Radar image segmentation method and device based on THAM-ResUNet

Abstract

The application discloses a radar image segmentation method and device based on THAM-ResUNet, and belongs to the technical field of ISAR image processing.The application is used to solve the problem that weak edge features of radar images relative to optical images result in difficulty in accurate segmentation.The ISAR space target image is obtained first, and then is sent to the THAM-ResUNet network for radar image segmentation; the THAM-ResUNet network is built based on the UNet network model, the THAM hybrid attention mechanism is arranged after each convolution module of the UNet network model, the output of the last THAM attention mechanism is subjected to a full connection layer to obtain the output vector of the THAM-ResUNet, and then the radar image segmentation is realized.

Description

Technical Field

[0001] This invention belongs to the field of ISAR image processing technology and relates to a radar image segmentation method and device based on a neural network model. Background Technology

[0002] Inverse Synthetic Aperture Radar (ISAR) is an important branch of synthetic aperture radar development. It can acquire detailed images of non-cooperative moving targets (such as aircraft, ships, and missiles) at long distances, in all weather conditions and at all times, making it of significant application value. In recent years, ISAR image segmentation techniques have played a crucial role in target feature extraction and recognition. However, radar images, compared to optical images, suffer from significant speckle noise, low signal-to-noise ratio, and weak edge features, making accurate segmentation difficult. Summary of the Invention

[0003] This invention aims to address the problem that radar images, compared to optical images, have weak edge features, making accurate segmentation difficult.

[0004] A radar image segmentation method based on THAM-ResUNet is proposed. First, an ISAR spatial target image is obtained, and then it is fed into the THAM-ResUNet network for radar image segmentation. The THAM-ResUNet network is built based on the UNet network model. Each convolutional module of the UNet network model is followed by a THAM hybrid attention mechanism. The output of the last THAM attention mechanism is passed through a fully connected layer to obtain the output vector of THAM-ResUNet, thereby realizing radar image segmentation.

[0005] The processing steps of the THAM attention mechanism include:

[0006] The input feature map G(a,x,y)=A(a,x,y)·exp(jφ(a,x,y)) is split into an amplitude spectrum A(a,x,y) and a phase spectrum φ(a,x,y), where a is the channel dimension, x is the range dimension, and y is the azimuth dimension; the phase change of adjacent frequency components is calculated:

[0007] Δφ(a,x,y)=φ(a,x,y-1)-φ(a,x,y)+φ(a,x,y+1)-φ(a,x,y)

[0008] Then, phase attention features are generated:

[0009] F φ =γ·Conv3D(Stack[φ(a,x,y),Δφ(a,x,y)])+b

[0010] Where Stack[·] is concatenation along the channel dimension, Conv3D represents three-dimensional convolution, b is the bias term, and γ is the weight parameter;

[0011] Then F φ The feature map G(a,x,y) is concatenated with the input feature map to obtain feature F. HAM attention processing is then applied to feature F to obtain THAM output features.

[0012] Furthermore, the process of performing HAM attention processing on feature F includes:

[0013] For feature F, average pooling and max pooling are respectively used to obtain and Based on parameters α and β, we obtain

[0014]

[0015] Then to One-dimensional fast convolution is performed to obtain channel refined feature maps; then a channel separation ratio parameter λ is introduced to divide the channel refined feature maps into important channel group F1′ and secondary important features F2′.

[0016] Average pooling and max pooling are performed on F1′, and the results of the two pooling operations are concatenated along the channel dimension to obtain a set of output features. The same processing is performed on F2′ to obtain another set of output features. For the two sets of output features, a shared convolutional layer with a kernel size of 7×7 is used for convolution to generate two sets of feature maps. These feature maps are then normalized and activated to obtain tensors. and The two tensors are multiplied by F1′ and F2′ respectively to obtain the spatial refinement features F1″ and F2″. The two are added element by element to obtain the THAM output features.

[0017] Furthermore, on In the process of performing one-dimensional fast convolution, the kernel size is... in Indicates closest Odd numbers.

[0018] Furthermore, the processing procedure of the THAM-ResUNet network built based on the UNet network model is as follows:

[0019] The input feature map first passes through a 7×7 convolutional module, and the output of the 7×7 convolutional module is processed by the THAM attention mechanism.

[0020] The output of the 7×7 convolutional module is fed into the bottleneck residual module. There are at least 4 bottleneck residual modules, and a THAM attention mechanism is introduced after each bottleneck residual module. Different bottleneck residual convolutional blocks include multiple bottleneck convolutional residual units. Each bottleneck residual convolutional unit has 3 convolutional layers. The input and the output of the last convolutional layer are concatenated. The residual concatenation in each bottleneck residual convolutional unit is processed by a ReLU activation layer.

[0021] The output of the last bottleneck residual module is fed into the basic residual convolutional modules. There are at least three basic residual convolutional modules, each with two convolutional layers. The residual connections in each basic residual convolutional module are processed by ReLU activation layers. A THAM attention mechanism is introduced after each basic residual convolutional block. After the THAM attention mechanism in the second-to-last and third-to-last basic residual convolutional modules, the skip connection dimension is adjusted by upsampling.

[0022] The THAM attention mechanism, set up after the 7×7 convolutional module and the first two bottleneck residual modules, performs skip connections with the last three basic residual convolutional blocks in the decoder.

[0023] Furthermore, the 7×7 convolution module is configured with a convolutional layer with a kernel size of 7×7, followed by a batch normalization layer, a max pooling layer, and a ReLU activation layer.

[0024] Furthermore, in the 7×7 convolution module, the convolution stride of the 7×7 convolution layer is 2 pixels, and the stride of the max pooling layer is 2 pixels.

[0025] Furthermore, the bottleneck residual module is set to 4, and the basic residual convolution module is set to 3;

[0026] The four bottleneck residual modules are denoted as Neck Residual Module 1 to Neck Residual Module 4. Bottleneck Residual Module 1 has three bottleneck convolutional residual units, with only the first convolutional unit being downsampled. Bottleneck Residual Module 2 has four bottleneck convolutional residual units, with only the first convolutional unit being downsampled. Bottleneck Residual Module 3 has six bottleneck convolutional residual units, with only the first convolutional unit being downsampled. Bottleneck Residual Module 4 has three bottleneck convolutional residual units.

[0027] Furthermore, the structure of the bottleneck residual convolutional unit is as follows:

[0028] When the bottleneck residual convolutional unit is set to downsample, the kernel size in the first convolutional layer is 1×1 pixels and the stride is 1 pixel; the kernel size in the second convolutional layer is 3×3 pixels and the stride is 2 pixels during downsampling, adjusting the residual connection dimension through the downsample branch; the kernel size in the third convolutional layer is 1×1 pixels and the stride is 1 pixel; each convolutional layer is followed by a batch normalization layer.

[0029] When the bottleneck residual convolutional unit does not have downsampling, the kernel size in the first convolutional layer is 1×1 pixels and the stride is 1 pixel; the kernel size in the second convolutional layer is 3×3 pixels and the stride is 1 pixel; the kernel size in the third convolutional layer is 1×1 pixels and the stride is 1 pixel; each convolutional layer is followed by a batch normalization layer.

[0030] Furthermore, the two convolutional layers of the basic residual convolutional module are as follows:

[0031] The kernel size in the first convolutional layer is 3×3 pixels, and the stride is 1 pixel.

[0032] The kernel size in the second convolutional layer is 3×3 pixels, and the stride is 1 pixel;

[0033] Each convolutional layer is followed by a batch normalization layer.

[0034] A radar image segmentation device based on THAM-ResUNet is disclosed. The device includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the radar image segmentation method based on THAM-ResUNet.

[0035] The beneficial effects of this invention are as follows: The THAM-ResUNet image segmentation framework proposed in this invention can segment ISAR images of space targets. First, this method designs a THAM (Triple Hybrid Attention Mechanism) based on the characteristics of radar images, innovatively designing phase attention to capture target defocus features. Then, key features are filtered through channel attention, spatial attention enhances the target edge response, and combined with residual structure, the detailed distribution characteristics of scattering points are effectively captured, resulting in better segmentation performance. Furthermore, this invention proposes a single-cycle strategy training optimization method, which can accelerate network training efficiency, reduce the possibility of model overfitting, and improve the model's generalization ability.

[0036] This invention utilizes deep learning technology to better meet the requirements of radar image segmentation under conditions such as significant speckle noise, low signal-to-noise ratio, and weak edge features. The network maintains a certain degree of stability in the face of changes in the viewpoint and noise of the input image. Attached Figure Description

[0037] Figure 1 This is a schematic diagram illustrating the generation of ISAR images of space targets and their corresponding labeled images;

[0038] Figure 2 This is an overall structural diagram of the THAM-ResUNet framework proposed in this invention;

[0039] Figure 3 This is the optimization effect of single-cycle strategy training. Detailed Implementation

[0040] To address the problems in the background technology, this patent proposes a radar image segmentation method based on THAM-ResUNet. This method introduces THAM to optimize feature selection, captures target defocus and texture features through phase attention, filters key features through channel attention, enhances target edge response through spatial attention, and combines residual structure to suppress gradient vanishing. It effectively captures the detailed distribution characteristics of scattering points using a deep network, significantly improving the segmentation accuracy of weakly scattering regions in radar images while suppressing background noise interference. Furthermore, the skip connection structure of UNet further enhances the fusion of multi-scale features, improving segmentation accuracy. Finally, an innovative single-cycle training optimization method effectively improves the model's generalization ability and increases network training efficiency.

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are also included.

[0042] Specific implementation method one: Combining Figures 1 to 2 This implementation method is described below.

[0043] The radar image segmentation method based on THAM-ResUNet described in this embodiment includes the following steps:

[0044] S1. Obtain the ISAR spatial target image dataset based on the three-dimensional model of the spatial target, and then obtain the corresponding label information through labelme. Divide the sample dataset composed of ISAR images and labeled images into training sample set and test sample set.

[0045] In the process of obtaining the ISAR space target image dataset, the ISAR imaging part uses the scattering point information of the three-dimensional model of the space target to perform ISAR simulation imaging according to the set radar parameters, and obtains the space target ISAR image dataset.

[0046] Labelme is used to annotate ISAR image datasets and obtain corresponding label information maps. During the annotation process, open Labelme via command line or application, import the ISAR image dataset, draw the annotation area on the image and enter the label name. After completion, a JSON-formatted annotation file is generated, containing the label information. The JSON file is then converted into a PNG-formatted label information map.

[0047] The ISAR image sequence dataset was then divided into a training set and a test set. The training set contained a label information map, in which each pixel had a label indicating whether the pixel belonged to the target. The test set did not contain a label information map.

[0048] S2. Construct a radar image segmentation network based on the THAM-ResUNet framework. Use ISAR images and corresponding label information maps as training samples to train THAM-ResUNet and obtain a training model. Then, use the trained segmentation model to perform segmentation tests on test ISAR image data and obtain the segmentation results of the test ISAR image data.

[0049] ISAR image data is input into the THAM-ResUNet network for processing. Label information maps are used as labels to train the THAM-ResUNet network. The ISAR image data size is 3×512×512, and the label information map data size is 1×512×512. Target image segmentation training and testing are performed on the THAM-ResUNet network architecture.

[0050] A THAM-ResUNet network based on the THAM triple hybrid attention mechanism is constructed for radar image segmentation.

[0051] S201 and THAM hybrid attention mechanism:

[0052] First, phase attention processing is performed. The input feature map G(a,x,y) = A(a,x,y)·exp(jφ(a,x,y)) can be decomposed into an amplitude spectrum A(a,x,y) and a phase spectrum φ(a,x,y), where a is the channel dimension, x is the range dimension, and y is the azimuth dimension. The phase spectrum reflects the phase shift caused by the Doppler effect of the target. This shift manifests as azimuth defocusing in the image, making the edges of the ISAR image more irregular. Severe defocusing can also lead to geometric distortion of the target. In addition, high-frequency phase changes correspond to the microstructure of scattering points, which is beneficial for capturing the texture features of the target. Defocusing features and target texture features can be captured by calculating the phase changes of adjacent frequency components.

[0053] Δφ(a,x,y)=φ(a,x,y-1)-φ(a,x,y)+φ(a,x,y+1)-φ(a,x,y) (1)

[0054] As shown in the above equation, only the azimuth dimension difference is used to extract phase changes while avoiding the introduction of irrelevant noise, and the boundaries are symmetrically filled. To address the impact of different defocusing phenomena on image segmentation results, a trainable parameter γ is used to generate phase attention features:

[0055] F φ =γ·Conv3D(Stack[φ(a,x,y),Δφ(a,x,y)])+b (2)

[0056] Where Stack[·] is the concatenation along the channel dimension, Conv3D uses a 1×3×3 three-dimensional convolution kernel, and b is the bias term.

[0057] For the weight parameter γ, phase difference is disabled when γ→0, degenerating to the baseline model; 0<γ<1, phase attention weights are weakened; γ>1, phase attention weights are strengthened.

[0058] Then perform channel attention processing: F φ The feature map G(a,x,y) is concatenated with the input feature map G(a,x,y) to obtain feature F. Feature F is then subjected to average pooling and max pooling respectively. and Average pooling and max pooling also play different roles at different stages of image feature extraction, and an adaptive mechanism was designed accordingly. and Each of these steps is multiplied by 1 / 2 by the trainable parameters α and β, then summed, and finally obtained by element-wise summation.

[0059]

[0060] Then to Perform one-dimensional fast convolution to capture the relationships between channels; the kernel size is [size missing]. in Indicates closest An odd number of channels are convolved to obtain a refined channel feature map. Then, a channel separation ratio parameter λ is introduced to divide the refined channel feature map into important channel group F1′ and secondary important features F2′.

[0061] Then, spatial attention is calculated based on F1′ and F2′. Average pooling and max pooling are then performed on F1′ and F2′ respectively, and the pooling results are concatenated along the channel dimension to obtain two sets of output features. These two sets of features are then convolved by a shared convolutional layer with a 7×7 kernel, generating two sets of feature maps of size H×W×1. These maps are then normalized and activated to obtain tensors. and The two tensors are multiplied by F1′ and F2′ respectively to obtain the spatial refinement features F1″ and F2″. The two are added element by element to obtain the final refinement features, namely the THAM output features.

[0062] S202. Constructing the network model framework:

[0063] like Figure 2 As shown, the THAM-ResUNet network contains three different convolutional modules: a 7×7 convolutional module, a bottleneck residual module, and a basic residual convolutional module.

[0064] The input image first passes through a 7×7 convolutional module, which is part of the encoder. It has one convolutional layer to expand the receptive field and capture coarse-grained features. The kernel size is 7×7 pixels and the stride is 2 pixels. Then it is followed by a batch normalization layer, a max pooling layer and a ReLU activation layer. The stride of the max pooling layer is 2 pixels. After the 7×7 convolutional module, the THAM attention mechanism is set.

[0065] The output of the 7×7 convolutional module is fed into the bottleneck residual module. Multiple bottleneck residual modules are configured; in this embodiment, four are used as part of the encoder. A THAM attention mechanism is introduced after each bottleneck residual module to dynamically calibrate features. Figure 2As shown, different bottleneck residual convolutional blocks each include multiple bottleneck convolutional residual units. Each bottleneck residual convolutional unit has three convolutional layers for the downsampling encoding part. The input and the output of the last convolutional layer (including the batch normalization layer) are residually connected. The kernel size of the first convolutional layer is 1×1 pixels, and the stride is 1 pixel. The kernel size of the second convolutional layer is 3×3 pixels, and the stride is 2 pixels during downsampling. The residual connection dimension is adjusted through the downsampling branch. The kernel size of the third convolutional layer is 1×1 pixels, and the stride is 1 pixel. Each convolutional layer is followed by a batch normalization layer. The residual connections in each bottleneck residual convolutional unit are processed by a ReLU activation layer.

[0066] In this embodiment, different bottleneck residual modules are provided with different numbers of bottleneck convolutional residual units. In this embodiment, bottleneck residual module 1 has 3 bottleneck convolutional residual units, and only the first convolutional unit is downsampled; bottleneck residual module 2 has 4 bottleneck convolutional residual units, and only the first convolutional unit is downsampled; bottleneck residual module 3 has 6 bottleneck convolutional residual units, and only the first convolutional unit is downsampled; bottleneck residual module 4 has 3 bottleneck convolutional residual units.

[0067] The output of the last bottleneck residual module is fed into the basic residual convolutional modules. Multiple basic residual convolutional modules are configured; in this embodiment, three are used as the decoder. Each basic residual convolutional module has two convolutional layers for upsampling the decoding part. The input and the output of the last convolutional layer are residually connected. The kernel size of the first convolutional layer is 3×3 pixels with a stride of 1 pixel, and the kernel size of the second convolutional layer is also 3×3 pixels with a stride of 1 pixel. Each convolutional layer is followed by a batch normalization layer. The residual connections in each basic residual convolutional module are processed by a ReLU activation layer. A THAM attention mechanism is introduced after each basic residual convolutional block to adaptively fuse features at multiple scales. After the THAM attention mechanisms in basic residual convolutional modules 1 and 2, the skip connection dimension is adjusted through upsampling.

[0068] A skip connection is designed between the downsampling encoding section and the upsampling decoding section, which can effectively fuse the multi-scale features of the encoding and decoding sections and alleviate the information loss caused by downsampling. In this embodiment, the THAM attention mechanism set after the 7×7 convolutional module and the first two bottleneck residual modules of the encoder is skipped to the last three basic residual convolutional blocks in the decoder.

[0069] The output of the final THAM attention mechanism is passed through a fully connected layer with an output dimension of 2 to obtain the output vector of THAM-ResUNet, thereby realizing the radar image segmentation method.

[0070] This implementation uses a residual network to construct the model. In a conventional neural network, each layer learns to map the input to the output. In residual learning, the goal of each layer is to learn the residual between the input mapping and the desired output; that is, the network attempts to fit the difference between the input and the desired output. Compared to ordinary networks, ResNet adds a short-circuit mechanism between every two or three convolutional layers, forming a residual learning structure. When a residual exists, it means that new features can be learned, thus enriching the feature information; conversely, it can maintain the network in a better state and alleviate the degradation problem. The basic residual block structure consists of two 3×3 convolutional layers, using skip connections to achieve residual learning. Deeper networks use a "bottleneck" residual block structure, with each residual block using 1×1, 3×3, and 1×1 convolutional layers to reduce the number of parameters and computational complexity.

[0071] The overall network architecture is similar to U. The encoder's main structure consists of four downsampling modules, a matched ReLU activation function, a BN layer, and a max-pooling layer, enabling the extraction of high-level image features and dimensionality reduction. For the decoder, its upsampling module comprises an upsampling convolutional layer, a skip connection, and a ReLU activation function. The upsampling convolutional layer progressively restores the image resolution, and the resulting feature map is concatenated with the feature information extracted from the corresponding downsampling layer through skip connections, outputting a feature map of the same size as the original image.

[0072] In some embodiments, the specific structural parameters of the THAM-ResUNet network are as follows:

[0073] In the downsampling encoding section, the output size of the 7×7 convolutional module is 64×256×256 pixels;

[0074] After being input into bottleneck residual module 1, it is passed to the first convolutional layer of the first bottleneck residual convolutional unit;

[0075] The first convolutional layer processes the input data and outputs a feature map of size 64×256×256 pixels. After being processed by the batch normalization layer, it is fed into the second convolutional layer.

[0076] The second convolutional layer performs downsampling with a stride of 2 pixels, outputting a feature map of size 64×128×128 pixels, which is then processed by a batch normalization layer and fed into the third convolutional layer.

[0077] The third convolutional layer performs dimension reorganization with a stride of 1 pixel, outputting a feature map of size 256×128×128 pixels. This map is then fused with the residual connections through the downsample branch, processed by the batch normalization layer and the ReLU layer, and passed to the other two bottleneck residual convolutional units of bottleneck residual module 1.

[0078] The other two bottleneck residual convolutional units of bottleneck residual module 1 are different from the first bottleneck residual convolutional unit only in the second convolutional layer. They do not require downsampling, have a stride of 1 pixel, and finally output a feature map of size 256×128×128 pixels through the THAM attention mechanism, which is then passed to bottleneck residual module 2.

[0079] Bottleneck residual module 2 has only one more bottleneck residual convolution unit than bottleneck residual module 1, and finally outputs a feature map of size 512×64×64 pixels, which is then passed to bottleneck residual module 3.

[0080] Bottleneck residual module 3 has only 3 more bottleneck residual convolution units than bottleneck residual module 1, and finally outputs a feature map of size 1024×32×32 pixels, which is then passed to bottleneck residual module 4.

[0081] Compared to bottleneck residual module 1, bottleneck residual module 4 does not perform downsampling processing in the second convolutional layer of the first bottleneck residual convolutional unit, with a stride of 1 pixel. Finally, it outputs a feature map of size 1024×64×64 pixels through the THAM attention mechanism and passes it to the upsampling decoding part.

[0082] In the upsampling decoding part, the input of each basic residual convolutional module is the result of the skip connection fusion of the output of the previous layer and the corresponding downsampling coding part;

[0083] The input to the first basic residual convolutional module is the 1024×64×64 pixel feature map from the previous layer plus the 512×64×64 pixel feature map before downsampling by the bottleneck residual module 3, which is then passed to the first convolutional layer of the first basic residual convolutional module.

[0084] The first convolutional layer performs upsampling with a stride of 1 pixel, outputting a feature map of size 256×64×64 pixels. After being processed by the batch normalization layer, it is fed into the second convolutional layer.

[0085] The second convolutional layer processes the input data and outputs a feature map of size 1024×64×64 pixels. After processing by the batch normalization layer and the ReLU layer, it outputs a feature map of size 512×128×128 pixels through upsampling and feeds it into the second basic residual convolutional module.

[0086] The input to the second basic residual convolution module is the 512×128×128 pixel feature map of the first basic residual convolution module plus the 256×128×128 pixel feature map before downsampling by the bottleneck residual module 2. The other steps are the same as the first basic residual convolution module. The output is a 256×256×256 pixel feature map to the third basic residual convolution module.

[0087] The input to the third basic residual convolutional module is the 256×256×256 pixel feature map of the second basic residual convolutional module plus the 64×256×256 pixel feature map before downsampling by the bottleneck residual module 1. The other steps are the same as the first basic residual convolutional module, and the output is a 64×512×512 pixel feature map to the fully connected layer.

[0088] The output size of the fully connected layer is 2×512×512.

[0089] Combination Figure 3 This invention demonstrates that the single-cycle strategy training optimization method can effectively accelerate network training efficiency. The training process is divided into four parts: two 30-epoch freeze training sessions and two 30-epoch unfreeze training sessions. Since the downsampling part has the most complex network structure, it is frozen during the freeze phase to keep the downsampling parameters unchanged and reduce model complexity. After multiple learning rate testing experiments, the learning rate was set to 0.001 during both the freeze and unfreeze phases. Due to the network's complexity and deep architecture, a weight decay value of 0.000001 was selected for both the freeze and unfreeze phases. In addition, to verify the model's convergence after four training sessions, an extra 10-epoch convergence test phase was set. This phase took place after the second unfreeze training, so the learning rate and weight decay were consistent with the unfreeze training to test whether the model was overfitting. Although THAM-ResUNet, which uses a single-cycle strategy to optimize training, does not improve as quickly as THAM-ResUNet in the early stages, it stabilizes after 120 epochs. This demonstrates that single-cycle training optimization is beneficial for accelerating model training efficiency.

[0090] S3. After training, the trained THAM-ResUNet model is obtained and used for actual image segmentation. In this embodiment, the ISAR image sequence in the test dataset is input into the model to obtain the image segmentation result. Specific Implementation Method Two:

[0092] This embodiment is a radar image segmentation device based on THAM-ResUNet. The device includes a processor and a memory. It should be understood that this includes any device including a processor and a memory described in this invention. The device may also include other units or modules that perform display, interaction, processing, control and other functions through signals or instructions.

[0093] The memory stores at least one instruction, which is loaded and executed by the processor to implement the radar image segmentation method based on THAM-ResUNet.

[0094] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions. Those skilled in the art will also understand that this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product implemented 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. The solutions in the embodiments of this application may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0095] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] 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.

[0097] 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.

[0098] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0099] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0100] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A radar image segmentation method based on THAM-ResUNet, characterized in that, First, ISAR spatial target images are obtained, and then fed into the THAM-ResUNet network for radar image segmentation. The THAM-ResUNet network is built based on the UNet network model. Each convolutional module of the UNet network model is followed by a THAM hybrid attention mechanism. The output of the last THAM attention mechanism is passed through a fully connected layer to obtain the output vector of THAM-ResUNet, thereby realizing radar image segmentation. The processing steps of the THAM attention mechanism include: Input feature map Decomposed into amplitude spectrum and phase spectrum Where 'a' is the channel dimension, 'x' is the range dimension, and 'y' is the azimuth dimension; calculate the phase change of adjacent frequency components: Then, phase attention features are generated: in, It is spliced ​​along the channel. Represents 3D convolution. It is a bias term; These are weight parameters; Then With the input feature map The concatenation process yields feature F, which is then subjected to HAM attention processing: feature F is obtained by average pooling and max pooling respectively. and Based on parameters and get : Then to One-dimensional fast convolution is performed to obtain channel refined feature maps, where the convolution kernel size is [size missing]. ,in Indicates closest The odd number; then introduce a channel separation ratio parameter. The refined feature map of the channel is divided into important channel groups. and secondary features ; right Average pooling and max pooling are performed, and the results of the two pooling methods are concatenated along the channel dimension to obtain a set of output features; The same processing is performed to obtain a set of output features. For both sets of output features, a shared convolutional layer with a 7×7 kernel is used to generate two sets of feature maps, which are then normalized and activated to obtain tensors. and The two tensors are respectively with and Multiplication yields refined spatial features. and The two are added element by element to obtain the THAM output features.

2. The radar image segmentation method based on THAM-ResUNet according to claim 1, characterized in that, The processing procedure of the THAM-ResUNet network, built based on the UNet network model, is as follows: The input feature map first passes through a 7×7 convolutional module, and the output of the 7×7 convolutional module is processed by the THAM attention mechanism. The output of the 7×7 convolutional module is fed into the bottleneck residual module. There are at least 4 bottleneck residual modules, and a THAM attention mechanism is introduced after each bottleneck residual module. Different bottleneck residual convolutional blocks include multiple bottleneck convolutional residual units. Each bottleneck residual convolutional unit has 3 convolutional layers. The input and the output of the last convolutional layer are concatenated. The residual concatenation in each bottleneck residual convolutional unit is processed by a ReLU activation layer. The output of the last bottleneck residual module is fed into the basic residual convolutional modules. There are at least three basic residual convolutional modules, each with two convolutional layers. The residual connections in each basic residual convolutional module are processed by ReLU activation layers. A THAM attention mechanism is introduced after each basic residual convolutional block. After the THAM attention mechanism in the second-to-last and third-to-last basic residual convolutional modules, the skip connection dimension is adjusted by upsampling. The THAM attention mechanism, set up after the 7×7 convolutional module and the first two bottleneck residual modules, performs skip connections with the last three basic residual convolutional blocks in the decoder.

3. The radar image segmentation method based on THAM-ResUNet according to claim 2, characterized in that, The 7×7 convolution module is configured with a convolutional layer with a kernel size of 7×7, followed by a batch normalization layer, a max pooling layer, and a ReLU activation layer.

4. The radar image segmentation method based on THAM-ResUNet according to claim 3, characterized in that, In the 7×7 convolution module, the convolution stride of the 7×7 convolution layer is 2 pixels, and the stride of the max pooling layer is 2 pixels.

5. The radar image segmentation method based on THAM-ResUNet according to claim 2, characterized in that, The bottleneck residual module is set to 4, and the basic residual convolution module is set to 3; The four bottleneck residual modules are denoted as Neck Residual Module 1 to Neck Residual Module 4. Bottleneck Residual Module 1 has three bottleneck convolutional residual units, with only the first convolutional unit being downsampled. Bottleneck Residual Module 2 has four bottleneck convolutional residual units, with only the first convolutional unit being downsampled. Bottleneck Residual Module 3 has six bottleneck convolutional residual units, with only the first convolutional unit being downsampled. Bottleneck Residual Module 4 has three bottleneck convolutional residual units.

6. The radar image segmentation method based on THAM-ResUNet according to claim 5, characterized in that, The structure of the bottleneck residual convolutional unit is as follows: When the bottleneck residual convolutional unit is set to downsample, the kernel size in the first convolutional layer is 1×1 pixels and the stride is 1 pixel; the kernel size in the second convolutional layer is 3×3 pixels and the stride is 2 pixels during downsampling, adjusting the residual connection dimension through the downsample branch; the kernel size in the third convolutional layer is 1×1 pixels and the stride is 1 pixel; each convolutional layer is followed by a batch normalization layer. When the bottleneck residual convolutional unit does not have downsampling, the kernel size in the first convolutional layer is 1×1 pixels and the stride is 1 pixel; the kernel size in the second convolutional layer is 3×3 pixels and the stride is 1 pixel; the kernel size in the third convolutional layer is 1×1 pixels and the stride is 1 pixel; each convolutional layer is followed by a batch normalization layer.

7. The radar image segmentation method based on THAM-ResUNet according to claim 5, characterized in that, The two convolutional layers of the basic residual convolutional module are as follows: The kernel size in the first convolutional layer is 3×3 pixels, and the stride is 1 pixel. The kernel size in the second convolutional layer is 3×3 pixels, and the stride is 1 pixel; Each convolutional layer is followed by a batch normalization layer.

8. A radar image segmentation device based on THAM-ResUNet, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement a radar image segmentation method based on THAM-ResUNet as described in any one of claims 1 to 7.

Images