with a joint Gaussian enhancement-structure preservation strategy for real and imaginary parts.

[0070] During network training, augmented image datasets are input into the network. Complex features of the training set images are input into the network and passed through all modules sequentially. The Adam optimization algorithm is used during training, with a batch size of 32, 100 training epochs, a learning rate of 0.0001, and a loss weight of [missing value]. , The values ​​are 0.3 and 0.2. During training, the loss function is calculated through forward propagation and the network weights are updated through backpropagation until the loss function converges.

[0071] S3, Target Recognition: Input the test dataset into the trained optimized network model to obtain the recognition results.

[0072] To verify the effectiveness of this invention in assisting target identification using SAR image phase information, the MSTAR ten-class dataset was used. The original training set was set with an observation elevation angle of 17°, and the original test set was set with an observation elevation angle of 15°. The number of samples for each type in the dataset is summarized in Table 1.

[0073] Table 1. Number of training and test samples in the MSTAR dataset

[0074] To verify the impact of each key component in the proposed model on SAR recognition performance, the effectiveness of the invention will be verified through ablation experiments, focusing on the joint Gaussian enhancement-structure preservation strategy of real and imaginary parts, the cross-channel feature unwrapping and dynamic recombination module of real and imaginary parts, and the complex domain statistical comparison constraint mechanism, as detailed below: Table 2 Ablation experiments for each module

[0075] Experimental results show that the model of this invention, with the three components removed—the joint Gaussian enhancement-structure preservation strategy for real and imaginary parts, the cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts, and the complex domain statistical comparison constraint mechanism—achieved an overall accuracy of 97.47% as the baseline model, demonstrating the effectiveness of the basic network architecture, although there is room for performance improvement. Introducing only the joint Gaussian enhancement-structure preservation strategy for real and imaginary parts improved the accuracy to 99.05%, a 1.58% improvement over the baseline model, indicating that the adaptive data augmentation method based on the statistical characteristics of the SAR complex domain effectively alleviated the problem of limited sample size. Adding only the cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts achieved an accuracy of 99.13%, a 1.66% improvement, proving that this module can accurately separate redundant coupled features of real and imaginary parts, and that the physically prior-guided dynamic recombination strategy enhances the discriminative power and physical interpretability of the features. When only the complex domain statistical contrast constraint mechanism is introduced, the accuracy rate is 97.91%, an improvement of 0.44%. Although the improvement is smaller than that of the first two components, it verifies the positive effect of the complex domain statistical contrast constraint mechanism on optimizing feature distribution. Through amplitude consistency and phase separability constraints, the features are made to better fit the physical laws of SAR scattering.

[0076] When the two components work together, performance is further improved: the accuracy of the combination of the real and imaginary parts joint Gaussian enhancement-structure preservation strategy and the real and imaginary parts cross-channel feature unwrapping and dynamic recombination module is 99.34%, an improvement of 0.21% compared to the highest value of a single component; the accuracy of the combination of the real and imaginary parts cross-channel feature unwrapping and dynamic recombination module and the complex domain statistical comparison constraint mechanism is 99.30%; and the accuracy of the combination of the real and imaginary parts joint Gaussian enhancement-structure preservation strategy and the complex domain statistical comparison constraint mechanism is 99.22%. This indicates that the high-quality samples provided by data augmentation lay the foundation for feature unwrapping and constraint optimization, and the pure features after feature unwrapping also make the optimization effect of statistical constraints more significant, with a benign synergistic effect among the modules. When all three components are introduced (i.e., the complete model of this invention), the accuracy reaches 99.63%, an improvement of 2.16% compared to the baseline model and 0.29% compared to the optimal two-component combination, fully verifying the rationality of the design of each component of the proposed framework and the effectiveness of synergistic optimization.

[0077] To verify the performance of the present invention. Comparative Experiment 1 uses "J. Ai, Y. Mao, Q. Luo, L. Jia, and M. Xing, SAR target classification using the multikernel-size feature fusion-based convolutional neural network," in IEEE Transactions on Geoscience and Remote Sensing, vol.60, pp.1-13, Sep. 2021," from existing related technologies; Comparative Experiment 2 uses "S. Chen, H. Wang, F. Xu, and YQ Jin, Target classification using the deep convolutional networks for SAR images," in IEEE Transactions on Geoscience and Remote Sensing, vol.54, no. 8, pp.4806–4817, Apr. 2016," from existing related technologies; and Comparative Experiment 2 uses "Y. Zhang, J. Xia, X. Gao, L. Xue, X. Zhang and X. Li, SM-CNN: Separability Measure-Based CNN for SAR Target Recognition," in IEEE Geoscience. and Remote Sensing Letters, vol. 20, pp. 1-5, 2023.” as comparative experiment 3; “Z. Zhang, H. Wang, F. Xu, and YQJin, "Complex-valued convolutional neural network and its application in polarimetric SAR image classification," in IEEE Transactions on Geoscience and Remote Sensing, vol. 55, no. 12, pp. 7177–7188, Dec. 2017.” as comparative experiment 4; “R. Wang, Z. Wang, K.” as comparative experiment 4.Xia, H. Zou and J. Li, "Target recognition in single-channel SAR images based on the complex-valued convolutional neural network with data augmentation," in IEEE Transactions on Aerospace and Electronic Systems, vol. 59, no.2, pp.796-804, Apr. 2023. This was used as Comparative Experiment 5. The existing related technology, "P. Ni, G. Xu, Z. Zhong, J. Chen and W. Hong, "SAR Target Recognition Using Complex Manifold Multiscale Feature Fusion Network," in IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, pp. 3532-3535, 2024.", was used as Comparative Experiment 6. Comparative experiments were conducted, and the recognition accuracy was used for evaluation. All networks had identical parameter settings.

[0078] Table 3 Comparison experiment with other algorithms

[0079] As shown in Table 3, in the real-valued models, Comparative Experiment 1 achieved a classification accuracy of 97.44%, Comparative Experiment 2 improved the accuracy to 99.13% through feature enhancement strategies, and Comparative Experiment 3 achieved 99.32% thanks to the feature extraction capabilities of deep networks, demonstrating the strong potential of real-valued deep models. However, due to limitations in real-domain processing, it cannot fully utilize the complete information of SAR complex data. In the complex-valued models, Comparative Experiment 4, as an early complex-valued network, achieved an accuracy of 98.10%, lower than advanced real-valued models. Comparative Experiments 5 and 6 improved the accuracy to 99.01% and 99.30% respectively by optimizing complex-valued operations and feature fusion strategies, demonstrating the advantages of complex-valued networks in mining complex information. However, due to the lack of targeted feature unwrapping and physical constraint design, there is still room for performance improvement. The accuracy of Example 1 is 99.63%, which is 0.31% higher than that of the best real-valued model in Experiment 3 and 0.33% higher than that of the best complex-valued model in Experiment 6. This fully demonstrates that the present invention, through the feature untangling, recombination and constraint mechanism driven by statistical physics, can more fully explore the physical statistical information in the complex domain and break through the performance bottleneck of traditional real-valued and complex-valued models.

[0080] In another embodiment of the present invention, a SAR image target recognition system based on complex domain statistical property unwrapping and dynamic reassembly network is provided, for implementing the SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network, comprising: The network building module is used to build a complex convolutional neural network based on ResNet34 as the initial network model. The module for unwrapping and dynamically recombining cross-channel features of real and imaginary parts is designed and integrated into the network architecture. The network training module is used to train the initial network model by taking the enhanced image as input and using amplitude consistency constraint loss, phase separability constraint loss and cross-entropy loss together to obtain the optimized network model. The image recognition module is used to identify the test image set using an optimized network model and obtain the recognition results.

[0081] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, 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, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can implement the operation of a SAR target recognition method based on complex domain statistical characteristics unwrapping and dynamic reassembly networks.

[0082] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device within a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium of the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space containing the terminal's operating system. Furthermore, this storage space also contains one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the SAR target recognition method based on complex domain statistical property unwrapping and dynamic reassembly networks in the above embodiments.

[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the functions specified in one or more boxes. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A SAR target recognition method based on complex domain statistical properties unwrapping and dynamic reassembly networks, characterized in that, Includes the following steps: A complex convolutional neural network based on ResNet34 was constructed as the initial network model, and a module for unwrapping and dynamically recombining cross-channel features of real and imaginary parts was designed and integrated into the network architecture. Using the enhanced image as input, the initial network model is jointly trained with amplitude consistency constraint loss, phase separability constraint loss and cross-entropy loss to obtain the optimized network model; The optimized network model is used to identify the test image set and obtain the identification results. During the identification process, the real and imaginary part cross-channel feature unwrapping and dynamic recombination module unwrappes the coupled complex features in the image under test into real part exclusive features, imaginary part exclusive features and shared features based on statistical thresholds. It also adaptively assigns feature weights to different regions based on SAR scattering physics priors and obtains recombined features through weighted fusion.

2. The SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network according to claim 1, characterized in that, The initial network model includes a pre-reinforced ResNet34 backbone network, a module for unwrapping and dynamically reassembling cross-channel features of real and imaginary parts, and a post-reinforced ResNet34 backbone network; The front-end ResNet34 backbone network includes a series of convolutional layers, pooling layers, batch normalization layers, and ReLU activation functions; The module for unwrapping and dynamically recombining real and imaginary parts across channels includes a real part-specific feature extraction unit, an imaginary part-specific feature extraction unit, a shared feature extraction unit, and a physical prior dynamic feature recombining unit; The ResNet34 backbone network consists of a series of convolutional layers, batch normalization layers, pooling layers, and ReLU activation functions.

3. The SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network as described in claim 1, characterized in that, The process of acquiring enhanced images: Based on the statistical property that the real and imaginary parts of SAR complex images follow a joint Gaussian distribution, a joint Gaussian enhancement-structure preservation strategy for real and imaginary parts is designed. Adaptive perturbations are generated through local structure correlation analysis, and the enhanced samples are screened for structural integrity to obtain enhanced images.

4. The SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network according to claim 3, characterized in that, Local structure correlation in the local structure adaptive perturbation strategy of the joint real and imaginary part Gaussian enhancement-structure preservation strategy. The calculation formula is: in,( , () represents the pixel coordinates of the window center; and These are the original complex images in The real and imaginary parts of the value, The global variance of the real and imaginary parts of this sample; x-coordinate relative to the center pixel of the window The lateral offset, The range of values ​​[ , ]; The vertical coordinate relative to the center pixel of the window The vertical offset, The range of values ​​[ , ]; It is the horizontal offset of the center pixel. Longitudinal offset The real grayscale value of the subsequent pixel; It is the horizontal offset of the center pixel. Longitudinal offset The imaginary grayscale value of the subsequent pixel; This represents the total number of pixels within the local window. This is the window half-width parameter, and the window size is... ; The closer the value is to 1, the smaller the difference in gray levels between the real and imaginary parts of each pixel in the window and the center pixel, the higher the consistency of pixel distribution, indicating that the structure in the window is more regular and is a high-structure region; the closer the value is to 0, the larger the difference in gray levels between the real and imaginary parts of each pixel in the window and the center pixel and the more disordered the distribution, the weaker the correlation between pixels, indicating that the structure is more random and is a low-structure region. The local structure adaptive perturbation strategy determines the perturbation coefficient based on the local structure correlation value. The perturbation term is further adjusted and superimposed with the original sample to obtain the initial enhanced sample. The enhancement process is as follows: in, The structural correlation threshold, For weak disturbance coefficients, This is the strong disturbance coefficient. , These are the real Gaussian perturbation term and the imaginary Gaussian perturbation term, respectively, generated based on the global statistical parameters of the real and imaginary parts. , These are the real and imaginary parts of the sample, respectively. The initial augmented samples are screened for structural integrity using the following formula: in, To enhance the local structural correlation index of the initial sample, The threshold for structural change; compared with the original sample ( If the comparison does not exceed the threshold, If the structure is deemed acceptable, then the perturbation term is regenerated and the verification is repeated until an acceptable enhanced sample is obtained. ).

5. The SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network according to claim 1, characterized in that, During the initial training of the network model, the Adam optimization algorithm was used, and the data batch, training epochs, learning rate, and weight parameter values ​​were set. During training, the loss function is calculated using forward propagation, and the network weights are updated using backpropagation until the loss function converges, thus obtaining the parameters of the optimized network model.

6. The SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network according to claim 5, characterized in that, The process of constructing the amplitude consistency constraint loss is as follows: in, , These are the global mean and variance of the amplitude characteristics within the target region of similar samples. , They represent the first The mean and variance of the amplitude of each sample target region. The total number of samples; Let the mean square error function be used. The process of constructing the phase separability constraint loss is as follows: in, The phase distance between any two out-of-class sample vectors. For the total number of out-of-class sample pairs, It is the phase separation threshold; The expression for the total loss function is: in, Indicates the total loss. For cross-entropy loss, For amplitude consistency constraint loss, For phase separability constraint loss, parameters , These are used to adjust the weights of amplitude consistency and phase separability constraint losses in the total loss, respectively.

7. The SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network according to claim 1, characterized in that, The cross-channel feature unwrapping and dynamic recombination module for real and imaginary parts unwrappes coupled complex features in the image under test into real-part specific features, imaginary-part specific features, and shared features based on statistical thresholds. It then adaptively assigns feature weights to different regions based on SAR scattering physics priors, and weightedly fuses these weighted features to obtain the recombined features, including: Step S100, Feature unwrapping: Extract real part specific features from the image to be tested. Virtual Part Specific Features and shared features The details are as follows: The real-dominant region was screened by calculating the local variance ratio of the real and imaginary parts, and real-part specific features were extracted. The specific process is as follows: in, , The input consists of the original real part features and the original imaginary part features. Indicates in a local window Internal variance calculation For the Sigmoid function, For the generated real part exclusive mask; Represents matrix dot product. For statistically constrained convolution; The edge response ratio is used to screen the imaginary part-dominant region and extract imaginary part-specific features. The specific process is as follows: in, This indicates the calculation of gradient magnitude. This is the mask for the generated imaginary part. Shared features The extraction process is as follows: in, It is achieved by calculating the cross-channel Pearson correlation coefficient matrix of the real and imaginary parts of the characteristic channels. ( (Based on the number of feature channels) a set of highly correlated feature channels whose absolute correlation coefficients are greater than a threshold are selected. , They are Features of the real and imaginary parts within, It is a two-dimensional Gaussian smoothing operator; Step S200, Feature Weighted Recombination: Assign differential weights based on the scattering type of the region where the pixel is located. Give the real part exclusive features Virtual Part Specific Features and shared features Acquire recombination characteristics The reorganization process is as follows: in, , , These are the recombination weights of the extracted real-part specific features, imaginary-part specific features, and shared features. This is the final reorganization characteristic; The feature weight combination is based on the scattering contribution and real / imaginary part characteristics of different regions of the SAR target, classifying them into different types of image regions. Predefined.

8. A SAR image target recognition system based on complex domain statistical property unwrapping and dynamic reassembly network, used to implement the SAR image target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network as described in any one of claims 1 to 7, characterized in that, include: The network building module is used to build a complex convolutional neural network based on ResNet34 as the initial network model. The module for unwrapping and dynamically recombining cross-channel features of real and imaginary parts is designed and integrated into the network architecture. The network training module is used to train the initial network model by taking the enhanced image as input and using amplitude consistency constraint loss, phase separability constraint loss and cross-entropy loss together to obtain the optimized network model. The image recognition module is used to identify the test image set using an optimized network model and obtain the recognition results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the SAR target recognition method based on complex domain statistical property unwrapping and dynamic reassembly network as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the SAR target recognition method based on unwrapping and dynamic reassembly of complex domain statistical properties as described in any one of claims 1 to 7.

Images