A spacecraft component identification method and system

Abstract

The present application discloses a method and system for identifying spacecraft components, which relates to the field of image recognition technology. The method includes: inputting the image to be identified into a spatial attention perception encoder to generate output features; inputting the output features into a multi-level adaptive spatial feature fusion decoder to obtain a spacecraft component mask; and determining the recognition result of the spacecraft component based on the spacecraft component mask. The present application uses a backbone network ConvNeXt‑B and a spatial attention perception feature fusion module to extract features, and inputs the extracted features into four cascaded adaptive spatial feature fusion modules to fuse low-level, middle-level, and high-level features of different scales, thereby obtaining richer semantic information of spacecraft components. Experimental results show that compared with other existing methods, the method of the present application achieves the best spacecraft component recognition effect.

Description

Technical Field

[0001] The present application relates to the field of image recognition technology, and in particular to a method and system for identifying spacecraft components. Background Art

[0002] With the rapid development of the aerospace industry, on-orbit servicing of space targets has become an important development technology in the aerospace field. On-orbit servicing of space targets includes the repair, reorganization and capture of spacecraft. The identification of spacecraft components is the most critical step in achieving on-orbit servicing of space targets, and is the prerequisite for subsequent satellites to conduct rendezvous and docking, repair and assembly, strike space debris and other tasks. Spacecraft components refer to observable local payloads that play an important role in the on-orbit operation of space targets. These components have obvious graphic features and textures in the images of space targets. In the spacecraft component identification task, five components need to be identified, including solar panels, antennas, optical payloads, instruments and thrusters. These components undertake key functions and tasks. Solar panels provide power, antennas communicate, instruments are responsible for recording and transmitting data, thrusters are responsible for spacecraft movement, and optical payloads perform observations.

[0003] In recent years, there are two main methods for spacecraft component recognition based on deep learning: target detection and semantic segmentation, both of which are key tasks in the field of computer vision.

[0004] Object detection methods classify and locate spacecraft components by using bounding boxes. However, object detection methods can only provide approximate locations within rectangular boundaries, which may not accurately reflect the specific shape of the components. Therefore, it is difficult for object detection-based methods to provide precise edge information for spacecraft components. In contrast, semantic segmentation-based methods are more effective in detecting and locating the entire spacecraft.

[0005] Compared with the target detection method, the semantic segmentation method assigns a category label to each image pixel, thereby generating an accurate component mask, providing richer and more detailed information for subsequent image analysis tasks. In addition, by adopting deep learning techniques such as convolutional neural networks, the semantic segmentation-based method can learn the characteristics of components of different scales and shapes, thereby flexibly adapting to components of different scales and shapes. Therefore, the semantic segmentation method is more suitable for the task of spacecraft component identification, and can achieve accurate segmentation of spacecraft component images, providing better support for subsequent on-orbit services of space targets.

[0006] Although semantic segmentation is more suitable for spacecraft component identification tasks, it also has many limitations. Due to the large differences in size and shape of spacecraft components, misclassification or omission occurs during the spacecraft component identification process, which makes it difficult to perform subsequent spacecraft rendezvous, capture, and repair tasks. Secondly, some spacecraft have complex surface textures and irregular shapes. In the process of identifying spacecraft components, the edges of the generated component segmentation mask are inaccurate and even discontinuous, which hinders the spacecraft from performing subsequent complex tasks such as on-orbit servicing of space targets. Summary of the invention

[0007] The embodiments of the present application provide a method and system for identifying spacecraft components, which are used to solve the problem of inaccurate recognition results in the process of spacecraft component recognition by the semantic segmentation method in the prior art.

[0008] On the one hand, an embodiment of the present application provides a method for identifying a spacecraft component, comprising:

[0009] Obtain an image to be recognized;

[0010] The image to be recognized is input into the spatial attention perception encoder, which includes a backbone network ConvNeXt-B and a first spatial attention perception feature fusion module. The backbone network ConvNeXt-B includes four stages connected in sequence, namely stage one, stage two, stage three and stage four. Two first spatial attention perception feature fusion modules are connected after stage two and stage three respectively. Stage one extracts the features of the image to be recognized, and stages two, three and four extract features from the features output by the previous stage or the first spatial attention perception feature fusion module. Each stage generates corresponding output features.

[0011] The output features are input into a multi-stage adaptive spatial feature fusion decoder. The multi-stage adaptive spatial feature fusion decoder includes a multi-stage adaptive spatial feature fusion module and a second spatial attention perception feature fusion module connected in sequence. The multi-stage adaptive spatial feature fusion module includes four cascaded adaptive spatial feature fusion modules. The output features generated by the first stage, the second stage, the third stage and the fourth stage are input into an adaptive spatial feature fusion module respectively. The adaptive spatial feature fusion module fuses the output features of the corresponding stage and the output features of the previous stage to generate fused features. The second spatial attention perception feature fusion module splices and cascades the fused features to obtain a spacecraft component mask.

[0012] An identification result of the spacecraft component is determined based on the spacecraft component mask.

[0013] On the other hand, an embodiment of the present application further provides a spacecraft component identification system, including:

[0014] An image acquisition module, used for acquiring an image to be identified;

[0015] A spatial attention perception encoder comprises a backbone network ConvNeXt-B and a first spatial attention perception feature fusion module. The backbone network ConvNeXt-B comprises four stages connected in sequence, namely stage one, stage two, stage three and stage four. Two first spatial attention perception feature fusion modules are connected after stage two and stage three respectively. Stage one extracts features of the image to be recognized. Stages two, three and four extract features from the features output by the previous stage or the first spatial attention perception feature fusion module. Each stage generates corresponding output features.

[0016] A multi-stage adaptive spatial feature fusion decoder comprises a multi-stage adaptive spatial feature fusion module and a second spatial attention perception feature fusion module connected in sequence, wherein the multi-stage adaptive spatial feature fusion module comprises four cascaded adaptive spatial feature fusion modules, wherein the output features generated by the first stage, the second stage, the third stage and the fourth stage are input into one adaptive spatial feature fusion module correspondingly, and the adaptive spatial feature fusion module fuses the output features of the corresponding stage and the output features of the previous stage to generate fused features, and the second spatial attention perception feature fusion module splices and cascades the fused features to obtain a spacecraft component mask;

[0017] An image recognition module is used to determine a recognition result of a spacecraft component based on a spacecraft component mask.

[0018] A spacecraft component identification method and system in this application has the following advantages:

[0019] The backbone network ConvNeXt-B and the Spatial Attention Feature Fusion Module (SA-FFM) are used to extract features, and the extracted features are input into four cascaded Adaptive Spatial Feature Fusion (ASFF) modules to fuse low-level, mid-level, and high-level features of different scales, thereby obtaining richer semantic information of spacecraft components. Experimental results show that compared with other existing methods, the method of this application achieves the best spacecraft component recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application 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 application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 An architectural diagram of a spacecraft component identification system provided in an embodiment of the present application.

[0022] Figure 2 This is an architecture diagram of the backbone network ConvNeXt-B provided in an embodiment of the present application.

[0023] Figure 3 This is an architectural diagram of the residual block ConvNeXt Block provided in an embodiment of the present application.

[0024] Figure 4 An architectural diagram of a downsampling block provided in an embodiment of the present application.

[0025] Figure 5 An architectural diagram of the spatial attention perception feature fusion module provided in an embodiment of the present application.

[0026] Figure 6 An architectural diagram of the spatial attention mechanism block provided for an embodiment of the present application.

[0027] Figure 7 An architectural diagram of the multi-level adaptive spatial feature fusion module provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0029] Figure 1 This is an architecture diagram of a spacecraft component identification system provided in an embodiment of the present application. This embodiment of the present application provides a spacecraft component identification method, including the following steps:

[0030] S100: Acquire an image to be recognized.

[0031] S110, input the image to be recognized into the Spatial Attention Encoder (SAE), the Spatial Attention Encoder includes a backbone network ConvNeXt-B and a first spatial attention perception feature fusion module, the backbone network ConvNeXt-B includes four stages connected in sequence, namely stage 1, stage 2, stage 3 and stage 4, and the two first spatial attention perception feature fusion modules are connected after stage 2 and stage 3 respectively, such as Figure 2As shown in the figure, stage one extracts the features of the image to be recognized, and stages two, three, and four extract features from the features output by the previous stage or the first spatial attention perception feature fusion module, and each stage generates corresponding output features.

[0032] For example, Figure 3 As shown, stage 1 contains a residual block ConvNeXt Block, and stage 2, stage 3, and stage 4 all contain a downsampling block and a residual block ConvNeXt Block connected in sequence. In the backbone network ConvNeXt-B, the number of repetitions of stage 1, stage 2, stage 3, and stage 4 are 3, 3, 27, and 3, respectively.

[0033] The residual blocks ConvNeXt Block in stage 1, stage 2, stage 3 and stage 4 perform depth 2D convolution with convolution kernel 7×7, layer normalization (LN), 2D convolution with convolution kernel 1×1, layer normalization, 2D convolution with convolution kernel 1×1, layer normalization and regularization in sequence, and the downsampling block performs layer normalization and 2D convolution with convolution kernel 2×2 in sequence, as shown in Fig. Figure 4 shown.

[0034] The first spatial attention perception feature fusion module contains three parallel branches, which respectively perform spatial feature enhancement, local feature extraction and retain original input features. The first spatial attention perception feature fusion module concatenates local features, original input features and enhanced spatial features.

[0035] Specifically, Figure 5 As shown, among the three branches in the first spatial attention perception feature fusion module, the first branch processes the input feature F After performing a 1×1 two-dimensional convolution with a convolution kernel, the first convolution feature is obtained. , the first convolution feature After the two-dimensional convolution with a convolution kernel of 3×3, the second convolution feature is obtained. The first convolution feature and the second convolution feature are combined to obtain the first combined convolution feature. The first combined convolution feature is subjected to a two-dimensional convolution with a convolution kernel of 3×3 to obtain the third convolution feature. The third convolution feature and the first combined convolution feature are combined to obtain the second combined convolution feature. , the second merged convolutional features After being processed by the spatial attention mechanism block, the spatial features are obtained , the spatial feature and the second merged convolution feature are convolved to obtain the fourth convolution feature, completing the enhancement of the spatial feature.

[0036] The second branch performs a two-dimensional convolution with a convolution kernel of 1×1 on the input feature to obtain the fifth convolution feature. The fifth convolution feature is combined with the fourth convolution feature to obtain the third combined convolution feature. The third combined convolution feature is subjected to a two-dimensional convolution with a convolution kernel of 3×3 to obtain the sixth convolution feature, thereby completing the extraction of local features.

[0037] The third branch combines the input features with the sixth convolutional features to obtain the output features. express.

[0038] like Figure 6 As shown in the figure, the spatial attention mechanism block performs maximum pooling and average pooling on the input features to obtain feature one and feature two respectively. The features obtained by merging feature one and feature two are convolved to obtain feature three. Feature three is activated by the Sigmoid function to obtain feature four, which is the output feature.

[0039] S120, input the output features into a multi-scale adaptive spatial feature fusion decoder (Multi-scale AdaptiveSpatial Feature Fusion Decoder, Multi-ASFFD), the multi-scale adaptive spatial feature fusion decoder includes a multi-scale adaptive spatial feature fusion module (Multi-scale Adaptive Spatial Feature Fusion Module, M-ASFFM) and a second spatial attention perception feature fusion module connected in sequence, the multi-level adaptive spatial feature fusion module includes four cascaded adaptive spatial feature fusion modules (ASFFM), the output features generated by stage one, stage two, stage three and stage four are input into an adaptive spatial feature fusion module corresponding to each other, the adaptive spatial feature fusion module fuses the output features of the corresponding stage and the output features of the previous stage to generate fused features, the second spatial attention perception feature fusion module splices and cascades the fused features to obtain a spacecraft component mask.

[0040] For example, Figure 7 As shown, the process of the adaptive spatial feature fusion module consists of two stages: constant scaling and adaptive fusion.

[0041] In the constant scaling stage, each adaptive spatial feature fusion module samples and performs convolution operations on the input features, so that the number of channels of the feature vectors output by other adaptive spatial feature fusion modules is adjusted to the same as the number of channels of the feature vectors output by the current adaptive spatial feature fusion module.

[0042] Specifically, during the constant scaling phase:

[0043] For the first-level ASFFM, the feature maps output by stages 2, 3, and 4 are downsampled and convolved to obtain the same number of channels as stage 1.

[0044] For the second-level ASFFM, the feature map output from stage one is upsampled and convolved to obtain the same number of channels as stage two; the feature maps output from stages three and four are downsampled and convolved to obtain the same number of channels as stage two.

[0045] For the third-level ASFFM, the feature maps output by stage one and stage two are upsampled and convolved to obtain the same number of channels as stage three; the feature maps output by stage four are downsampled and convolved to obtain the same number of channels as stage three.

[0046] For the fourth level ASFFM, the feature maps output by stages one, two, and three are upsampled and convolved to obtain the same number of channels as stage four.

[0047] In the adaptive fusion stage:

[0048] Adaptive fusion is performed according to the formula:

[0049]

[0050] In the formula, is the fused feature, Respectively, it indicates the adjustment from stage 1, stage 2, stage 3 and stage 4 to stage l The eigenvectors of , , and The sum is , . , , and is the weight coefficient of the feature map output by the four stages after softmax processing, and . The calculation formula is:

[0051]

[0052] In the formula, e is a natural constant, , , and Respectively represent l The adaptive fusion mechanism in the layer is used as the feature vector , , and The assigned weight coefficients are processed through a softmax function to ensure that their sum is 1.

[0053] The second spatial attention perception feature fusion module also includes three parallel branches, which respectively perform spatial feature enhancement, local feature extraction, and retain original input features. The second spatial attention perception feature fusion module concatenates local features, original input features, and enhanced spatial features. The processing flow of the three branches is the same as that of the three branches in the first spatial attention perception feature fusion module, and will not be repeated here.

[0054] The spatial attention-aware encoder and the multi-level adaptive spatial feature fusion decoder are trained before processing the image to be recognized. During the training process, a composite loss function consisting of a cross entropy loss function and a boundary loss function is used for optimization.

[0055] Specifically, the composite loss function can be written as:

[0056]

[0057] In the formula, is the composite loss function, is the cross entropy loss function, is the boundary loss function.

[0058] The cross entropy loss function is a pixel-based loss function that calculates the loss during training by comparing the probability distribution of each pixel output by the network with the similarity between the probability distribution of the pixel of the true label. The formula is as follows:

[0059]

[0060] In the formula, N Represents the total number of pixels in the image. M represents the number of categories, is the observed sample i Belongs to category c The predicted probability of is a sign function, if the sample i The true category is c ,but ,otherwise .

[0061] The boundary loss function is mainly used to improve the accuracy of the network on the boundary of the segmented target. By using the integral method to calculate the change of the boundary, the local differential calculation on the boundary point is avoided. The formula is as follows:

[0062]

[0063] In the formula, is the boundary loss function, which is defined by the domain of the image The boundary integral on , Represents the domain of the image, i.e., the space containing all pixels. The integral of the boundary loss is calculated within this domain. It's the border The level set function is directly based on the ground truth area G Calculated in advance, is the groundtruth region G The set of all points on the boundary. q ,but ,otherwise . Represents any point q With border The distance between the closest points on . It is a region S The binary indicator function of: If q ,but =1, otherwise =0.

[0064] The image to be recognized is processed by two-dimensional convolution and layer normalization in turn before being used as the input of stage one. The output features of stage four are processed by global average pooling, layer normalization and linear transformation in turn before being input into the adaptive spatial feature fusion module.

[0065] The output features generated in each stage are sequentially batch normalized and activated by Gaussian error linear units before being input into the corresponding adaptive spatial feature fusion module;

[0066] The concatenated features output by the second spatial attention perception feature fusion module are sequentially activated by two-dimensional convolution, batch normalization, and Gaussian error linear unit to obtain the spacecraft component mask.

[0067] S130, determining a recognition result of the spacecraft component based on the spacecraft component mask.

[0068] The present application also provides a spacecraft component identification system, including:

[0069] An image acquisition module, used for acquiring an image to be identified;

[0070] A spatial attention perception encoder comprises a backbone network ConvNeXt-B and a first spatial attention perception feature fusion module. The backbone network ConvNeXt-B comprises four stages connected in sequence, namely stage one, stage two, stage three and stage four. Two first spatial attention perception feature fusion modules are connected after stage two and stage three respectively. Stage one extracts features of the image to be recognized. Stages two, three and four extract features from the features output by the previous stage or the first spatial attention perception feature fusion module. Each stage generates corresponding output features.

[0071] A multi-stage adaptive spatial feature fusion decoder comprises a multi-stage adaptive spatial feature fusion module and a second spatial attention perception feature fusion module connected in sequence, wherein the multi-stage adaptive spatial feature fusion module comprises four cascaded adaptive spatial feature fusion modules, wherein the output features generated by the first stage, the second stage, the third stage and the fourth stage are input into one adaptive spatial feature fusion module correspondingly, and the adaptive spatial feature fusion module fuses the output features of the corresponding stage and the output features of the previous stage to generate fused features, and the second spatial attention perception feature fusion module splices and cascades the fused features to obtain a spacecraft component mask;

[0072] An image recognition module is used to determine a recognition result of a spacecraft component based on a spacecraft component mask.

[0073] In order to accurately evaluate the segmentation performance of the method proposed in this application, the mean intersection over union (mIoU) is selected as the evaluation indicator. This indicator is used to quantitatively analyze the test set. The mIoU value range is [0, 1]. The higher the mIoU, the more accurate the segmentation result of the system. The calculation formula of mIoU is as follows:

[0074]

[0075] in, k Indicates the number of categories, v represents the true value of the pixel category, w Represents the prediction result of the pixel. represents the number of correctly classified pixels, express w Predicted to be v The number of pixels, express v Predicted to be w The number of pixels.

[0076] Simulation experiment

[0077] Simulation conditions:

[0078] A single NVIDIA GeForce GTX 3080Ti with 12G memory is used for training and testing.

[0079] Simulation content:

[0080] In order to prove the effectiveness of the method proposed in this application, seven popular methods in recent years were used in the experiment for comparison. The quantitative detection accuracy mIoU value and IoU value of each component of the method proposed in this application and other 7 comparison algorithms on two data sets are shown in Table 1 and Table 2. Among them:

[0081] 1. U-Net is described in detail in “Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), Munich, Germany, 5–9 October 2015.”

[0082] 2. DANet (Dual Attention Network) is described in “Fu, J.; Liu, J.; Tian, ​​H.; Li, Y.; Bao, Y.; Fang, Z.; Lu, H. Dual Attention Network for Scene Segmentation. In Proceedings of the 2019 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 15–20 June 2020.”

[0083] 3. UperNet (ResNet-50) is introduced in detail in “Wang R, Jiang H, Li Y. UPerNet with ConvNeXtfor Semantic Segmentation[C] / / 2023 IEEE 3rd International Conference on Electronic Technology, Communication and Information (ICETCI). IEEE, 2023.”

[0084] 4. DeepLabv3+ "Chen, LCE; Zhu, YK; Papandreou, G.; Schroff, F.; Adam, H. Encoder-Decoder with Atrous Separable Convolution for Semantic ImageSegmentation. In Proceedings of the 15th European Conference on ComputerVision (ECCV), Munich, Germany, 8–14September 2018." Detailed introduction.

[0085] 5. OCRNet (Object-Contextual Representations Network) is introduced in detail in "Yu, HY, Xiao, KC, Xi, LC, Jing, DW, 2020. SegmentationTransformer: Object-Contextual Representations for Semantic Segmentation. In: Proc. European Conference on Computer Vision(ECCV).".

[0086] 6. SegNeXt is introduced in detail in "Guo MH, Lu CZ, Hou Q, et al. Segnext: Rethinking convolutional attention design for semantic segmentation[J]. Advances in Neural Information Processing Systems, 2022."

[0087] 7. ConvNeXt-B is introduced in detail in “Zhao Y, Zhong R, Cui L. Intelligent recognition of spacecraft components from photorealistic images based on Unreal Engine 4[J]. Advances in Space Research, 2023, 71(9).”

[0088] Table 1 mIoU results of different methods on dataset 1

[0089]

[0090] Table 2 mIoU results of different methods on dataset 2

[0091]

[0092] It can be seen from Table 1 and Table 2 that the effectiveness of the proposed method has been verified on two spacecraft component datasets. The experimental results show that compared with other existing methods, the proposed method achieves the best spacecraft component recognition results.

[0093] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0094] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for identifying a spacecraft component, characterized in that: include: Obtain an image to be recognized; Input the image to be identified into a spatial attention perception encoder, the spatial attention perception encoder includes a backbone network ConvNeXt-B and a first spatial attention perception feature fusion module, the backbone network ConvNeXt-B includes four stages connected in sequence, namely stage one, stage two, stage three and stage four, two of the first spatial attention perception feature fusion modules are connected to stage two and stage three respectively, the stage one extracts the features of the image to be identified, the stage two, the stage three and the stage four extract features from the features output by the previous stage or the first spatial attention perception feature fusion module, and each stage generates corresponding output features; the stage one includes a residual block ConvNeXt Block, the stage two, the stage three and the stage four all include a downsampling block and the residual block ConvNeXt Block connected in sequence; the first spatial attention perception feature fusion module includes three parallel branches, the three branches respectively perform spatial feature enhancement, local feature extraction and retain the original input features, the first spatial attention perception feature fusion module splices and cascades the local features, the original input features and the enhanced spatial features; The output features are input into a multi-stage adaptive spatial feature fusion decoder, the multi-stage adaptive spatial feature fusion decoder comprises a multi-stage adaptive spatial feature fusion module and a second spatial attention perception feature fusion module connected in sequence, the multi-stage adaptive spatial feature fusion module comprises four cascaded adaptive spatial feature fusion modules, the output features generated by the stage one, the stage two, the stage three and the stage four are input into one adaptive spatial feature fusion module respectively, the adaptive spatial feature fusion module fuses the output features of the corresponding stage and the output features of the previous stage to generate fused features, the second spatial attention perception feature fusion module splices and cascades the fused features to obtain a spacecraft component mask; the second spatial attention perception feature fusion module also comprises three parallel branches, the three branches respectively perform spatial feature enhancement, local feature extraction and retain the original input features, the second spatial attention perception feature fusion module splices and cascades the local features, the original input features and the enhanced spatial features; An identification result of a spacecraft component is determined based on the spacecraft component mask.

2. A method for identifying a spacecraft component according to claim 1, characterized in that: Each of the adaptive spatial feature fusion modules performs sampling and convolution operations on the input features, so that the number of channels of the feature vectors output by other adaptive spatial feature fusion modules is adjusted to be the same as the number of channels of the feature vectors output by the current adaptive spatial feature fusion module.

3. A method for identifying a spacecraft component according to claim 1, characterized in that: The residual block ConvNeXtBlock performs depth two-dimensional convolution, layer normalization, two-dimensional convolution, layer normalization, two-dimensional convolution, layer normalization and regularization in sequence, and the downsampling block performs layer normalization and two-dimensional convolution in sequence.

4. A method for identifying a spacecraft component according to claim 1, characterized in that: Among the three branches in the first spatial attention perception feature fusion module and the second spatial attention perception feature fusion module, the first branch performs two-dimensional convolution on the input feature to obtain the first convolution feature, the first convolution feature is again subjected to two-dimensional convolution to obtain the second convolution feature, the first convolution feature and the second convolution feature are merged to obtain the first merged convolution feature, the first merged convolution feature is again subjected to two-dimensional convolution to obtain the third convolution feature, the third convolution feature and the first merged convolution feature are merged to obtain the second merged convolution feature, the second merged convolution feature is processed by the spatial attention mechanism block to obtain the spatial feature, the spatial feature and the second merged convolution feature are convolved to obtain the fourth convolution feature, the second branch performs two-dimensional convolution on the input feature to obtain the fifth convolution feature, the fifth convolution feature and the fourth convolution feature are merged to obtain the third merged convolution feature, the third merged convolution feature is subjected to two-dimensional convolution to obtain the sixth convolution feature, and the third branch merges the input feature with the sixth convolution feature to obtain the output feature.

5. A method for identifying a spacecraft component according to claim 4, characterized in that: The spatial attention mechanism block performs maximum pooling and average pooling on the input features respectively to obtain feature one and feature two. The feature obtained by combining feature one and feature two is subjected to convolution processing to obtain feature three. Feature three is activated by a Sigmoid function to obtain feature four, i.e., the output feature.

6. A method for identifying a spacecraft component according to claim 1, characterized in that: Before processing the image to be recognized, the spatial attention perception encoder and the multi-level adaptive spatial feature fusion decoder are first trained, and a composite loss function composed of a cross entropy loss function and a boundary loss function is used for optimization during the training process.

7. A method for identifying a spacecraft component according to claim 1, characterized in that: The image to be identified is processed by two-dimensional convolution and layer normalization in sequence and then used as the input of the stage one. The output features of the stage four are processed by global average pooling, layer normalization and linear transformation in sequence and then input into the adaptive spatial feature fusion module. The output features generated in each stage are sequentially batch normalized and activated by Gaussian error linear units and then input into the corresponding adaptive spatial feature fusion module; The concatenated and cascaded features output by the second spatial attention perception feature fusion module are sequentially activated by two-dimensional convolution, batch normalization, and Gaussian error linear unit to obtain the spacecraft component mask.

8. A system using a method for identifying a spacecraft component according to any one of claims 1 to 7, characterized in that: include: An image acquisition module, used for acquiring an image to be identified; A spatial attention perception encoder comprises a backbone network ConvNeXt-B and a first spatial attention perception feature fusion module, wherein the backbone network ConvNeXt-B comprises four stages connected in sequence, namely stage one, stage two, stage three and stage four, wherein two of the first spatial attention perception feature fusion modules are respectively connected after the stage two and the stage three, wherein the stage one extracts the features of the image to be identified, and the stage two, the stage three and the stage four extract features of the features output by the previous stage or the first spatial attention perception feature fusion module, and each stage generates corresponding output features; the stage one comprises a residual block ConvNeXt Block, and the stage two, the stage three and the stage four all comprise a downsampling block and the residual block ConvNeXt Block connected in sequence; the first spatial attention perception feature fusion module comprises three parallel branches, wherein the three branches respectively perform spatial feature enhancement, local feature extraction and retain original input features, and the first spatial attention perception feature fusion module concatenates and cascades local features, original input features and enhanced spatial features; A multi-stage adaptive spatial feature fusion decoder comprises a multi-stage adaptive spatial feature fusion module and a second spatial attention perception feature fusion module connected in sequence, wherein the multi-stage adaptive spatial feature fusion module comprises four cascaded adaptive spatial feature fusion modules, wherein the output features generated by the first stage, the second stage, the third stage and the fourth stage are input into one adaptive spatial feature fusion module correspondingly, wherein the adaptive spatial feature fusion module fuses the output features of the corresponding stage and the output features of the previous stage to generate fused features, wherein the second spatial attention perception feature fusion module performs splicing and cascading on the fused features to obtain a spacecraft component mask; wherein the second spatial attention perception feature fusion module also comprises three parallel branches, wherein the three branches respectively perform spatial feature enhancement, local feature extraction and retain original input features, and wherein the second spatial attention perception feature fusion module performs splicing and cascading on the local features, the original input features and the enhanced spatial features; An image recognition module is used to determine a recognition result of a spacecraft component based on the spacecraft component mask.

Images