Methods, apparatus, equipment and media for attitude angle estimation of inverse synthetic aperture radar images

By using a heterogeneous model collaborative architecture and hybrid loss function fusion training, the problems of insufficient accuracy and real-time performance in ISAR image attitude angle estimation are solved, achieving efficient and accurate attitude angle prediction, which is suitable for attitude angle estimation of non-cooperative targets.

CN123118752APending Publication Date: 2026-11-17NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511480646.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing ISAR image attitude angle estimation methods suffer from problems such as accuracy dependence on prior models, difficulty in obtaining them, and insufficient real-time performance on non-cooperative targets. In particular, traditional methods are prone to losing information during feature extraction.

Method used

A heterogeneous model collaborative architecture is adopted, in which a first pre-trained model captures global features and a second pre-trained model captures local detailed features. The two models are then combined with a hybrid loss function for fusion training to construct a fusion network for pose angle prediction.

Benefits of technology

It improves the accuracy and real-time performance of attitude angle estimation, enables efficient feature extraction without relying on prior models, achieves performance complementarity in different angle ranges, and enhances the robustness of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118752A_ABST
    Figure CN123118752A_ABST
Patent Text Reader

Abstract

This invention relates to a method, apparatus, device, and medium for attitude angle estimation of inverse synthetic aperture radar (ISAR) images. The method includes: acquiring a dataset; constructing a first pre-trained model and a second pre-trained model; inputting ISAR images and corresponding attitude labels into the first and second pre-trained models respectively for training, capturing global features of the ISAR images through the first pre-trained model and local detail features through the second pre-trained model, resulting in trained first and second models; fusing the first and second models based on a pre-set hybrid loss function, fusing local detail features with global features to obtain a trained fusion network; and outputting attitude angle prediction values ​​through the fusion network. This invention enables complementary performance in prediction across different angle intervals, significantly improving the accuracy of attitude angle estimation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of radar technology, and in particular to a method, apparatus, device, and medium for estimating the attitude angle of inverse synthetic aperture radar images. Background Technology

[0002] With the iterative evolution of space technology and the increasing diversification of space missions, accurate estimation of spacecraft attitude has become a crucial support for ensuring the smooth operation of complex tasks such as on-orbit servicing and maintenance, occupying an important position in the aerospace field. However, for some spacecraft, especially non-cooperative spacecraft, their surfaces often lack distinctive features for attitude calculation, which poses a challenge to attitude estimation. ISAR technology, with its all-weather, all-time operation and high imaging resolution, has become an effective technical means for spacecraft attitude and pointing estimation.

[0003] Spacecraft attitude and pointing estimation methods based on ISAR images can be divided into two categories: The first is model matching methods. The technical logic involves pre-constructing a 3D model of the target, traversing possible attitude parameters of the target and projecting them onto the imaging plane to generate a simulated image. Then, by measuring the feature similarity between the simulated image and the measured image, the attitude parameters corresponding to the optimal similarity are used as the estimation result. However, this method has a limitation: its accuracy largely depends on the accurate prior model of the target, and obtaining such a model is often difficult for non-cooperative targets. The second category is attitude estimation methods based on 3D reconstruction, which reconstruct the 3D structure of the target from a multi-view image sequence and then calculate the target's attitude. The most representative method is factorization. This method first extracts and tracks the dominant scatterer from the ISAR image sequence and constructs a trajectory matrix. Then, it obtains the 3D geometry and attitude by decomposing the trajectory matrix. However, because this method relies on a long observation interval to obtain the disparity angle, it is difficult to apply to slowly rotating targets and lacks real-time performance in practical applications. Compared with traditional feature extraction methods, target axis estimation based on neural networks shows advantages in both accuracy and speed. However, given the characteristics of ISAR images, existing methods, due to their use of multi-level max pooling operations, may result in the loss of feature information, thereby affecting the comprehensiveness of feature extraction. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, device, and medium for inverse synthetic aperture radar image attitude angle estimation that does not rely on prior models, can directly use ISAR images for axial regression estimation, and ensures high estimation accuracy, in order to address the above-mentioned technical problems.

[0005] An attitude angle estimation method for inverse synthetic aperture radar images, the method comprising: Obtain a dataset, which includes ISAR images and attitude labels corresponding to the ISAR images; Construct the first pre-trained model and the second pre-trained model; The ISAR image and the pose label corresponding to the ISAR image are respectively input into the first pre-trained model and the second pre-trained model for training. The first pre-trained model captures the global features of the ISAR image, and the second pre-trained model captures the local detail features of the ISAR image, thus obtaining the trained first model and the second model. Based on a pre-set hybrid loss function, the first model and the second model are fused and trained, and local detail features and global features are fused to obtain a trained fusion network. The fusion network outputs the attitude angle prediction value.

[0006] An inverse synthetic aperture radar image attitude angle estimation device, the device comprising: A dataset acquisition module is used to acquire a dataset, which includes ISAR images and attitude labels corresponding to the ISAR images; The model building module is used to build the first pre-trained model and the second pre-trained model; The pre-training module is used to input the ISAR image and the pose label corresponding to the ISAR image into the first pre-training model and the second pre-training model respectively for training. The first pre-training model captures the global features of the ISAR image, and the second pre-training model captures the local detail features of the ISAR image, so as to obtain the trained first model and the second model. The fusion training module is used to perform fusion training on the first model and the second model based on a pre-set fusion loss function, and to fuse local detail features with global features to obtain a trained fusion network. The attitude angle estimation module is used to output the attitude angle prediction value through the fusion network.

[0007] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the inverse synthetic aperture radar image attitude angle estimation method.

[0008] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the inverse synthetic aperture radar image attitude angle estimation method.

[0009] Compared with existing technologies, the inverse synthetic aperture radar image attitude angle estimation method, apparatus, device, and medium provided by this invention have the following beneficial effects: 1. Breaking through the limitations of traditional single-model feature capture, this approach employs a heterogeneous model collaborative architecture to construct dual-processing branches and simultaneously implement differentiated preprocessing. The first model focuses on global features to avoid overall pose misjudgment caused by local interference; the second model focuses on local detailed features to compensate for the loss of detailed information in global features. Finally, a hybrid loss function is used to achieve collaborative learning between the two models, which enhances the robustness of the models, achieves complementary performance in prediction across different angle intervals, enables comprehensive feature extraction, and significantly improves the accuracy of pose angle estimation.

[0010] 2. By leveraging the ability of neural networks to directly extract features, axial regression estimation can be performed directly using ISAR images without relying on prior models. This optimizes the estimation process, improves computational efficiency, and enhances real-time performance. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention, and those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0012] Figure 1 This is an application environment diagram of the inverse synthetic aperture radar image attitude angle estimation method provided in Example 1; Figure 2 This is a schematic diagram of the imaging results of the ALOS satellite at different attitude angles provided in Example 1; Figure 3 This is a schematic diagram of the structure of the first pre-trained model provided in Example 1; Figure 4 This is a schematic diagram of the structure of the second pre-trained model provided in Example 1; Figure 5 This is a schematic diagram of the prediction results of the first model provided in Example 1, wherein, Figure 5 (a) is a schematic diagram comparing the predicted and actual values ​​of the first yaw angle. Figure 5 (b) is a schematic diagram comparing the predicted value and the actual value of the first pitch angle; Figure 6 This is a schematic diagram of the prediction results of the second model provided in Example 1, wherein, Figure 6 (a) is a schematic diagram comparing the predicted and actual values ​​of the second yaw angle. Figure 6 (b) is a schematic diagram comparing the predicted value and the actual value of the second pitch angle; Figure 7 This is a schematic diagram of the prediction results of the fusion network provided in Example 1, wherein, Figure 7 (a) is a schematic diagram comparing the final predicted yaw angle with the actual value. Figure 7(b) is a schematic diagram comparing the final pitch angle prediction value with the actual value; Figure 8 This is a schematic diagram of the estimation error of each interval of the first model provided in Example 1, wherein, Figure 8 (a) is a schematic diagram of the yaw angle prediction error in different pitch angle ranges. Figure 8 (b) is a schematic diagram of pitch angle prediction error in different pitch angle ranges; Figure 9 This is a schematic diagram of the estimation error of each interval of the second model provided in Example 1, wherein, Figure 9 (a) is a schematic diagram of the second yaw angle prediction error in different pitch angle ranges. Figure 9 (b) is a schematic diagram of the second pitch angle prediction error in different pitch angle intervals; Figure 10 This is a schematic diagram of the estimation error of each interval of the fusion network provided in Example 1, wherein, Figure 10 (a) is a schematic diagram of the final yaw angle prediction error in different pitch angle ranges. Figure 10 (b) is a schematic diagram of the final pitch angle prediction error in different pitch angle ranges; Figure 11 This is a structural block diagram of an inverse synthetic aperture radar image attitude angle estimation device in one embodiment; Figure 12 This is an internal structural diagram of a computer device in one embodiment.

[0013] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0015] It should be noted that in this invention, the use of terms such as "first," "second," etc., is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0016] It is understood that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Example 1 like Figure 1 As shown, an attitude angle estimation method for inverse synthetic aperture radar images is provided, including the following steps: Step 201: Obtain the dataset, which includes ISAR images and pose labels corresponding to the ISAR images.

[0019] Step 202: Construct the first pre-trained model and the second pre-trained model.

[0020] Step 203: Input the ISAR image and the pose label corresponding to the ISAR image into the first pre-trained model and the second pre-trained model respectively for training. The first pre-trained model captures the global features of the ISAR image, and the second pre-trained model captures the local detail features of the ISAR image, so as to obtain the trained first model and the second model.

[0021] Step 204: Based on the pre-set hybrid loss function, perform fusion training on the first model and the second model, and fuse local detail features with global features to obtain the trained fusion network.

[0022] Step 205: Output the attitude angle prediction value through the fusion network.

[0023] The aforementioned inverse synthetic aperture radar (ISAR) image attitude angle estimation method overcomes the limitations of traditional single-model feature capture. It employs a heterogeneous model collaborative architecture to construct a dual-processing branch, simultaneously implementing differentiated preprocessing. The first model focuses on global features, avoiding overall attitude misjudgment caused by local interference; the second model focuses on local detail features, compensating for the loss of detailed information in global features. Finally, a hybrid loss function enables collaborative learning between the two models, enhancing model robustness, achieving complementary performance in predictions across different angle intervals, and significantly improving attitude angle estimation accuracy. This dual-model fusion method can simultaneously make judgments based on overall structure and key details, improving the comprehensiveness of feature extraction, reducing estimation bias caused by incomplete features, and has wide application scenarios. Furthermore, this invention eliminates the need to construct complex prior models, leveraging the ability of neural networks to directly extract features, enabling direct axial regression estimation using ISAR images, optimizing the estimation process, improving computational efficiency, and enhancing real-time performance.

[0024] The method provided by this invention is characterized by high efficiency, accuracy, and strong adaptability. It can be flexibly adapted to more types of ISAR imaging targets and has a wide range of applications.

[0025] In the specific implementation of step 201, the dataset used is the SAR image dataset from the ALOS satellite, containing 1387 images with a resolution of 256×256. When it is necessary to set the instantaneous attitude of the satellite target, it can be achieved by adjusting its Euler angles, namely yaw, pitch, and roll. These three angles are determined based on the angle between the target's coordinate system and the orbital coordinate system; when the two coordinate systems coincide, all Euler angles are 0°.

[0026] To reduce data volume, the roll angle in the dataset is fixed at 0°, while the yaw angle is set to 0° to 90° and the pitch angle to 0° to 360°, with each angle varying in 5° increments. In the dataset, the content of the ISAR image is represented by "X," and the attitude label corresponding to the ISAR image is the yaw angle. and pitch angle Imaging results of the ALOS satellite at different attitude angles are as follows: Figure 2 As shown. Further, the dataset undergoes preprocessing such as normalization and noise addition to achieve data augmentation.

[0027] In the specific implementation of steps 202 and 203, the first pre-trained model includes a first backbone network, a first branch network, and a second branch network.

[0028] The ISAR image is input into the backbone network for global feature extraction, and the first feature map is output. The first feature map is then input into the first branch network and the second branch network for processing, and the first yaw angle prediction value and the first pitch angle prediction value are output.

[0029] Based on the pre-set first loss function, the loss of the first yaw angle prediction value, the first pitch angle prediction value, and the attitude label is calculated; the first pre-trained model is optimized according to the loss to obtain the trained first model.

[0030] Specifically, such as Figure 3As shown, the first backbone network mainly includes multiple convolutional layers, batch normalization layers, and corrected linear units. In this embodiment, there are four convolutional layers, namely, convolutional layer 1 (3→16, 5×5 kernel, stride=2), convolutional layer 2 (16→32, 3×3 kernel, stride=2), convolutional layer 2_1 (32→64, 3×3 kernel, stride=2), and convolutional layer 3 (64→64, 3×3 kernel, stride=2) connected in sequence. Among them, convolutional layer 1, convolutional layer 2, and convolutional layer 2_1 are all connected to batch normalization layers and corrected linear units, and the feature map size is halved and the number of channels is increased by the convolution stride; convolutional layer 3 is connected to the first branch network and the second branch network respectively. During training, mean squared error (MSE) is used as the first loss function. The angle value directly output by the linear layer can directly measure the squared difference between the predicted value and the true value, so that the model learns the continuous value mapping. In the first backbone network, the modified linear unit adopts the ReLU activation function, thereby introducing nonlinearity and mitigating gradient vanishing.

[0031] The first and second branch networks have the same structure, both using 1×1 skip-connection convolutional layers for channel adjustment (64→128, stride=2), followed by skip-connection batch normalization layers and skip-connection corrected linear units for processing, and then processed by flattening layers connected to fully connected layers, finally outputting one-dimensional predicted values ​​for the first yaw angle and the first pitch angle, respectively. The input to the first branch network comes directly from the output of convolutional layer 4, while the second branch network obtains features from convolutional layer 4 through skip connections, achieving the fusion and utilization of features from different levels. In both branch networks, the corrected linear units also use the ReLU activation function, thus introducing non-linearity and mitigating gradient vanishing.

[0032] In the first pre-trained model, basic features of the input image, such as shallow information like edges and textures, are quickly extracted through continuous convolutional operations (combined with stride for dimensionality reduction). Then, the first yaw angle prediction and the first pitch angle prediction are directly output through two branch networks. Furthermore, the design of skip connections between the first backbone network and the second branch network enhances the fusion capability of features at different levels.

[0033] Understandably, the first pre-trained model uses a lightweight convolutional structure to construct an end-to-end direct regression estimation framework. By directly utilizing ISAR images for axial regression estimation, it completes the basic mapping from image to angle and extracts global features.

[0034] The second pre-trained model includes a second backbone network, a third branch network, and a fourth branch network.

[0035] The second backbone network includes a residual module and an attention module. The residual module and attention module enhance the input ISAR image to obtain the second feature map.

[0036] The third branch network has the same structure as the fourth branch network, both including a coarse-grained and fine-grained dual-layer output structure. The coarse-grained and fine-grained dual-layer output structure is used to extract local detail features from the second feature map and output the second yaw angle prediction value and the second pitch angle prediction value.

[0037] Based on the pre-set second loss function, the loss of the second yaw angle prediction value, the second pitch angle prediction value, and the attitude label is calculated; the second pre-trained model is optimized according to the loss to obtain the trained second model.

[0038] Specifically, such as Figure 4 As shown, the second backbone network mainly includes multiple pose residual modules, adaptive average pooling layers, and random activation layers. In this embodiment, there are four residual modules, namely pose residual block 1, pose residual block 2, pose residual block 3, and pose residual block 4 connected in sequence, with the output of the previous pose residual block serving as the input of the next pose residual block. Each pose residual block has the same structure, containing a global average pooling layer, two fully connected layers, and a logistic function. During processing, the features are first compressed through the global average pooling layer, and then the number of channels is changed through the fully connected layers to output residual maps of different dimensions. The logistic function is the Mish activation function, used to smooth nonlinear transformations. The four residual modules sequentially increase the number of channels from 3→16→32→64→128, and all use convolutions with stride=2 to reduce the dimensionality of the feature maps.

[0039] After each pose residual block, an attention module is connected. This module generates channel weights, which are then multiplied channel-by-channel with the residual map output from each pose residual block to enhance key feature representations and output an enhanced feature map. The enhanced feature map is then passed through a global average pooling layer to unify the features to 1×1×128 and is evaluated. Finally, a random activation layer (p=0.4) is used to suppress overfitting, outputting a second feature map. In the second backbone network, the residual module uses the PoseBlock residual module, the random activation layer uses a dropout layer, and the attention module uses an SE channel attention mechanism.

[0040] The third and fourth branch networks have the same structure, the difference being that the third branch network is used for yaw angle prediction, while the fourth branch network is used for pitch angle prediction. Both the third and fourth branch networks adopt a coarse-fine (Coarse+Fine) two-layer output structure. The coarse-grained fully connected layers for yaw and pitch angle are both linear layers (128→36), while the fine-grained fully connected layers for yaw and pitch angle are each composed of 36 parallel linear layers (128→36). The final output is the sum of the results from the coarse-grained and fine-grained fully connected layers for yaw and pitch angle, respectively (36 dimensions), i.e., the joint distribution results for yaw and pitch angle. This hierarchical prediction framework of "coarse classification - fine regression" can improve the accuracy of angle prediction.

[0041] Finally, the joint distribution results of yaw angle and pitch angle are temperature-scaled and two-stage decoded by the distribution decoder to output the second yaw angle prediction value and the second pitch angle prediction value.

[0042] In this embodiment, the second model introduces multiple residual modules, attention modules, and a coarse-fine granular dual-layer output structure to capture deep and complex features, generating second yaw angle prediction values ​​and second pitch angle prediction values ​​with local details. In the subsequent fusion stage, the preliminary results output by the first model are corrected and optimized using the second yaw angle prediction values ​​and the second pitch angle prediction values, which can ensure the accuracy of angle prediction.

[0043] In addition, regarding angle prediction, unlike the first model which directly regresses continuous angle values, the second model adopts a "coarse-grained hierarchical prediction" method (interval + offset), which maps continuous angles to 36 discrete intervals and assigns probabilities through linear interpolation. Then, the probabilities are diffused to adjacent intervals through dispersion. The sum of probabilities of the 36 intervals is calculated, and the expected value is taken as the interval center. Finally, the offset distribution of each interval is calculated, and the expected value is taken as the fine offset within the interval. This achieves coarse-grained hierarchical prediction of "coarse-grained interval division -> fine-grained offset optimization -> distributed decoding encoding", thus completing the pose estimation of the target.

[0044] Furthermore, in the training of the second model, a second loss function was designed to be specifically optimized for angular characteristics, and its expression is: ; ; ; In the formula, Indicates coarse-grained loss; Indicates fine-grained loss; Indicates dynamic weights; This represents the total number of ISAR image samples; Indicates the first The yaw angle of each sample belongs to the true range. The probability of; Indicates the first The pitch angle of each sample belongs to the true range. The probability of; This represents the actual yaw angle offset; This indicates the predicted yaw angle offset; This represents the actual pitch angle offset; This represents the predicted pitch angle offset.

[0045] It is understandable that the second loss function focuses on the boundary jump problem. By correcting the angle cycle error and improving the discretization accuracy, the prediction stability of the trained second model in the small angle range is improved, thus making the method proposed in this invention perform better in scenarios with high angle continuity requirements.

[0046] In addition, when training the first pre-trained model and the second pre-trained model, the input ISAR image dimension is 256×256×3, the dimension of the output data of the first pre-trained model is 2×1, and the dimension of the output data of the second pre-trained model is 36×36.

[0047] The training, validation, and test sets were divided in a 7:1:2 ratio. The Adam optimizer was used with a learning rate of 0.001, and the total training epochs were set to 200. By pre-training the two models separately, we can ensure that each model achieves optimal convergence, preparing for subsequent fusion.

[0048] In the specific implementation of step 204, the first pre-trained model and the second pre-trained model are loaded respectively, and fusion inference training is performed through a pre-set hybrid loss function to obtain the trained fusion network.

[0049] The expression for the hybrid loss function is: ; in: ; In the formula, Represents dynamic weights, with values ​​ranging from 1 to 2. ; Represents the first loss function; This represents the second loss function; Represents the consistency coefficient; This represents the consistency loss function, used to constrain the differences in the outputs of two models and avoid conflicts. This represents the total number of ISAR image samples in the training set of the dataset; This represents the predicted value of the first yaw angle; This represents the predicted value of the second yaw angle; This represents the predicted value of the first pitch angle; This represents the predicted value of the second pitch angle.

[0050] Furthermore, during the fusion training of the first and second models, it is necessary to jointly fine-tune the first and second models, and then reload the two finely-tuned models to finally obtain the trained fusion network.

[0051] It's understandable that in the standalone pre-training stage, the first pre-trained model uses a "convolutional layer + skip connection + continuous value regression" architecture to capture global image features. While it performs stably in the large angle range of 60° to 360°, its sensitivity to fine features in the small angle range of 0° to 60° is too low, making it prone to smoothing errors. The second pre-trained model, on the other hand, uses a "residual block + attention mechanism + coarse-grained hierarchical prediction" architecture. It achieves discrete angle estimation through probability distribution decoding, thus demonstrating stronger ability to capture local details in the small angle range of 0° to 60°. It can be seen that the first pre-trained model is better suited to its dependence on global brightness features and exhibits better stability in scenes with varying illumination over large angle ranges. The second pre-trained model enhances robustness to local features, making it particularly suitable for learning texture features in small angle ranges.

[0052] This embodiment achieves differentiated data processing through two independent models, amplifying the processing capabilities of global and local features respectively. In the fusion stage, by fusing the trained first and second models, they become complementary, improving prediction accuracy and robustness, achieving complementary performance across different angle intervals, and enabling the synergistic application of regression and binning prediction models in pose estimation. This solves the problem of uneven performance of a single model across different angle intervals.

[0053] In one embodiment, the effectiveness of the method proposed in this invention is verified.

[0054] The test environment was an NVIDIA RTX A2000 graphics card equipped with the PyTorch framework. As shown in Table 1, the fused network showed a significant advantage over the single-model network in terms of the average estimation error of yaw and pitch angles, with RMSE values ​​of 3.4101 and 28.0224, respectively.

[0055] Table 1. Average estimation errors for yaw and pitch

[0056] right Figures 5 to 7Analysis revealed that the structure of the fusion network is superior to the first and second models.

[0057] To further explore the advantages of the fusion network, the pitch angle range was subdivided into 12 intervals, each interval being 30°. Figures 8 to 10 The estimation error results for each interval are shown in detail.

[0058] Analysis revealed that the first model exhibited relatively large estimation errors for yaw angles in the pitch angle ranges of 0°–30° and 30°–60°, while achieving relatively good estimation results in other ranges. The second model showed reduced estimation errors in the pitch angle ranges of 0°–30° and 30°–60°, but its higher network complexity prevented optimal parameter optimization, resulting in a larger overall error. The fusion network, combining the advantages of both models, achieved superior performance in pitch and yaw angle prediction across all ranges.

[0059] This invention employs two complementary network models to perform differentiated preprocessing on input images. The first model uses a lightweight convolutional structure to achieve end-to-end direct regression, while the second model introduces a residual attention mechanism to construct a hierarchical prediction framework of "coarse classification – fine regression." Based on this, a fusion network is built using the first and second models, resulting in complementary performance in prediction across different angle intervals and improved pose prediction accuracy.

[0060] This invention employs a phased training strategy of "independent pre-training - joint fine-tuning" to optimize parameters. First, the single model converges to a local optimum, and then the dual model is collaboratively learned through a hybrid loss function. Ultimately, the fusion network outperforms the prediction performance of using a single model in the full angular range estimation of pitch and yaw angles.

[0061] It should be understood that, although this embodiment Figure 1 The steps are shown sequentially as indicated by the arrows, but they are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0062] Example 2 Based on the inverse synthetic aperture radar image attitude angle estimation method in Example 1, this example discloses an inverse synthetic aperture radar image attitude angle estimation device, such as... Figure 11 As shown, the inverse synthetic aperture radar image attitude angle estimation device includes: a dataset acquisition module 401, a model building module 402, a pre-training module 403, a fusion training module 404, and an attitude angle estimation module 405, wherein: The dataset acquisition module 401 is used to acquire a dataset, which includes ISAR images and attitude labels corresponding to the ISAR images.

[0063] The model building module 402 is used to build the first pre-trained model and the second pre-trained model.

[0064] The pre-training module 403 is used to input the ISAR image and the pose label corresponding to the ISAR image into the first pre-training model and the second pre-training model for training, respectively. The first pre-training model captures the global features of the ISAR image, and the second pre-training model captures the local detail features of the ISAR image, so as to obtain the trained first model and the second model.

[0065] The fusion training module 404 is used to perform fusion training on the first model and the second model based on a pre-set fusion loss function, and to fuse local detail features with global features to obtain a trained fusion network.

[0066] The attitude angle estimation module 405 is used to output the attitude angle prediction value through the fusion network.

[0067] In this embodiment, the specific working process and working principle of the dataset acquisition module 401, model building module 402, pre-training module 403, fusion training module 404, and attitude angle estimation module 405 are the same as those in Embodiment 1, and therefore will not be described again in this embodiment. Each unit module can be implemented entirely or partially through software, hardware, or a combination thereof. Each unit module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each of the above unit modules.

[0068] Example 3 This embodiment provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 12As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores inverse synthetic aperture radar (InSAR) image attitude angle estimation data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements an InSAR image attitude angle estimation method.

[0069] It is important to note that the aforementioned memory can be either standalone or integrated with the processor. When the memory is set up independently, the computer device also includes a bus for connecting the memory and the processor.

[0070] Example 4 This embodiment discloses a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the method in Embodiment 1 above.

[0071] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for estimating attitude angles in inverse synthetic aperture radar images, characterized in that, The method includes: Obtain a dataset, which includes ISAR images and attitude labels corresponding to the ISAR images; Construct the first pre-trained model and the second pre-trained model; The ISAR image and the pose label corresponding to the ISAR image are respectively input into the first pre-trained model and the second pre-trained model for training. The first pre-trained model captures the global features of the ISAR image, and the second pre-trained model captures the local detail features of the ISAR image, thus obtaining the trained first model and the second model. Based on a pre-set hybrid loss function, the first model and the second model are fused and trained, and local detail features and global features are fused to obtain a trained fusion network. The fusion network outputs the attitude angle prediction value.

2. The inverse synthetic aperture radar image attitude angle estimation method according to claim 1, characterized in that, The first pre-trained model includes a first backbone network, a first branch network, and a second branch network; The ISAR image is input into the first backbone network for global feature extraction, and the first feature map is output. The first feature map is input into the first branch network and the second branch network respectively for processing, and the first yaw angle prediction value and the first pitch angle prediction value are output. Based on a pre-set first loss function, the loss between the first yaw angle prediction value, the first pitch angle prediction value, and the attitude label is calculated; the first pre-trained model is optimized according to the loss to obtain a trained first model.

3. The inverse synthetic aperture radar image attitude angle estimation method according to claim 1, characterized in that, The second pre-trained model includes a second backbone network, a third branch network, and a fourth branch network; The second backbone network includes a residual module and an attention module. The residual module and the attention module are used to enhance the input ISAR image to obtain a second feature map. The third branch network has the same structure as the fourth branch network, both including a coarse-grained and fine-grained dual-layer output structure. The coarse-grained and fine-grained dual-layer output structure is used to extract local detail features from the second feature map and output the second yaw angle prediction value and the second pitch angle prediction value. Based on the pre-set second loss function, the loss between the second yaw angle prediction value, the second pitch angle prediction value, and the attitude label is calculated; the second pre-trained model is optimized according to the loss to obtain the trained second model.

4. The inverse synthetic aperture radar image attitude angle estimation method according to claim 3, characterized in that, The second loss function expression is: ; ; ; In the formula, Indicates coarse-grained loss; Indicates fine-grained loss; Indicates dynamic weights; This represents the total number of ISAR image samples; Indicates the first The yaw angle of each sample belongs to the true range. The probability of; Indicates the first The pitch angle of each sample belongs to the true range. The probability of; This represents the actual yaw angle offset; This indicates the predicted yaw angle offset; This represents the actual pitch angle offset; This represents the predicted pitch angle offset.

5. The inverse synthetic aperture radar image attitude angle estimation method according to any one of claims 1 to 4, characterized in that, The expression for the hybrid loss function is: ; In the formula, Indicates dynamic weights; Represents the first loss function; This represents the second loss function; Represents the consistency coefficient; This represents the consistency loss function.

6. The inverse synthetic aperture radar image attitude angle estimation method according to claim 5, characterized in that, The consistency loss function is expressed as follows: ; In the formula, This represents the total number of ISAR image samples in the training set of the dataset; This represents the predicted value of the first yaw angle; This represents the predicted value of the second yaw angle; This represents the predicted value of the first pitch angle; This represents the predicted value of the second pitch angle.

7. The inverse synthetic aperture radar image attitude angle estimation method according to any one of claims 1 to 4, characterized in that, The process of fusing and training the first and second models also includes: The first model and the second model are jointly fine-tuned to obtain the trained fusion network.

8. An attitude angle estimation device for inverse synthetic aperture radar images, characterized in that, The device includes: A dataset acquisition module is used to acquire a dataset, which includes ISAR images and attitude labels corresponding to the ISAR images; The model building module is used to build the first pre-trained model and the second pre-trained model; The pre-training module is used to input the ISAR image and the pose label corresponding to the ISAR image into the first pre-training model and the second pre-training model respectively for training. The first pre-training model captures the global features of the ISAR image, and the second pre-training model captures the local detail features of the ISAR image, so as to obtain the trained first model and the second model. The fusion training module is used to perform fusion training on the first model and the second model based on a pre-set fusion loss function, and to fuse local detail features with global features to obtain a trained fusion network. The attitude angle estimation module is used to output the attitude angle prediction value through the fusion network.

9. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the inverse synthetic aperture radar image attitude angle estimation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the inverse synthetic aperture radar image attitude angle estimation method according to any one of claims 1 to 7.