Axial estimation method, device and equipment for space target based on convolutional neural network

Abstract

The present invention relates to a method, device and equipment for axial estimation of space targets based on convolutional neural networks. The method comprises: obtaining a data set; constructing a space target axial estimation model, the space target axial estimation model comprising a backbone network and a branch network; inputting an ISAR image and a corresponding label into the backbone network, the backbone network extracts features by combining a channel attention mechanism in a feature inverse mapping manner to obtain weighted optimization features; there are two branch networks, the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target; the loss function is calculated based on the yaw angle and the pitch angle, the space target axial estimation model is iteratively optimized by the loss function to obtain a trained space target axial estimation model; posture estimation is performed by the trained space target axial estimation model. The method of the present invention improves the performance of the model in complex tasks, and has higher accuracy, robustness and adaptability.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a method, device and equipment for axial estimation of space targets based on convolutional neural networks. Background Art

[0002] Space targets are divided into two categories: cooperative and non-cooperative. Cooperative targets refer to those space targets that actively cooperate with observation or operation systems. They usually have a variety of sensors (such as accelerometers, gyroscopes, GPS, attitude sensors, etc.), and can transmit data in real time to ground stations or other space equipment through communication systems. For example, the vast majority of communication satellites, scientific exploration satellites, and navigation satellites are cooperative targets. Non-cooperative targets refer to those space targets that cannot actively transmit status information. Such targets may lose communication functions or are not designed to have cooperative capabilities, resulting in the inability to obtain key information such as their position and attitude. For example, space debris, failed satellites, etc. are usually regarded as non-cooperative targets.

[0003] At present, optical imaging telescopes or ISAR are usually used as the basis for attitude estimation of non-cooperative targets. However, in practical applications, due to the influence of factors such as sunlight, earth shadow, atmospheric turbulence, and weather, the timeliness and quality of adaptive optical images are difficult to guarantee. ISAR has all-day and all-weather imaging capabilities and is not affected by light. It has stronger adaptability. However, ISAR images lack texture information and have angular glint problems, which makes it difficult to extract features based on deep learning methods.

[0004] Although the traditional space target axial estimation method can perform space target axial estimation based on adaptive optical images or ISAR images. Space target axial estimation methods are mainly divided into methods based on three-dimensional reconstruction and methods based on template matching. Among them, the three-dimensional reconstruction-based methods, whether factorization or neural radiation field, are realized through multiple ISAR images. However, due to the angular scintillation phenomenon of the target scattering center in ISAR imaging, and the special structures such as cavities, the projection matrix cannot be accurately obtained, which will cause the reconstructed three-dimensional structure to have a large error; in addition, this method has higher requirements for the signal-to-noise ratio of the imaging location. When the signal-to-noise ratio is low, it is difficult to obtain the desired three-dimensional structure. Template matching-based methods are usually based on manual feature extraction techniques, such as Hough transform, inverse Jordan transform, etc., which are used to match the specified target structure features, such as line edges, rectangular sailboards, but for the complex structure of satellites and the diversity of possible satellite attitudes, the difficulty of feature extraction based on template matching methods will increase, thereby affecting the robustness of space target attitude estimation. Summary of the invention

[0005] Based on this, it is necessary to provide a method, device and equipment for axial estimation of spatial targets based on convolutional neural networks, which can comprehensively extract features and reduce feature information loss, in order to address the above technical problems.

[0006] A method for estimating the axial direction of a space target based on a convolutional neural network, the method comprising:

[0007] Acquire a data set, the data set comprising ISAR images at different attitude angles and labels corresponding to the ISAR images;

[0008] A space target axial estimation model is constructed, wherein the space target axial estimation model includes a backbone network and a branch network; the ISAR image and the corresponding label are input into the backbone network as a group of training samples, and the backbone network extracts features of the ISAR image by combining a channel attention mechanism with feature inverse mapping to obtain weighted optimization features; there are two branch networks, and the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target;

[0009] Calculating a loss function based on the yaw angle and the pitch angle, and iteratively optimizing the space target axial estimation model through the loss function to obtain a trained space target axial estimation model;

[0010] The position and posture of the space target in the SAR image to be detected is estimated by using the trained space target axial estimation model.

[0011] A spatial target axial estimation device based on a convolutional neural network, the device comprising:

[0012] A data acquisition module, used to acquire a data set, wherein the data set includes ISAR images at different attitude angles and labels corresponding to the ISAR images;

[0013] A data processing module is used to construct an axial estimation model for a space target, wherein the axial estimation model for a space target includes a backbone network and a branch network; the ISAR image and the corresponding label are input into the backbone network as a group of training samples, and the backbone network extracts features of the ISAR image by combining a channel attention mechanism with feature inverse mapping to obtain weighted optimization features; there are two branch networks, and the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target;

[0014] A model training module, used for calculating a loss function based on the yaw angle and the pitch angle, and iteratively optimizing the space target axial estimation model through the loss function to obtain a trained space target axial estimation model;

[0015] The pose estimation module is used to estimate the pose of the space target in the SAR image to be detected by using the trained space target axial estimation model.

[0016] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the spatial target axial estimation method based on a convolutional neural network when executing the computer program.

[0017] The above-mentioned convolutional neural network-based space target axial estimation method, device and equipment first obtain a data set, which includes ISAR images at different attitude angles and labels corresponding to the ISAR images; then build a space target axial estimation model, which includes a backbone network and a branch network; the ISAR image and the corresponding label are input into the backbone network as a group of training samples, and the backbone network extracts features from the ISAR image by combining the channel attention mechanism with feature inverse mapping to obtain weighted optimization features; there are two branch networks, and the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target; the loss function is calculated based on the yaw angle and the pitch angle, and the space target axial estimation model is iteratively optimized through the loss function to obtain a trained space target axial estimation model; finally, the posture of the space target in the SAR image to be detected is estimated through the trained space target axial estimation model.

[0018] The space target axial estimation model constructed by the present invention can make the model perform feature mapping more accurately by means of feature inverse mapping, so that the feature dimension reduction process will not lose important information for the regression task; combined with the channel attention mechanism, the overall trend of global information is captured, and the significant features of the target area are highlighted; the combination of the two can fully capture the global and local characteristics of the ISAR image, thereby achieving good performance. The loss function designed by the present invention can adjust the weight according to the complexity of the sample; at the same time, it also takes into account the periodic characteristics of the space target posture, effectively avoiding the non-periodic prediction of the angle error; through the design of the loss function, the model's understanding of the correlation between different tasks is enhanced, and it can handle the uncertainty of different tasks, improve the performance of the model in complex tasks, and have higher accuracy, robustness and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0020] Figure 1 A schematic diagram of the flow of the spatial target axial estimation method based on a convolutional neural network provided in Example 1;

[0021] Figure 2 This is a structural block diagram of the space target axial estimation model provided in Example 1;

[0022] Figure 3 This is a structural block diagram of the residual layer provided in Example 1;

[0023] Figure 4 This is a block diagram of the channel attention layer structure provided in Example 1;

[0024] Figure 5 Schematic diagram of the yaw angle and pitch angle estimation error results under different pitch angle ranges provided in Example 1; wherein, Figure 5 (a) is a schematic diagram of the yaw angle estimation error results under different pitch angle ranges. Figure 5 (b) is a schematic diagram of the pitch angle estimation error results under different pitch angle ranges;

[0025] Figure 6 This is the thermal diagram when the yaw angle is 30° provided in Example 1; wherein, Figure 6 (a) is the heat map when the yaw angle is 30° in the first convolutional layer. Figure 6 (b) is the heat map when the yaw angle is 30° in the last convolutional layer;

[0026] Figure 7 This is the thermal diagram when the pitch angle is 150° provided in Example 1; wherein, Figure 7 (a) is the heat map when the pitch angle is 150° in the first convolutional layer. Figure 7 (b) is the heat map when the pitch angle is 150° in the last convolutional layer;

[0027] Figure 8 It is a structural block diagram of the space target axial estimation device based on convolutional neural network provided in Example 2;

[0028] Fig. 9 This is a diagram of the internal structure of the computer device provided in Example 3.

[0029] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] It should be noted that in the present invention, descriptions such as "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0032] It can be understood that the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0033] The following will describe the implementation of the present invention in detail with reference to the accompanying drawings in the embodiment diagram of the present invention.

[0034] Example 1

[0035] This embodiment discloses a method for estimating the axial direction of a space target based on a convolutional neural network. Without the need to reconstruct the three-dimensional result of the target, the ISAR image is directly used for axial regression estimation, and an end-to-end estimation method is established. Without increasing the complexity of the system, the feature extraction capability of the neural network is effectively utilized to avoid the traditional method based on manual feature extraction, and it has the characteristics of high efficiency and robustness. It mainly constructs a space target axial estimation model, and then through the method of feature inverse mapping, the model can be more accurately mapped, so that the dimension reduction process of the feature will not lose important information for the regression task; at the same time, it combines the channel attention mechanism to capture the overall trend of global information and highlight the significant features of the target area; the combination of the two can fully capture the global and local characteristics of the ISAR image, thereby achieving good performance. The loss function designed by the present invention can adjust the weight according to the complexity of the sample; at the same time, it also considers the periodic characteristics of the attitude of the space target, effectively avoiding the non-periodic prediction of the angle error; through the design of the loss function, the model's understanding of the correlation between different tasks is enhanced, and the uncertainty of different tasks can be handled, and the performance of the model in complex tasks is improved, with higher accuracy, robustness and adaptability.

[0036] like Figure 1 As shown, the spatial target axial estimation method based on convolutional neural network provided in this embodiment includes the following steps:

[0037] Step 201: Acquire a data set, where the data set includes ISAR images at different attitude angles and labels corresponding to the ISAR images.

[0038] Step 202, constructing a space target axial estimation model, the space target axial estimation model includes a backbone network and a branch network; the ISAR image and the corresponding label are input into the backbone network as a group of training samples, the backbone network extracts features of the ISAR image by combining the channel attention mechanism through feature inverse mapping, and obtains weighted optimization features; there are two branch networks, and the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target.

[0039] Step 203, calculating a loss function based on the yaw angle and the pitch angle, iteratively optimizing the space target axial estimation model through the loss function, and obtaining a trained space target axial estimation model.

[0040] Step 204: estimate the position and posture of the space target in the to-be-detected SAR image using the trained space target axial estimation model.

[0041] In the specific implementation process of step 201, the data set mainly uses the ALOS satellite image data set, which contains 1387 images with a resolution of 512×512. The data set is processed. For example, to set the instantaneous attitude of the satellite target, its Euler angles can be adjusted: yaw, pitch, and roll. These three angles are determined according to the angle between the target body coordinate system and the orbital coordinate system. When the two coordinate systems coincide, all Euler angles are 0°. It is worth noting that in order to reduce the amount of data, the roll angle of the data set is fixed to 0°. At the same time, the yaw angle is set to 0° to 90°, and the pitch angle is set to 0° to 360°, and the change interval of both is 5°. After such processing, ISAR images at different attitude angles are obtained, and then data amplification is performed on the data set using methods such as normalization and noise addition to obtain a data set for model training. Among them, the ISAR image is used Indicates that the corresponding label yaw angle is and the pitch angle is .

[0042] In the specific implementation process of step 202, if Figures 2 to 4As shown in the figure, the structure of the space target axial estimation model is shown. The space target axial estimation model includes a backbone network and two branch networks. In the figure, C, K, S, P and B represent the number of output channels, convolution kernel, stride, padding size and bias respectively; the activation layer is the ReLU activation function.

[0043] The backbone network consists of several residual layers and several convolutional layers connected in sequence; the residual layer extracts features through low-to-high dimensional channels to obtain deep features; the convolutional layer reduces the dimensionality of the deep features from high to low to perform feature inverse mapping. After each convolutional layer, there is also a channel attention layer; the channel attention layer captures global and local features and finally outputs weighted optimized features.

[0044] The residual layer includes a parallel first residual branch and a second residual branch; after the image features are respectively input into the first residual branch and the second residual branch for processing, the features output by the first residual branch are added to the features output by the second residual branch, and the deep features are output after passing through the activation layer.

[0045] The channel attention layer includes a parallel first attention branch and a second attention branch; after the image features are input into the first attention branch and the second attention branch for processing, the features output by the first attention branch are added to the features output by the second attention branch, and after passing through the activation layer, the weighted optimization features are output.

[0046] Specifically, Figure 2 As shown, the backbone network includes four convolutional layers and three residual layers. Among them, convolutional layer 1 is in the first place, and it also has a batch normalization layer and an activation layer, which is mainly used to extract shallow features. After convolutional layer 1, the maximum pooling layer is connected, and then residual layer 1, residual layer 2 and residual layer 3 are connected in sequence. The structures of residual layers 1 to 3 are the same, and the difference lies in the different channel dimensions. Since deep features are extracted through the residual layer, the channel dimensions of residual layers 1 to 3 are arranged from low to high. After residual layer 3, convolutional layer 2, convolutional layer 3 and convolutional layer 4 are connected in sequence. Then two branch networks are connected respectively. The two branch networks have the same structure, both of which include an adaptive maximum pooling layer, a convolutional layer and a linear layer; after the convolutional layer, a batch normalization layer and an activation layer are connected to facilitate normalization, activation and other processing. Among them, the first branch network is used to output the yaw angle, and the second branch network is used to output the pitch angle.

[0047] When processing data, convolution layer 1 outputs the extracted shallow features to the residual layer. Residual layers 1 to 3 gradually increase the channel dimension from 64 to 512 to obtain deep features. Then convolution layers 2 to 4 reduce the channel dimension to 64 to achieve the feature inverse mapping process, which greatly reduces the amount of calculation of the linear layer in the later stage. In convolution layers 2 to 4, the channel attention layer is used to capture global and local features, avoiding the problem that important information may be lost in the feature dimensionality reduction process. In both branch networks, an adaptive maximum pooling layer is used to compress the image features of each channel into a single value. In addition, for full connection, the output dimension is 1.

[0048] More specifically, if Figure 3 As shown, the first residual branch in the residual layer includes a convolutional layer 8 and a convolutional layer 9 connected in sequence; the second residual branch includes a convolutional layer 7; the convolutional layer in the first residual branch has the same structure as the convolutional layer in the second residual branch, except that the size of the convolution kernel is different; each convolutional layer is followed by a batch normalization layer and an activation layer for normalization, activation and other processing. It is worth noting that the first residual branch may also include more than two convolutional layers, and the second residual branch may include more than one convolutional layer, which are set according to the needs.

[0049] When processing data, in each residual layer, the image feature x is input into the first residual branch and the second residual branch respectively. After the convolution layer 8 and the convolution layer 9 perform spatial downsampling and channel dimension upscaling operations on the input image feature x, the feature is added to the feature output by the convolution layer 7 and passes through the activation layer to obtain the deep feature y. The residual structure designed in this embodiment enables the feature to be transferred to a deeper structure, thereby performing deeper feature extraction to enhance the feature acquisition capability of the entire model.

[0050] In convolutional layers 2 to 4, each activation layer is followed by a channel attention layer. Figure 4 As shown in Figure 1, the channel attention layer includes a parallel first attention branch and a second attention branch. The structures of the two attention branches are roughly the same, both of which include a pooling layer, a convolution layer, an activation layer, and a convolution layer connected in sequence. The difference is that the first attention branch uses an adaptive average pooling layer, and the second attention branch uses an adaptive maximum pooling layer.

[0051] When processing data, in each channel attention layer, the image features Input the first attention branch and the second attention branch respectively, and obtain image features through the adaptive average pooling layer and the adaptive maximum pooling layer and . Then the image features and Both use two convolutional layers to explore the intrinsic relationship in the channel dimension and obtain the features of the output of the first attention branch. and the features of the output of the second attention branch Finally, and After adding, the activation function is used to input image features Perform channel weighting ( ), and obtain the weighted optimization feature . Its specific expression is:

[0052] ;

[0053] ;

[0054] .

[0055] The channel attention layer can help the model recover more accurate feature mapping, so that the dimension reduction process of the feature will not lose important information for the regression task. At the same time, in the channel attention layer, the maximum pooling can highlight the significant features of the target area. In this way, the global and local characteristics of the ISAR image can be fully captured, thus achieving good performance.

[0056] In the specific implementation process of step 203, based on the constructed space target axial estimation model, the total loss function calculated includes three parts: dynamic perception weighted loss, periodic relative error constraint loss, and multi-task uncertainty loss.

[0057] Specifically, the expression of the total loss function is:

[0058] ;

[0059] In the formula, represents the balance coefficient of the dynamic perceptual weighted loss; The balance factor representing periodic losses; The balance coefficient representing the multi-task uncertainty loss; represents the dynamics-aware weighted loss; represents the periodic relative error constraint loss; represents the multi-task uncertainty loss. Among them, , , They are all hyperparameters obtained from experimental experience.

[0060] More specifically, the calculation expression of the dynamic perception weighted loss is:

[0061] ;

[0062] in,

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] The calculation expression of the periodic relative error constraint loss is:

[0068] ;

[0069] The calculation expression of multi-task uncertainty loss is:

[0070] ;

[0071] In the formula, The balance factor representing the yaw loss; The balance coefficient representing the pitch loss; Indicates yaw loss; Indicates pitch loss; Indicates The predicted value of samples in the yaw task; Indicates The predicted value of samples in the pitch task; Indicates The actual yaw value of samples in the yaw task; Indicates The true pitch value of samples in the pitch task; represents the total number of samples; Indicates The dynamic weight of samples in the yaw task; Indicates The dynamic weight of samples in the pitch regression task; represents the standard deviation of the yaw prediction task; represents the standard deviation of the pitch prediction task. , All of them are hyperparameters, obtained from experimental experience; represents the standard deviation of the yaw prediction task; represents the standard deviation of the pitch prediction task; represents the raw loss of yaw; Indicates the raw loss of pitch.

[0072] It is understandable that weights are usually adjusted according to the complexity, difficulty, or other domain-related information of the sample, ensuring that the model pays more attention to difficult samples. Therefore, in the dynamic perceptual weighted loss, by setting the dynamic weight and , which enables the weights in each channel attention layer to be dynamically adjusted, thereby learning the channel weights that are more important for yaw and pitch angle prediction, enhancing the expressiveness of these channels and reducing interference with irrelevant information, thereby improving the robustness and discriminability of feature representation. In addition, in the dynamic weight and In the and , which can control the extension range of the error. By dynamically perceiving the weighted loss, the weight can be adjusted according to the complexity of the sample, enhancing the model's ability to handle difficult samples and improving the robustness of the overall prediction.

[0073] For the angle prediction task, this embodiment takes into account the periodic characteristics of the space target posture, sets the periodic relative error constraint loss to deal with the errors of the yaw angle and pitch angle, emphasizes the periodic characteristics of the angle difference, and uses the relative amplification or reduction loss of the absolute value of the error to help capture this periodic behavior and limit the error, avoiding the non-periodic prediction of the angle error, so as to better improve the model accuracy.

[0074] For the uncertainty in multi-task learning, this embodiment sets a multi-task uncertainty loss, which also sets the standard deviation and , which is used to reflect the uncertainty of the model in the yaw angle and pitch angle prediction tasks. and The unweighted mean square error is used. It is used to penalize the uncertainty of the prediction. In this embodiment, setting the multi-task uncertainty loss can help the model remain stable when facing uncertain tasks, and avoid excessive uncertainty by penalizing the logarithm of the standard deviation.

[0075] In this embodiment, some common problems of axial estimation in ISAR images are solved through the designed loss function, such as dynamic adjustment of sample weights, processing of periodic errors, and modeling of uncertainty. It not only has advantages in the accuracy of model training, but also can enhance the model's understanding of the correlation between different tasks, and can handle the uncertainty of different tasks, thereby improving the performance of the model in complex tasks, especially when dealing with non-cooperative space targets, with higher accuracy, robustness and adaptability.

[0076] In one of the embodiments, the method proposed in the present invention is verified. First, the data set is divided into a training set, a validation set, and a test set in a ratio of 7:1:2; the optimizer uses the adam optimizer, and the learning rate is 0.001. The total training cycle is set to 200. The average attitude estimation error in the yaw and pitch directions is used as the key indicator for evaluating network performance to accurately evaluate the performance of the network. The dimension of the input data is 512×512×3. The dimension of the output data is 2×1.

[0077] The test environment is an NVIDIA RTX A5000 graphics card equipped with a Pytorch framework. As shown in Table 1, the proposed method shows significant advantages over the traditional model in terms of the average estimation errors of the yaw angle and pitch angle, which are 0.85° and 6.6° respectively.

[0078] In order to further explore the specific reasons for this advantage, this method subdivides the pitch angle range into 12 intervals, with each 30° being a subdivision interval. Figure 5 The estimation error results in each interval are shown in detail. The analysis shows that the estimation error of the yaw angle is relatively large in the three intervals of 90-120° and 330-360°, while relatively good estimation results are obtained in other intervals. The pitch angle estimation error in the 0-30° interval is the most significant, up to 36.9°, followed by the 330-360° interval, and the estimation errors in other intervals are generally less than 5°.

[0079] Table 1 Average estimation errors of yaw and pitch for two models

[0080]

[0081] In addition, the present invention uses Grad-CAM technology to implement inverse activation operations for the predicted values ​​of yaw angle and pitch angle, and successfully obtains heat maps of multiple convolutional layers on the feature extraction area. These heat maps serve as an important basis for the partial interpretability of the network model proposed in the present invention. Specifically, Figure 6 and Figure 7 The heatmaps shown are based on sample data with a yaw angle of 30° and a pitch angle of 150°.

[0082] Through a detailed analysis of these heat maps, it can be clearly observed that in convolutional layer 1, the model mainly performs shallow, global feature extraction, focusing on the overall outline and basic shape of the satellite. As the network level goes deeper, convolutional layers 5 and 6 gradually focus on the key features of the satellite, which are crucial for accurately predicting the yaw and pitch angles.

[0083] It can be understood that although this embodiment Figure 1The steps in the process are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps can be executed in other orders. Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed 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 part of the sub-steps or stages of other steps.

[0084] Example 2

[0085] Based on the space target axial estimation method based on convolutional neural network in Example 1, this embodiment discloses a space target axial estimation device based on convolutional neural network, such as Figure 8 As shown, the spatial target axial estimation device based on convolutional neural network includes: a data acquisition module 401, a data processing module 402, a model training module 403 and a posture estimation module 404, wherein:

[0086] The data acquisition module 401 is used to acquire a data set, which includes ISAR images at different attitude angles and labels corresponding to the ISAR images.

[0087] The data processing module 402 is used to construct an axial estimation model for space targets, which includes a backbone network and a branch network. The ISAR image and the corresponding labels are input into the backbone network as a group of training samples. The backbone network extracts features of the ISAR image by combining the channel attention mechanism with feature inverse mapping to obtain weighted optimized features. There are two branch networks, and the weighted optimized features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target.

[0088] The model training module 403 is used to calculate the loss function based on the yaw angle and the pitch angle, and iteratively optimize the space target axial estimation model through the loss function to obtain a trained space target axial estimation model.

[0089] The pose estimation module 404 is used to estimate the pose of the space target in the to-be-detected SAR image by using the trained space target axial estimation model.

[0090] In this embodiment, the specific working process and working principle of the data acquisition module 401, the data processing module 402, the model training module 403 and the pose estimation module 404 are the same as those in the method of Embodiment 1, so they are not described in detail in this embodiment. Each unit module can be implemented in whole or in part by software, hardware and a combination thereof, and each unit module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above unit modules.

[0091] Example 3

[0092] like Fig. 9 The terminal device disclosed in this embodiment includes a transmitter, a receiver, a memory, and a processor. The transmitter is used to send instructions and data, the receiver is used to receive instructions and data, the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions stored in the memory to implement the method in the above-mentioned embodiment 1.

[0093] It should be noted that the above memory can be independent or integrated with the processor. When the memory is independently provided, the terminal device further includes a bus for connecting the memory and the processor.

[0094] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may 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 many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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.

[0096] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A method for estimating the axial direction of a space target based on a convolutional neural network, characterized in that: The method comprises: Acquire a data set, the data set comprising ISAR images at different attitude angles and labels corresponding to the ISAR images; A space target axial estimation model is constructed, wherein the space target axial estimation model includes a backbone network and a branch network; the ISAR image and the corresponding label are input into the backbone network as a group of training samples, and the backbone network extracts features of the ISAR image by combining a channel attention mechanism with feature inverse mapping to obtain weighted optimization features; there are two branch networks, and the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target; Calculating a loss function based on the yaw angle and the pitch angle, and iteratively optimizing the space target axial estimation model through the loss function to obtain a trained space target axial estimation model; The position and posture of the space target in the SAR image to be detected is estimated by using the trained space target axial estimation model.

2. The method for estimating the spatial target axis based on a convolutional neural network according to claim 1, characterized in that: The backbone network includes a plurality of residual layers and a plurality of convolutional layers connected in sequence; The residual layer extracts features through a low-to-high dimensional channel to obtain deep features; the convolution layer reduces the dimensionality of the deep features from high to low and performs feature inverse mapping; A channel attention layer is also set after each convolution layer; the channel attention layer captures global and local features and finally outputs weighted optimized features.

3. The method for estimating the spatial target axis based on a convolutional neural network according to claim 2, characterized in that: The residual layer includes a first residual branch and a second residual branch in parallel; After the image features are respectively input into the first residual branch and the second residual branch for processing, the features output by the first residual branch and the features output by the second residual branch are added, and after passing through an activation layer, deep features are output.

4. The method for estimating the spatial target axis based on a convolutional neural network according to claim 2, characterized in that: The channel attention layer includes a parallel first attention branch and a second attention branch; After the image features are input into the first attention branch and the second attention branch for processing, the features output by the first attention branch are added to the features output by the second attention branch, and the weighted optimized features are output after passing through the activation layer.

5. The method for estimating the axial direction of a space target based on a convolutional neural network according to any one of claims 1 to 4, characterized in that: The loss function is calculated based on the yaw angle and the pitch angle, and the loss function expression is: ； In the formula, represents the balance coefficient of the dynamic perceptual weighted loss; The balance factor representing periodic losses; The balance coefficient representing the multi-task uncertainty loss; represents the dynamics-aware weighted loss; represents the periodic relative error constraint loss; represents the multi-task uncertainty loss.

6. The method for estimating the spatial target axis based on a convolutional neural network according to claim 5, characterized in that: The calculation expression of the dynamic perception weighted loss is: ； in, ； ； ； ； In the formula, The balance factor representing the yaw loss; The balance coefficient representing the pitch loss; Indicates yaw loss; Indicates pitch loss; Indicates The predicted value of samples in the yaw task; Indicates The predicted value of samples in the pitch task; Indicates The actual yaw value of samples in the yaw task; Indicates The true pitch value of samples in the pitch task; represents the total number of samples; Indicates The dynamic weight of samples in the yaw task; Indicates The dynamic weight of samples in the pitch regression task; represents the standard deviation of the yaw prediction task; represents the standard deviation of the pitch prediction task.

7. The method for estimating the spatial target axis based on a convolutional neural network according to claim 5, characterized in that: The calculation expression of the periodic relative error constraint loss is: ； In the formula, Indicates The predicted value of samples in the yaw task; Indicates The predicted value of samples in the pitch task; Indicates The actual yaw value of samples in the yaw task; Indicates The true pitch value of samples in the pitch task; Represents the total number of samples.

8. The method for estimating the spatial target axis based on a convolutional neural network according to claim 5, characterized in that: The calculation expression of the multi-task uncertainty loss is: ； In the formula, represents the standard deviation of the yaw prediction task; represents the standard deviation of the pitch prediction task; represents the raw loss of yaw; Indicates the raw loss of pitch.

9. A spatial target axial estimation device based on convolutional neural network, characterized in that: The device comprises: A data acquisition module, used to acquire a data set, wherein the data set includes ISAR images at different attitude angles and labels corresponding to the ISAR images; A data processing module is used to construct an axial estimation model for a space target, wherein the axial estimation model for a space target includes a backbone network and a branch network; the ISAR image and the corresponding label are input into the backbone network as a group of training samples, and the backbone network extracts features of the ISAR image by combining a channel attention mechanism with feature inverse mapping to obtain weighted optimization features; there are two branch networks, and the weighted optimization features are respectively input into the two branch networks for processing to obtain the yaw angle and pitch angle of the space target; A model training module, used for calculating a loss function based on the yaw angle and the pitch angle, and iteratively optimizing the space target axial estimation model through the loss function to obtain a trained space target axial estimation model; The pose estimation module is used to estimate the pose of the space target in the SAR image to be detected by using the trained space target axial estimation model.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method for axial estimation of a spatial target based on a convolutional neural network as described in any one of claims 1 to 8 are implemented.

Images