Space target ISAR image segmentation method based on complex shape prior learning

Abstract

The application provides a spatial target ISAR image segmentation method based on complex morphological prior learning, comprising: acquiring ISAR complex image data of a spatial target to be segmented, the ISAR complex image data of the spatial target to be segmented containing real part information and imaginary part information; inputting the ISAR complex image data of the spatial target to be segmented into a pre-trained complex morphological prior segmentation network for segmentation processing to obtain a spatial target ISAR image component segmentation result; wherein the pre-trained complex morphological prior segmentation network is jointly constituted based on a complex U-Net backbone network and a shape prior module. Based on this, a complete technical closed loop from "full use of complex data" to "global morphological prior embedding" and then to "local-global feature collaborative optimization" is realized, the inherent defects of existing real number CNN models in ISAR image segmentation are fundamentally overcome, and the precision and robustness of complex spatial target segmentation are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of ISAR image segmentation technology, specifically to a spatial target ISAR image segmentation method based on complex morphological prior learning. Background Technology

[0002] Inverse Synthetic Aperture Radar (ISAR) utilizes the relative motion between the target and the radar to acquire high-resolution two-dimensional images of non-cooperative targets, playing a crucial role in space target recognition, situational awareness, and status monitoring. Component segmentation of ISAR images aims to accurately separate pixel-level regions of different components such as the satellite body and solar panels from the imaging results, providing fundamental data support for subsequent target feature extraction, attitude estimation, and structural analysis. It is worth noting that the raw echo data from ISAR imaging is essentially in complex form, with its amplitude and phase information jointly reflecting the target's electromagnetic scattering characteristics and geometric features. However, current mainstream processing workflows typically only convert complex data into amplitude or grayscale images before inputting them into the segmentation network. This not only loses the fine structural and boundary clues contained in the phase information but also makes the network unable to handle the segmentation challenges of weakly scattering components, slender structures, and irregular contours commonly found in ISAR images, thus limiting further improvements in segmentation accuracy.

[0003] To address the challenges of ISAR image segmentation, deep learning-based segmentation methods have become the mainstream research approach. Existing technologies disclose various solutions based on improved U-Net architectures, such as fusing multi-scale features with attention mechanisms to enhance the model's robustness under small sample conditions; or introducing high-resolution feature pyramids and deformable convolutions to enhance the network's adaptability to multi-view and geometrically deformed targets. Furthermore, some schemes attempt to capture pixel-level correlations using 4D convolutions, or explicitly learn the scattering center distribution characteristics of different components by constructing scattering characteristic-guided mechanisms. These methods mostly focus on optimizing local feature extraction and fusion strategies, attempting to solve the problem of accurate identification of target components in complex backgrounds by enhancing the network's sensitivity to texture and intensity information.

[0004] While the aforementioned existing technologies have improved the segmentation performance of ISAR images to some extent, they still have inherent limitations. Existing deep learning models are generally built on real-valued convolutional neural networks, and their feature extraction process mainly relies on the amplitude information of the image. This fails to fully exploit the geometric structure and boundary information contained in the phase components of ISAR complex data, resulting in limited feature representation completeness. Furthermore, although existing network architectures possess strong local feature extraction capabilities, they are limited by the local receptive field of convolution operations, making it difficult to effectively model the overall geometry of the target. This lack of global structural constraints often leads to segmentation results exhibiting phenomena such as broken component outlines, missing slender structures, and missegmentation of local regions when dealing with complex geometric structures of spatial targets, making it difficult to guarantee the integrity of the target structure and the stability of its boundaries. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a spatial target ISAR image segmentation method based on complex morphological prior learning.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a spatial target ISAR image segmentation method based on complex morphological prior learning, comprising: Acquire ISAR complex image data of the spatial target to be segmented, which includes real and imaginary part information; The ISAR complex image data of the spatial target to be segmented is input into a pre-trained complex morphology prior segmentation network for segmentation processing to obtain the segmentation results of the spatial target ISAR image components. The pre-trained complex morphology prior segmentation network is composed of a complex U-Net backbone network and a shape prior module. The complex U-Net backbone network has an encoding stage and a decoding stage. The encoding stage is based on shallow complex feature encoding units and a complex domain residual downsampling module, while the decoding stage is based on multiple decoding layers and a segmentation head. The feature encoding unit includes complex convolutional layers, complex normalization layers, and complex activation layers; the complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps; the shape prior module operates on the residual connection process in the encoding and decoding stages, and is used to interact and fuse the learnable complex domain global shape prior with the multi-scale complex feature maps to impose target overall shape constraints on the multi-scale complex feature maps; the shape prior module is equipped with multiple cascaded shape prior sub-modules.

[0007] In a second aspect, the present invention provides a space target ISAR image segmentation device based on complex morphological prior learning, comprising: an acquisition unit and a segmentation processing unit; The acquisition unit is used to acquire ISAR complex image data of the spatial target to be segmented, which includes real part information and imaginary part information. The segmentation processing unit is used to input the ISAR complex image data of the space target to be segmented into a pre-trained complex morphology prior segmentation network for segmentation processing, thereby obtaining the segmentation results of the space target ISAR image components. The pre-trained complex morphology prior segmentation network is composed of a complex U-Net backbone network and a shape prior module. The complex U-Net backbone network has an encoding stage and a decoding stage. The encoding stage is based on shallow complex feature encoding units and a complex domain residual downsampling module, while the decoding stage is based on multiple decoding layers and a segmentation head. The shallow complex feature encoding unit includes a complex convolutional layer, a complex normalization layer, and a complex activation layer; the complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps; the shape prior module operates on the residual connection process in the encoding and decoding stages, and is used to interact and fuse the learnable complex domain global shape prior with the multi-scale complex feature maps to impose target overall shape constraints on the multi-scale complex feature maps; the shape prior module has multiple cascaded shape prior sub-modules.

[0008] This invention provides a method for segmenting ISAR images of space targets based on complex morphological prior learning, comprising: acquiring complex ISAR image data of a space target to be segmented, the complex ISAR image data of the space target to be segmented containing real and imaginary information; inputting the complex ISAR image data of the space target to be segmented into a pre-trained complex morphological prior segmentation network for segmentation processing to obtain segmentation results of space target ISAR image components; wherein, the pre-trained complex morphological prior segmentation network is composed of a complex U-Net backbone network and a shape prior module; the complex U-Net backbone network is configured with an encoding stage and a decoding stage, the encoding stage being based on shallow complex morphological prior learning. The system consists of a complex feature encoding unit and a complex domain residual downsampling module. The decoding stage is based on multiple decoding layers and a segmentation head. The shallow complex feature encoding unit includes a complex convolutional layer, a complex normalization layer, and a complex activation layer. The complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps. The shape prior module operates on the residual connection process of the encoding and decoding stages. It is used to interact and fuse the learnable complex domain global shape prior with the multi-scale complex feature maps to impose target overall shape constraints on the multi-scale complex feature maps. The shape prior module has multiple cascaded shape prior sub-modules. In this invention, firstly, by directly inputting ISAR complex image data containing both real and imaginary parts, a shallow feature encoding unit consisting of complex convolutional layers, complex normalization layers, and complex activation layers is constructed. This allows the network to utilize both amplitude and phase information simultaneously during the feature extraction stage, thus solving the problem of existing methods relying solely on amplitude information and failing to extract geometric structure and boundary information in the phase, significantly improving the completeness of feature representation. Then, a cascaded complex domain residual downsampling module is introduced in the encoding stage. Through the complex residual bottleneck unit, multi-scale complex feature maps are extracted step-by-step, enhancing the network's ability to capture local details. Simultaneously, a shape prior module operates on the residual connection process of encoding-decoding, utilizing learnable complex domain global shape priors and multi-scale features for interactive fusion. This introduces explicit constraints on the overall geometric shape of the target to the network, thus solving the problem that convolutional operations are difficult to model globally due to limited local receptive fields. Finally, relying on the step-by-step guidance of multiple cascaded sub-modules in the shape prior module, global shape constraints are integrated into the reconstruction process of multi-scale features, which effectively suppresses local missegmentation and enhances the continuity of component contours and the ability to maintain slender structures. This solves practical problems such as contour breakage, structural loss and boundary instability in the segmentation results, and ensures the integrity of the target structure and the stability of the boundary.In summary, this invention achieves a complete technical loop from "full utilization of complex data" to "global morphological prior embedding" and then to "local-global feature collaborative optimization," fundamentally overcoming the inherent defects of existing real-valued CNN models in ISAR image segmentation and significantly improving the accuracy and robustness of complex spatial target segmentation.

[0009] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a spatial target ISAR image segmentation method based on complex morphological prior learning provided in an embodiment of the present invention; Figure 2 An exemplary schematic diagram of the overall structure of a pre-trained complex morphological prior segmentation network is shown. Figure 3 A schematic diagram of the complex domain residual downsampling module is shown as an example; Figure 4 An exemplary structural diagram of the shape prior submodule is shown; Figure 5 A schematic diagram of ISAR image segmentation results based on the method of the present invention is shown as an example; Figure 6 A schematic diagram of a spatial target ISAR image segmentation device based on complex morphological prior learning provided in this embodiment of the invention. Detailed Implementation

[0011] This invention aims to address the shortcomings of ISAR image component segmentation in both complex information utilization and structural modeling. By introducing a complex convolutional neural network to jointly model amplitude and phase information, the network learns the spatial correlations and structural changes between scattering points, improving problems such as structural occlusion caused by missing phase information. Furthermore, by combining the global feature modeling advantages of the SPM module, the segmentation process is guided by the overall morphology of the target, effectively improving the accuracy of ISAR image component segmentation results.

[0012] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0013] To improve the accuracy and robustness of complex space target segmentation, this invention provides a space target ISAR image segmentation method based on complex morphological prior learning. Figure 1 This is a flowchart illustrating a spatial target ISAR image segmentation method based on complex morphological prior learning, as provided in an embodiment of the present invention. Figure 1 As shown, it includes: S101. Obtain ISAR complex image data of the spatial target to be segmented.

[0014] The ISAR complex image data of the spatial target to be segmented contains real and imaginary information.

[0015] S102. Input the ISAR complex image data of the space target to be segmented into the pre-trained complex morphology prior segmentation network for segmentation processing to obtain the segmentation results of the space target ISAR image components.

[0016] The pre-trained complex morphology prior segmentation network is composed of a complex U-Net backbone network and a shape prior module. The complex U-Net backbone network has an encoding stage and a decoding stage. The encoding stage is composed of shallow complex feature encoding units and complex domain residual downsampling modules, while the decoding stage is composed of multiple decoding layers and a segmentation head. The shallow complex feature encoding units include complex convolutional layers, complex normalization layers, and complex activation layers. The complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps. The shape prior module operates on the residual connection process of the encoding and decoding stages. It is used to interact and fuse the learnable complex domain global shape prior with the multi-scale complex feature maps to apply target overall morphological constraints to the multi-scale complex feature maps. The shape prior module has multiple cascaded shape prior sub-modules.

[0017] Figure 2 An exemplary schematic diagram of the overall structure of a pre-trained complex morphological prior segmentation network is shown, such as... Figure 2 As shown, the pre-trained Complex Morphology Prior Segmentation Network (CSPM-UNet) can effectively extract global and local features of amplitude and phase. This network structure employs complex convolution and complex feature extraction in both the encoder and decoder. By replicating features and encoding them sequentially across different network depths, shallow layers capture local features, while deeper layers enhance the ability to represent the overall geometric morphology. Through the ResNet-skip module (shallow complex feature encoding unit + complex domain residual downsampling module), CSPM-UNet can more effectively analyze the complex scattering characteristics of targets, making it suitable for high-precision component segmentation tasks.

[0018] The Shape Prior Module (SPM) provides explicit structural morphological constraints to the network, enhancing its ability to preserve the overall shape of the target during segmentation. In SPM, the network first generates an initial shape prior representation, then performs pixel-by-pixel alignment and structural consistency correction between this prior and the current features, establishing a connection between detailed regions and the overall geometry. This process allows the network to capture local texture while maintaining the integrity of the target's overall contour. Furthermore, SPM models the structural relationships between different regions through a shape consistency interaction mechanism between features, effectively suppressing boundary breaks and feature loss in weakly scattering regions, further improving the accuracy and structural coherence of the segmentation results.

[0019] Optionally, the shallow complex feature coding unit is composed of two cascaded initial combination modules; Each initial assembly module is configured to cascade sequentially. Complex convolutional layers, complex normalization layers, and complex activation layers; Each complex convolutional layer is equipped with a real part convolutional kernel and an imaginary part convolutional kernel; the real part convolutional kernel is used to represent the weights of the real part information, and the imaginary part convolutional kernel is used to represent the weights of the imaginary part information.

[0020] It should be noted that, in order to adapt to the complex number distribution, the complex number normalization layer in this invention adopts the batch normalization (BN) strategy, which normalizes the real part channel and the imaginary part channel of the input respectively, thus maintaining the statistical distribution stability of each component of the complex number.

[0021] The complex activation layers used in this invention each contain two independent ReLU activation units. During forward propagation, nonlinear activation is performed on the real and imaginary parts separately. Notably, in-place operations are explicitly disabled in this invention to avoid conflicts in gradient calculation paths during backpropagation.

[0022] Complex BN is used to accelerate model convergence and alleviate gradient vanishing, while complex ReLU activation function can introduce nonlinear transformations to improve the model's generalization ability.

[0023] Optionally, the output of the real part convolution kernel is represented as: ; The output of the imaginary part convolution kernel is represented as: ; in, This represents the output of the real part convolution kernel. This represents the output of the imaginary part convolution kernel. This represents the real part information of the ISAR complex image data of the spatial target to be segmented. Represents the real part convolution kernel. Represents the imaginary part of the convolution kernel. This represents the imaginary part information of the ISAR complex image data of the spatial target to be segmented. This represents a two-dimensional convolution operation.

[0024] In addition, to improve numerical stability, this invention performs non-in-situ normalization on the convolution kernel in each forward propagation, avoiding in-situ modification of the learnable parameters themselves, thereby ensuring the correctness of backpropagation.

[0025] In this embodiment, complex features are represented by tensor quantization by concatenating the real and imaginary part channels. For any complex feature map, its shape in the network is represented as follows: The four-dimensional tensor Indicates batch size. This represents the number of channels in a single section, where The front of the middle Each channel represents the real part, and then... Each channel represents the imaginary part. Indicates the height of the feature map. Indicates the width of the feature map.

[0026] Figure 3 A schematic diagram of the complex domain residual downsampling module is shown as an example. Figure 3 As shown, the complex domain residual downsampling module (ResNet-skip) includes a root processing unit (Root Block) and a multi-stage residual body connected in sequence; The root processing units are cascaded sequentially. Complex convolutional layers, complex normalization layers, complex activation layers, and Complex max pooling layers are constructed.

[0027] The root processing unit is used to rapidly reduce the spatial resolution of the feature map and increase the number of channels, preparing for subsequent deep feature extraction. The root processing unit uses a 7×7 convolutional kernel with a stride of 2 and padding of 3. To alleviate the difficulty of training deep complex networks, this... Complex convolutional layers integrate weight normalization techniques. During each forward propagation, the mean and variance of the real and imaginary weight matrices are calculated and normalized, which helps to smooth the loss surface and accelerate convergence. This maps the shallow features of the input to a higher-dimensional space while halving the dimensions of the feature map.

[0028] This is followed by a complex normalization layer. Unlike ordinary batch normalization, this complex normalization layer divides the channels into 32 groups for statistical normalization, making it insensitive to batch size. This layer contains two parallel normalization operators that independently normalize the real and imaginary channels of the feature map, preserving the phase statistical characteristics of the complex signal.

[0029] Next, nonlinear activation is performed through a complex activation layer, and ReLU operations are performed on the real and imaginary parts respectively.

[0030] Finally, to further compress the spatial dimensionality and extract salient features, a complex max-pooling layer with a 3×3 pooling kernel and a stride of 2 is used. The pooling operation is performed independently on the real and imaginary parts of the input; the output real part is a local maxima of the input real part, and the output imaginary part is a local maxima of the input imaginary part. This step halves the resolution of the feature map again.

[0031] The multi-stage residual volume is composed of a first complex residual bottleneck unit (Block1), a second complex residual bottleneck unit (Block2), and a third complex residual bottleneck unit (Block3) cascaded in sequence; each complex residual bottleneck unit is composed of multiple complex residual bottleneck sub-units. The complex residual bottleneck subunit is composed of a complex normalization layer, a complex activation layer, and a complex convolutional layer cascaded in sequence.

[0032] Optionally, the first complex residual bottleneck unit is provided with two complex residual bottleneck sub-units, and the stride of the complex convolutional layer in all complex residual bottleneck sub-units is 1. The second complex residual bottleneck unit has three complex residual bottleneck sub-units. The stride of the complex convolutional layer in the first complex residual bottleneck sub-unit is 2, and the stride of the complex convolutional layer in the other complex residual bottleneck sub-units is 1. The third complex residual bottleneck unit contains five complex residual bottleneck sub-units. The stride of the complex convolutional layer in the first complex residual bottleneck sub-unit is 2, while the stride of the complex convolutional layer in the other complex residual bottleneck sub-units is 1.

[0033] Furthermore, to support multi-scale feature fusion in the complex U-Net architecture decoder, this multi-stage residual construct employs a specific output mechanism. The network not only returns the output of the last layer but also collects the feature maps at the end of each stage (Root Block, Block1, Block2) through a list of features. In the implementation, to ensure that subsequent computations do not corrupt the gradient computation map due to in-place operations, a deep copy of the tensor is performed using the clone() operation when collecting features. These multi-scale complex feature maps contain rich information ranging from fine-grained texture to high-level semantics, and fully preserve phase features, ultimately being fed into the various layers of the decoder for reconstruction.

[0034] Continue to refer to Figure 2 The decoding stage consists of a first decoding layer, a second decoding layer, a third decoding layer, a fourth decoding layer, and a segmentation head; the first decoding layer, the second decoding layer, the third decoding layer, and the fourth decoding layer are all based on cascaded complex convolutional layers and upsampling layers; The shape prior module includes a first shape prior submodule, a second shape prior submodule, and a third shape prior submodule; The output of the shallow complex feature coding unit is connected to the input of the root processing unit and the input of the fourth decoding layer, respectively. The input of the first shape prior submodule is connected to the output of the root processing unit and the output of the second shape prior submodule, respectively; the output of the first shape prior submodule is connected to the input of the third decoding layer. The input of the second shape prior submodule is connected to the output of the first complex residual bottleneck unit and the output of the third shape prior submodule, respectively; the output of the second shape prior submodule is connected to the input of the second decoding layer. The input of the third shape prior submodule is connected to the output of the second complex residual bottleneck unit and the complex domain global shape prior, respectively; the output of the third shape prior submodule is connected to the input of the first decoding layer. The input of the first decoding layer is also connected to the output of the third complex residual bottleneck unit; the output of the first decoding layer is connected to the input of the second decoding layer. The output of the second decoding layer is connected to the input of the third decoding layer; the output of the third decoding layer is connected to the input of the fourth decoding layer; and the output of the fourth decoding layer is connected to the input of the segmentation head.

[0035] The Shape Prior Module (SPM) is a key module for solving boundary ambiguity problems in ISAR image segmentation. Embedded as an independent shape constraint flow within the segmentation network, it aims to address segmentation uncertainties in weak boundary regions by modeling the topological prior of the target structure in a high-dimensional parameter space. Traditional segmentation networks often rely solely on local texture and contrast information, easily leading to topologically incorrect segmentation results. SPM, however, introduces a set of learnable global shape parameters decoupled from the input image, forcing the network to reference standard target morphology during segmentation.

[0036] Optionally, the processing procedure of each shape prior submodule in the shape prior module is the same; the processing procedure of the shape prior submodule includes: The global shape prior in the complex domain is sequentially flattened, and the updated shape prior is obtained by feature refinement and global modeling through a self-attention mechanism and a multilayer perceptron. The updated shape prior is upsampled to the same spatial resolution as the current multi-scale complex feature map to obtain the upsampled feature map. The multi-scale complex feature map is used as the query and the upsampled feature map is used as the key and value. Cross-modal interaction is performed through a cross-attention mechanism, and the weighted fusion result is superimposed on the updated shape prior to obtain the superimposed result. The superposition result is then concatenated with the corresponding current multi-scale complex feature map to obtain the enhanced feature map. The enhanced feature map is input into the shape back-projection operator to calculate the shape residual update amount. The shape residual update amount is then superimposed on the updated shape prior to obtain a new prior for iterative updates of the next-level shape prior submodule.

[0037] The processing procedure is explained using formulas, taking the third shape prior submodule (SPM3) as an example.

[0038] ; in, This represents the updated shape prior. This represents a multilayer perceptron. This represents the prior global shape of the complex field. Represents the self-attention operator. Represents the complex number level normalization operator; ; The spatial reconstruction operator represents the shape prior. Represents the upsampled feature map; ; This indicates the result of the superposition. This represents the current multi-scale complex feature map. This represents the Softmax operation. Represents the dot product operation of complex numbers; ; Represents enhanced feature maps; ; This represents the shape residual update amount. Indicates the shape return operator; ; This represents the new prior that is iteratively updated in the next level shape prior submodule.

[0039] The formula for the self-attention mechanism is as follows: ; in , , These are matrices representing queries, keys, and values. Self-attention calculates the similarity between the query and all keys using dot products, and then weights each value accordingly. express or The feature dimension, which is used as a scaling factor to adjust the dot product result. The numerical range is set to prevent SoftMax gradient saturation, thereby stabilizing the final attention output.

[0040] Figure 4 An exemplary structural diagram of the shape prior submodule is shown. Now refer to... Figure 4 The processing flow of the shape prior submodule is explained in detail.

[0041] First, in the self-updating unit (the dashed box above), the input is the initialized global shape prior of the complex field (dimension ). , Represents the prior embedding dimension. Indicates spatial height, (Representing spatial width). This prior is first serialized and flattened, transforming its spatial dimensions to... The sequence vector. This sequence vector is used as the query ( ),key( ) and value ( Enter the self-attention mechanism. and Matrix multiplication (dot product) is performed to generate a self-affinity matrix describing the prior internal structural relationships. This matrix is ​​then processed using scaling and normalized exponential functions (Scale & Softmax) to obtain attention weights. These weights are applied to... The matrix is ​​used to perform feature weighting. Subsequently, the weighted features are superimposed onto the original input sequence through residual connections, and then further modeled for global context using complex-layer normalization and a multilayer perceptron (MLP). Finally, a reshaping operation restores the sequence features to a spatial tensor, generating an updated "global shape prior".

[0042] Secondly, in the cross-update unit (the dashed box below), the shape prior after self-update is first upsampled to adjust its spatial resolution to match the multi-scale complex image features of the current level. Figure 1 The dimensions are consistent, resulting in an upsampled feature map. At this point, the current multi-scale complex feature map is used as the query ( The above-sampled shape prior is used as the key. ) and value ( This involves calculating a pixel-level similarity matrix (i.e., cross-attention). After scaling and Softmax normalization, this similarity matrix acts as a weighted spatial attention mask. The weighted fusion result is superimposed onto the upsampled shape prior through residual connection to obtain the superimposed result.

[0043] Next, the superposition result is combined with the current multi-scale complex feature map (the combination processing after the output of the corresponding cross-update unit in the figure), and after processing by the residual block, an enhanced feature map that incorporates structural constraints is obtained.

[0044] Finally, to maintain the dynamic evolution of shape priors across levels, the enhanced feature map is input into the shape back-projection operator (corresponding to the Downsample arrow in the figure), and the shape residual update is calculated through a downsampling operation. This update is then superimposed back onto the initial (or previous level) shape prior at the input of the update unit through residual concatenation, resulting in a new prior for iterative updates of the next-level shape prior submodule. This cascaded design achieves dynamic constraints and iterative refinement of shape priors from coarse to fine.

[0045] Optionally, the training process of the pre-trained complex morphological prior segmentation network includes: Acquire ISAR complex image sample data of space targets; The ISAR complex image sample data of space targets are input into the initial complex morphological prior segmentation network, and iterative training is performed using a preset loss function; The initial complex form prior segmentation network that satisfies the iteration stopping condition is used as the pre-trained complex form prior segmentation network. The preset loss function is composed of both binary cross-entropy loss and Dice loss.

[0046] Optionally, the iteration stopping condition includes the number of iterations meeting an iteration threshold or the value of a preset loss function being continuously less than a loss threshold.

[0047] This invention provides a spatial target ISAR image segmentation method based on complex morphological prior learning. First, by directly inputting ISAR complex image data containing both real and imaginary parts, a shallow feature encoding unit is constructed, consisting of complex convolutional layers, complex normalization layers, and complex activation layers. This allows the network to utilize both amplitude and phase information simultaneously during feature extraction, thus solving the problem of existing methods relying solely on amplitude information and failing to extract geometric structure and boundary information in the phase, significantly improving the completeness of feature representation. Then, a cascaded complex domain residual downsampling module is introduced in the encoding stage. Through the complex residual bottleneck unit, multi-scale complex feature maps are extracted step-by-step, enhancing the network's ability to capture local details. Simultaneously, a shape prior module operates on the residual connection process of encoding-decoding, utilizing learnable complex domain global shape priors and multi-scale features for interactive fusion. This introduces explicit constraints on the overall geometric shape of the target to the network, thus solving the problem that convolutional operations are difficult to model globally due to limited local receptive fields. Finally, guided by the cascading sub-modules within the shape prior module, global morphological constraints are integrated throughout the multi-scale feature reconstruction process. This effectively suppresses local missegmentation and enhances the continuity of component contours and the preservation of slender structures. Consequently, it solves practical problems such as contour breaks, structural gaps, and boundary instability in the segmentation results, ensuring the integrity of the target structure and the stability of its boundaries. In summary, this invention achieves a complete technical loop from "full utilization of complex data" to "global morphological prior embedding" and then to "local-global feature collaborative optimization," fundamentally overcoming the inherent defects of existing real-valued CNN models in ISAR image segmentation and significantly improving the accuracy and robustness of complex spatial target segmentation.

[0048] To verify the effectiveness of the method of the present invention, simulation experiments were also conducted as follows.

[0049] The space targets selected were models of components such as Tiangong-1. To fully verify the effectiveness of the proposed method, ISAR images were acquired using high-fidelity electromagnetic computational imaging, with the horizontal and vertical axes representing the Doppler and range axes, respectively. This experiment used CSPM-UNet as the segmentation network and selected different components of the space targets as segmentation tasks for comparison, verifying the network's effectiveness at different scales and structures.

[0050] The target input consists of 4500 complex image data points, each 256×256 pixels. During training, the network parameters are set as follows: 200 training epochs, a batch size of 8, and the Adam optimizer is used to update the network parameters. The initial learning rate is set to 0.001, and the update strategy for this parameter is continuous. Adjustable Dice Similarity Coefficient If there is no improvement, the learning rate will decrease to half of its original value in the next round.

[0051] To quantify the segmentation results of this invention, the intersection-over-union ratio (IoU) is used. Mean Intersection over Union (MIU) ), Sensitivity ), Specificity ), accuracy Mean Absolute Error ) and mean squared error ( ) is used as an evaluation indicator. The definitions of different evaluation indicators are as follows: ; ; ; ; ; ; ; ; in, This represents the number of samples that were correctly predicted as positive. This represents the number of samples that were incorrectly predicted as positive. This represents the number of samples that were correctly predicted as negative. This represents the number of samples that were incorrectly predicted as negative. It is the number of pixels in the predicted region; This is the actual number of pixels in the label; It is the total number of pixels; Represents the total number of categories.

[0052] Analysis of experimental results.

[0053] Figure 5 An exemplary schematic diagram of ISAR image segmentation results based on the method of the present invention is shown, wherein... Figure 5 Figure (a) shows a schematic diagram of the original ISAR complex image sample. Figure 5Figure (b) shows the label map corresponding to the original ISAR complex image samples. Figure 5 Figure (c) shows a schematic diagram of the segmentation results based on the method of the present invention. Experimental indicators are shown in Table 1.

[0054] Table 1 Experimental Indicators

[0055] from Figure 5 As can be seen from the visualization results and Table 1, the CSPM-UNet proposed in this invention outperforms traditional semantic segmentation networks in ISAR image component segmentation tasks. This network can fully utilize phase information, segment components clearly, and effectively identify and extract the main body and the solar panel, demonstrating excellent segmentation performance.

[0056] This invention also includes an ablation experiment targeting SPM. During the experiment, four network variants were created for comparison: 1) a baseline model (without SPM); 2) a single-layer SPM model (SPM added only after the root processing unit, S1); 3) a two-layer SPM model (SPM added after the root processing unit and Block 1, S2); and 4) a complete model (SPM added after the root processing unit, Block 1, and Block 2, S3).

[0057] All models were trained using the same training parameters and training set. The experimental results are shown in Table 2.

[0058] Table 2 Ablation Experiment Results

[0059] As shown in Table 2, the performance-enhancing effect of SPM gradually increases with the number of layers. This indicates that SPM can provide explicit structural constraints to the network, further improving the accuracy and structural coherence of the segmentation results, thereby improving the model's performance.

[0060] The method provided in this embodiment of the invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc., and this embodiment of the invention does not limit the application to such devices.

[0061] Based on the same inventive concept, embodiments of the present invention also provide a space target ISAR image segmentation device based on complex morphological prior learning. Figure 6 A schematic diagram of a spatial target ISAR image segmentation device based on complex morphological prior learning provided in this embodiment of the invention is shown below. Figure 6 As shown, it includes: an acquisition unit 601 and a segmentation processing unit 602; The acquisition unit 601 is used to acquire ISAR complex image data of the spatial target to be segmented, which includes real part information and imaginary part information. The segmentation processing unit 602 is used to input the ISAR complex image data of the space target to be segmented into a pre-trained complex morphology prior segmentation network for segmentation processing, and obtain the segmentation results of the space target ISAR image components. The pre-trained complex morphology prior segmentation network is composed of a complex U-Net backbone network and a shape prior module. The complex U-Net backbone network has an encoding stage and a decoding stage. The encoding stage is based on a shallow complex feature encoding unit and a complex domain residual downsampling module, and the decoding stage is based on multiple decoding layers and a segmentation head. The shallow complex feature encoding unit includes a complex convolutional layer, a complex normalization layer, and a complex activation layer; the complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps; the shape prior module operates on the residual connection process in the encoding and decoding stages, and is used to interact and fuse the learnable complex domain global shape prior with the multi-scale complex feature maps to impose target overall shape constraints on the multi-scale complex feature maps; the shape prior module is equipped with multiple cascaded shape prior sub-modules.

[0062] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0063] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings and the disclosure, will understand and implement other variations of the disclosed embodiments in carrying out the claimed invention. In this description, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0064] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A spatial target ISAR image segmentation method based on complex morphological prior learning, characterized in that, include: Acquire ISAR complex image data of the spatial target to be segmented, wherein the ISAR complex image data of the spatial target to be segmented includes real part information and imaginary part information; The ISAR complex image data of the spatial target to be segmented is input into a pre-trained complex morphology prior segmentation network for segmentation processing to obtain the segmentation results of the spatial target ISAR image components. The pre-trained complex morphology prior segmentation network is composed of a complex U-Net backbone network and a shape prior module. The complex U-Net backbone network has an encoding stage and a decoding stage. The encoding stage is based on shallow complex feature encoding units and a complex domain residual downsampling module. The decoding stage is based on multiple decoding layers and... The structure comprises a shallow complex feature encoding unit including a complex convolutional layer, a complex normalization layer, and a complex activation layer; a complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps; a shape prior module operates on the residual connection process of the encoding and decoding stages, and is used to interactively fuse the learnable complex domain global shape prior with the multi-scale complex feature maps to apply target overall shape constraints to the multi-scale complex feature maps; The shape prior module is configured with multiple cascaded shape prior sub-modules.

2. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 1, characterized in that, The shallow complex feature encoding unit is composed of two cascaded initial combination modules; Each of the initial combination modules is configured to be cascaded sequentially. Complex convolutional layers, complex normalization layers, and complex activation layers; Each of the complex convolutional layers is provided with a real part convolutional kernel and an imaginary part convolutional kernel; wherein, the real part convolutional kernel is used to represent the weight of the real part information, and the imaginary part convolutional kernel is used to represent the weight of the imaginary part information.

3. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 2, characterized in that, The output of the real part convolution kernel is expressed as follows: ； The output of the imaginary part convolution kernel is expressed as follows: ； in, This represents the output of the real part convolution kernel. This represents the output of the imaginary part convolution kernel. This represents the real part information of the ISAR complex image data of the spatial target to be segmented. Represents the real part convolution kernel. Represents the imaginary part of the convolution kernel. This represents the imaginary part information of the ISAR complex image data of the spatial target to be segmented. This represents a two-dimensional convolution operation.

4. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 1, characterized in that, The complex domain residual downsampling module includes a root processing unit and a multi-stage residual body connected in sequence. The root processing units are cascaded sequentially. Complex convolutional layers, complex normalization layers, complex activation layers, and Complex max pooling layer construction; The multi-stage residual body is composed of a first complex residual bottleneck unit, a second complex residual bottleneck unit, and a third complex residual bottleneck unit cascaded in sequence; each of the complex residual bottleneck units is composed of multiple complex residual bottleneck sub-units; The complex residual bottleneck subunit is composed of a complex normalization layer, a complex activation layer, and a complex convolutional layer cascaded in sequence.

5. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 4, characterized in that, The first complex residual bottleneck unit is provided with two complex residual bottleneck sub-units, and the stride of the complex convolutional layer in all the complex residual bottleneck sub-units is 1. The second complex residual bottleneck unit is provided with 3 complex residual bottleneck sub-units, wherein the stride of the complex convolutional layer in the first complex residual bottleneck sub-unit is 2, and the stride of the complex convolutional layer in the other complex residual bottleneck sub-units is 1. The third complex residual bottleneck unit is provided with 5 complex residual bottleneck sub-units, wherein the stride of the complex convolutional layer in the first complex residual bottleneck sub-unit is 2, and the stride of the complex convolutional layer in the remaining complex residual bottleneck sub-units is 1.

6. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 4, characterized in that, The decoding stage includes a first decoding layer, a second decoding layer, a third decoding layer, a fourth decoding layer, and a segmentation head; the first decoding layer, the second decoding layer, the third decoding layer, and the fourth decoding layer are all based on cascaded complex convolutional layers and upsampling layers; The shape prior module includes a first shape prior submodule, a second shape prior submodule, and a third shape prior submodule; The output of the shallow complex feature encoding unit is connected to the input of the root processing unit and the input of the fourth decoding layer, respectively. The input terminal of the first shape prior submodule is connected to the output terminal of the root processing unit and the output terminal of the second shape prior submodule, respectively; the output terminal of the first shape prior submodule is connected to the input terminal of the third decoding layer. The input terminal of the second shape prior submodule is connected to the output terminal of the first complex residual bottleneck unit and the output terminal of the third shape prior submodule, respectively; the output terminal of the second shape prior submodule is connected to the input terminal of the second decoding layer. The input of the third shape prior submodule is connected to the output of the second complex residual bottleneck unit and the complex domain global shape prior, respectively; the output of the third shape prior submodule is connected to the input of the first decoding layer. The input of the first decoding layer is also connected to the output of the third complex residual bottleneck unit; the output of the first decoding layer is connected to the input of the second decoding layer. The output of the second decoding layer is connected to the input of the third decoding layer; the output of the third decoding layer is connected to the input of the fourth decoding layer; and the output of the fourth decoding layer is connected to the input of the segmentation head.

7. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 1, characterized in that, The processing procedure is the same for each shape prior submodule in the shape prior module; The processing procedure of the shape prior submodule includes: The global shape prior in the complex domain is sequentially flattened, and features are refined and global modeled using a self-attention mechanism and a multilayer perceptron to obtain the updated shape prior. The updated shape prior is upsampled to the same spatial resolution as the current multi-scale complex feature map to obtain an upsampled feature map. The current multi-scale complex feature map is used as the query and the upsampled feature map is used as the key and value. Cross-modal interaction is performed through a cross-attention mechanism, and the weighted fusion result is superimposed on the updated shape prior to obtain the superimposed result. The superposition result is combined with the corresponding current multi-scale complex feature map to obtain an enhanced feature map. The enhanced feature map is input into the shape back-projection operator to calculate the shape residual update amount. The shape residual update amount is then superimposed on the updated shape prior to obtain a new prior for iterative updates of the next-level shape prior submodule.

8. The space target ISAR image segmentation method based on complex morphological prior learning according to claim 1, characterized in that, The training process of the pre-trained complex morphological prior segmentation network includes: Acquire ISAR complex image sample data of space targets; The ISAR complex image sample data of the space target is input into the initial complex morphological prior segmentation network, and iterative training is performed using a preset loss function; The initial complex form prior segmentation network that satisfies the iteration stopping condition is used as the pre-trained complex form prior segmentation network. The preset loss function is composed of both binary cross-entropy loss and Dice loss.

9. A spatial target ISAR image segmentation method based on complex morphological prior learning according to claim 8, characterized in that, The iteration stopping conditions include the number of iterations meeting an iteration threshold or the value of a preset loss function being continuously less than a loss threshold.

10. A space target ISAR image segmentation device based on complex morphological prior learning, characterized in that, include: Acquisition unit and segmentation processing unit; The acquisition unit is used to acquire ISAR complex image data of the spatial target to be segmented, wherein the ISAR complex image data of the spatial target to be segmented includes real part information and imaginary part information; The segmentation processing unit is used to input the ISAR complex image data of the space target to be segmented into a pre-trained complex morphology prior segmentation network for segmentation processing, to obtain the segmentation result of the space target ISAR image components; wherein, the pre-trained complex morphology prior segmentation network is composed of a complex U-Net backbone network and a shape prior module; the complex U-Net backbone network is configured with an encoding stage and a decoding stage, the encoding stage is composed of a shallow complex feature encoding unit and a complex domain residual downsampling module, and the decoding stage is composed of a multi-dimensional complex feature encoding unit and a complex domain residual downsampling module. The system consists of a decoding layer and a segmentation head. The shallow complex feature encoding unit includes a complex convolutional layer, a complex normalization layer, and a complex activation layer. The complex domain residual downsampling module is used to downsample the input complex feature map step by step through cascaded complex residual bottleneck units to extract multi-scale complex feature maps. The shape prior module operates on the residual connection process of the encoding stage and the decoding stage, and is used to interact and fuse the multi-scale complex feature map with the learnable complex domain global shape prior to apply a target overall shape constraint to the multi-scale complex feature map. The shape prior module is configured with multiple cascaded shape prior sub-modules.

Images