Adaptive channel screening and feature fusion network for spacecraft component recognition and spacecraft component recognition method

Abstract

The application provides an adaptive channel screening and feature fusion network for spacecraft component identification and a spacecraft component identification method, and is used for solving the technical problem that the existing spacecraft component identification method based on a deep learning model has low identification precision for small components in a complex scene where components of different sizes coexist.The adaptive channel screening and feature fusion network for spacecraft component identification comprises an encoder including a backbone network and a channel screening module; the channel screening module relies on a DCT algorithm and an RPCA algorithm, accurately evaluates the response performance of each channel in an important frequency band, quantifies and determines the information importance of each channel, and then adaptively screens out key feature channels suitable for large components and small components, effectively eliminates redundant feature interference, significantly enhances the identification precision and scene adaptation capability of the adaptive channel screening and feature fusion network for components of various sizes, and lays an efficient and accurate foundation for subsequent feature extraction.

Description

Technical Field

[0001] This invention relates to deep learning models and methods for spacecraft component identification, specifically to an adaptive channel filtering and feature fusion network and method for spacecraft component identification. Background Technology

[0002] With the rapid development of aerospace technology, on-orbit servicing technology for spacecraft has become one of the fastest-growing cutting-edge technologies in the space field. Spacecraft components are localized payloads with clearly defined physical boundaries, capable of fulfilling specific sub-functions of a space mission. They play a crucial role in the on-orbit operation of spacecraft, and the precise positioning and identification of these components is a vital aspect of on-orbit servicing technology. Key spacecraft components include solar panels, antennas, optical payloads, instruments, and thruster nozzles. Accurate identification, classification, and labeling of these components ensures timely and accurate monitoring and full lifecycle management, a crucial prerequisite for the successful implementation of subsequent core missions such as on-orbit inspection, on-orbit docking, and fault repair. This technology holds significant engineering application value and broad development prospects in the aerospace engineering field.

[0003] In recent years, deep learning technology has been widely applied and achieved remarkable results in the field of target detection and recognition due to its powerful feature extraction and fitting capabilities. Compared with traditional recognition methods, deep learning technology has higher recognition accuracy. In complex space scenarios, it not only demonstrates better recognition performance for various targets, but also exhibits stronger robustness to typical interference factors such as low-light scenes, target incompleteness, and motion blur. Therefore, it is widely used in space target detection and recognition tasks. However, existing deep learning-based spacecraft component recognition methods still face many technical challenges that urgently need to be overcome: On the one hand, spacecraft components vary significantly in size. When components of different sizes coexist, deep learning models tend to focus more on feature extraction from large components, while the weak features of small components are easily overwhelmed by redundant information in the shallow stages of the deep learning model; on the other hand, during the fusion of deep and shallow features, the effective features of small-sized components contained in the shallow features are further diluted. The combination of these problems ultimately leads to the low recognition accuracy of existing deep learning-based spacecraft component recognition methods for small components in complex scenarios where components of different sizes coexist. Therefore, how to accurately identify small components in complex scenarios where components of different sizes coexist has become a key technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to address the technical problem that existing deep learning-based spacecraft component identification methods have low accuracy in identifying small components in complex scenarios where components of different sizes coexist. The invention provides an adaptive channel filtering and feature fusion network and a spacecraft component identification method for spacecraft component identification.

[0005] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0006] An adaptive channel filtering and feature fusion network for spacecraft component identification includes an encoder and a decoder. The encoder includes a backbone network and a channel filtering module. The backbone network includes four sequentially arranged stages and three downsampling layers. Each downsampling layer is connected between two adjacent stages to extract features from the input spacecraft component image four times in sequence, resulting in four feature maps of different sizes.

[0007] The channel filtering module includes a DCT transform unit, an RPCA decomposition filtering unit, a low-rank weight generation unit, a sparse weight generation unit, a low-rank feature output unit, and a sparse feature output unit. The input of the DCT transform unit is connected to the output of the first stage, and is used to perform channel-by-channel discrete DCT transform and expansion on the first feature map to obtain a matrix-form response signal. The input of the RPCA decomposition filtering unit is connected to the output of the DCT transform unit, and is used to perform singular value decomposition on the matrix-form response signal to obtain a low-rank matrix and a sparse matrix. Then, the low-rank matrix and the sparse matrix are filtered to obtain low-rank frequency band response features and sparse frequency band responses. Features: The inputs of the low-rank weight generation unit and the sparse weight generation unit are respectively connected to the two outputs of the RPCA decomposition and filtering unit, and are used to perform weight generation processing on the low-rank frequency band response features and the sparse frequency band response features to obtain the low-rank channel weights and the sparse channel weights; The input of the low-rank feature output unit is connected to the output of the first stage and the output of the low-rank weight generation unit, and the input of the sparse feature output unit is connected to the output of the first stage and the output of the sparse weight generation unit, and is used to multiply the first feature map with the low-rank channel weights and the sparse channel weights element-wise to obtain the low-rank feature map and the sparse feature map;

[0008] The decoder includes four feature fusion modules, a feature concatenation layer, a first convolutional layer, and a segmentation head. The input of the first feature fusion module is connected to the outputs of the third and fourth stages, respectively, and is used to fuse the third and fourth feature maps to obtain a first feature fusion map. The input of the second feature fusion module is connected to the outputs of the first and second stages, respectively, and is used to fuse the first and second feature maps to obtain a second feature fusion map. The input of the third feature fusion module is connected to the outputs of the second feature fusion module and the low-rank feature output unit, respectively, and is used to fuse the second feature map and the low-rank feature map to obtain a third feature fusion map. The first feature fusion map consists of two layers: a feature fusion layer and a feature stitching layer. The first feature fusion layer is connected to the output of the third feature fusion module and the output of the sparse feature output unit, respectively. It is used to fuse the third feature fusion map and the sparse feature map to obtain the fourth feature fusion map. The input of the feature stitching layer is connected to the output of the four feature fusion modules, respectively. It is used to stitch the first to fourth feature fusion maps to obtain multi-scale fusion features. The input of the first convolutional layer is connected to the output of the feature stitching layer, and it is used to perform convolution processing on the obtained multi-scale fusion features to obtain refined multi-scale fusion features. The input of the segmentation head is connected to the output of the first convolutional layer, and it is used to generate predictive mask labels based on the refined multi-scale fusion features to achieve the identification of spacecraft components.

[0009] Furthermore, the feature fusion module includes a feature fusion preprocessing module, a multi-priority extraction module, a hybrid expert feature extraction module, a block gating module, and a weighted fusion module;

[0010] In the first feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the fourth stage and the output of the third stage, respectively. This module normalizes the fourth and third feature maps respectively, and then concatenates the two normalized feature maps to obtain the first concatenated feature map. In the second feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the first feature fusion module and the output of the second stage, respectively. This module normalizes the first fused feature map and the second feature map respectively, and then concatenates the two normalized feature maps to obtain the second concatenated feature map. The third feature fusion... In the feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the second feature fusion module and the output of the low-rank feature output unit, respectively. It is used to normalize the second feature fusion map and the low-rank feature map, and then concatenate the two normalized feature maps to obtain the third concatenated feature map. In the fourth feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the third feature fusion module and the output of the sparse feature output unit, respectively. It is used to normalize the third feature fusion map and the sparse feature map, respectively. Then, it concatenates the two normalized feature maps to obtain the fourth concatenated feature map.

[0011] The input of the multi-priority extraction module is connected to the first output of the feature fusion preprocessing module in the same feature fusion module. It is used to extract priors from the corresponding spliced ​​feature map to obtain the standard deviation feature map and the extreme value difference feature map. Then, the spliced ​​feature map, the standard deviation feature map and the extreme value difference feature map are spliced ​​together in the channel dimension to obtain a multi-dimensional spliced ​​feature map.

[0012] The hybrid expert feature extraction module includes a global expert module, a local expert module, and an edge expert module. The inputs of the global expert module, the local expert module, and the edge expert module are respectively connected to the first output of the feature fusion preprocessing module in the same feature fusion module. They are used to perform global expert feature extraction, local expert feature extraction, and edge expert feature extraction on the spliced ​​feature map respectively, so as to obtain a global expert feature map, a local expert feature map, and an edge expert feature map respectively.

[0013] The input of the block gating module is connected to the output of the multi-prior extraction module in the same feature fusion module. It is used to obtain a multi-dimensional feature weight map based on block weights based on the multi-dimensional spliced ​​feature map, and guide the global expert feature map, local expert feature map, and edge expert feature map to perform weighted summation to obtain the fusion logic feature map.

[0014] The input of the weighted fusion module is connected to the output of the block gating module and the second and third outputs of the feature fusion preprocessing module in the same feature fusion module, respectively. It is used to guide the two normalized feature maps to perform weighted summation through the fusion logic feature map to obtain the corresponding feature fusion map. The output of the weighted fusion module in the four feature fusion modules is connected to the input of the feature splicing layer.

[0015] Furthermore, the feature fusion preprocessing modules in the first to third feature fusion modules all include an interpolation module, a first normalization layer, a second normalization layer, and a first splicing module, and the fourth feature fusion module includes a first normalization layer, a second normalization layer, and a first splicing module.

[0016] In the first feature fusion module, the input of the interpolation module is connected to the output of the fourth stage, and the input of the first normalization layer is connected to the output of the interpolation module in the same feature fusion module. This is used to increase the size of the fourth feature map to be the same as the third feature map through interpolation and then perform normalization processing. The input of the second normalization layer is connected to the output of the third stage and is used to normalize the third feature map. The two inputs of the first stitching module are connected to the outputs of the first and second normalization layers, respectively. This is used to stitch the normalized fourth and third feature maps together to obtain the first stitched feature map. In the second feature fusion module, the input of the interpolation module is connected to the output of the first feature fusion module, and the input of the first normalization layer is connected to the output of the interpolation module within the same feature fusion module. This is used to increase the size of the first feature fusion map to be the same as the second feature map through interpolation and then perform normalization processing. The input of the second normalization layer is connected to the output of the second stage and is used to normalize the second feature map. The two inputs of the first stitching module are respectively connected to the outputs of the first and second normalization layers within the same feature fusion module, and are used to process the normalized first feature fusion map and the second feature map. Row concatenation yields a second concatenated feature map. In the third feature fusion module, the input of the interpolation module is connected to the output of the second feature fusion module, and the input of the first normalization layer is connected to the output of the interpolation module within the same feature fusion module. This is used to increase the size of the second fused feature map to be the same as the low-rank feature map through interpolation and to perform normalization processing. The input of the second normalization layer is connected to the first output of the channel filtering module and is used to normalize the low-rank feature map. The two inputs of the first concatenation module are respectively connected to the outputs of the first and second normalization layers within the same feature fusion module and are used to normalize the low-rank feature map. The processed second feature fusion map and low-rank feature map are concatenated to obtain the third concatenated feature map. In the fourth feature fusion module, the input of the first normalization layer is connected to the output of the third feature fusion module to normalize the third feature fusion map. The input of the second normalization layer is connected to the second output of the channel filtering module to normalize the sparse feature map. The two inputs of the first concatenation module are respectively connected to the outputs of the first and second normalization layers in the same feature fusion module to concatenate the normalized third feature fusion map and the sparse feature map to obtain the fourth concatenated feature map.

[0017] The multi-prior extraction module includes a multi-prior extractor and a second stitching module. The input of the multi-prior extractor is connected to the output of the first stitching module in the same feature fusion module, and is used to perform prior extraction on the received stitched feature map to obtain a standard deviation feature map and an extreme value difference feature map. The first input of the second stitching module is connected to the output of the first stitching module in the same feature fusion module, and the second input is connected to the output of the multi-prior extractor, and is used to stitch the received stitched feature map, standard deviation feature map, and extreme value difference feature map along the channel dimension to obtain a multi-dimensional stitched feature map.

[0018] The global expert module includes a global average pooling layer, a multi-layer fully connected network, a first multiplier, and a head convolutional layer connected in sequence; the local expert module includes a second convolutional layer, a third normalization layer, a first activation layer, and a third convolutional layer connected in sequence; the edge expert module includes a grouped convolutional module, a fourth convolutional layer, a fourth normalization layer, a second activation layer, and a fifth convolutional layer connected in sequence, wherein the grouped convolutional module includes a horizontal Sobel convolutional kernel, a vertical Sobel convolutional kernel, and a fusion computation module;

[0019] The input of the global average pooling layer is connected to the output of the first concatenation module in the same feature fusion module, and is used to perform adaptive average pooling on the received concatenated feature map to extract the global feature map; the multilayer fully connected network is used to generate channel attention weights based on the extracted global feature map; the other input of the first multiplier is connected to the output of the first concatenation module in the same feature fusion module, and is used to multiply the channel attention weights with the concatenated feature map to obtain the global expert feature map;

[0020] The input of the second convolutional layer is connected to the output of the first stitching module in the same feature fusion module, and is used to convolve the received stitched feature map to obtain a local feature map; the third normalization layer, the first activation layer, and the third convolutional layer are used to perform normalization processing, activation processing, and feature integration on the local feature map in sequence to obtain a local expert feature map;

[0021] The inputs of the horizontal and vertical Sobel convolutional kernels are both connected to the output of the first stitching module in the same feature fusion module. This is used to extract gradients from the received stitched feature maps, obtaining horizontal and vertical gradient feature maps. The input of the fusion calculation module is connected to the outputs of the corresponding horizontal and vertical Sobel convolutional kernels, respectively. This is used to perform square root and squaring operations on the horizontal and vertical gradient feature maps, obtaining a fused gradient feature map. The input of the fourth convolutional layer is connected to the output of the fusion calculation module, used to convolve the fused gradient feature map to obtain edge features. The fourth normalization layer, the second activation layer, and the fifth convolutional layer are used to perform normalization, activation, and feature integration on the edge features sequentially, obtaining an edge expert feature map.

[0022] The block gating module includes a block convolutional layer, a fifth normalization layer, a third activation layer, a sixth convolutional layer, a first softmax layer, a seventh convolutional layer, a bilinear upsampling layer, and a gated weighted fusion module connected in sequence.

[0023] The input of the block convolutional layer is connected to the output of the second stitching module in the same feature fusion module, and is used to perform skip convolution operations on the multidimensional stitched feature map to obtain the initial block feature map; the fifth normalization layer, the third activation layer, and the sixth convolutional layer are used to perform normalization, activation, and feature integration on the initial block feature map in sequence to obtain the block logical value feature map; the first softmax layer is used to perform softmax normalization on the block logical value feature map to obtain the block weights; the seventh convolutional layer is used to perform further convolution in combination with the block weights and to enlarge the size through the bilinear upsampling layer to obtain a multidimensional feature weight map based on the block weights; the input of the gated weighted fusion module is connected to the output of the bilinear upsampling layer, the head convolutional layer, the third convolutional layer, and the fifth convolutional layer in the same feature fusion module, and is used to guide the global expert feature map, the local expert feature map, and the edge expert feature map to perform weighted summation through the multidimensional feature weight map based on the block weights to obtain the fused logical feature map;

[0024] The input of the weighted fusion module is connected to the output of the gated weighted fusion module, the first normalization layer, and the second normalization layer in the same feature fusion module. It is used to guide the feature maps output by the first normalization layer and the second normalization layer to perform weighted summation through the fusion logic feature map to obtain the corresponding feature fusion map.

[0025] Furthermore, when filtering low-rank matrices, the RPCA decomposition and filtering unit retains the 12 low-rank frequency bands with the largest L2 norm, and when filtering sparse matrices, it retains the 24 sparse frequency bands with the largest L1 norm.

[0026] Furthermore, the low-rank weight generation unit includes a first learning parameter configuration module, a first frequency band index loading module, a first frequency band response loading module, a second multiplier, a second softmax layer, a third multiplier, a sixth normalization layer, and a first fully connected network; the sparse weight generation unit includes a second learning parameter configuration module, a second frequency band index loading module, a second frequency band response loading module, a fourth multiplier, a third softmax layer, a fifth multiplier, a seventh normalization layer, and a second fully connected network.

[0027] The inputs of the first learning parameter configuration module, the first frequency band index loading module, and the first frequency band response loading module are respectively connected to the first output of the RPCA decomposition and filtering unit, and are used to configure learnable parameter vectors for the low-rank frequency band response features, load the frequency band indexes of the low-rank frequency band response features, and load the frequency band responses of the low-rank frequency band response features; the input of the second multiplier is respectively connected to the output of the first learning parameter configuration module and the output of the first frequency band index loading module, and is used to perform element-wise multiplication of the frequency band indexes of the low-rank frequency band response features with the corresponding learnable parameter vectors to obtain the low-rank feature frequency band weighted response; the input of the second softmax layer is connected to the second multiplier. The output of the first multiplier is used to perform a softmax function operation on the weighted response of the low-rank feature frequency bands. The input of the third multiplier is connected to the output of the second softmax layer and the output of the first frequency band response loading module, respectively. It is used to multiply the weighted response of the low-rank feature frequency bands after the softmax function operation with the frequency band response of the low-rank frequency band response feature again element-wise to obtain the intermediate value of the low-rank feature weights. The input of the sixth normalization layer is connected to the output of the third multiplier, and its output is connected to the input of the first fully connected network. It is used to normalize the intermediate value of the low-rank feature weights and then input it into the first fully connected network for feature dimension adjustment to obtain the low-rank channel weights.

[0028] The inputs of the second learning parameter configuration module, the second frequency band index loading module, and the second frequency band response loading module are respectively connected to the second output of the RPCA decomposition and filtering unit. These modules are used to configure learnable parameter vectors for the sparse frequency band response features, load the frequency band indexes of the sparse frequency band response features, and load the frequency band responses of the sparse frequency band response features. The input of the fourth multiplier is connected to the outputs of the second learning parameter configuration module and the second frequency band index loading module. This multiplier performs element-wise multiplication of the frequency band indexes of the sparse frequency band response features with the corresponding learnable parameter vectors to obtain the sparse feature frequency band weighted response. The input of the third softmax layer is connected to the fourth multiplier. The output of the fifth multiplier is used to perform a softmax function operation on the sparse feature frequency band weighted response; the input of the fifth multiplier is connected to the output of the third softmax layer and the output of the second frequency band response loading module, respectively, and is used to multiply the sparse feature frequency band weighted response after the softmax function operation with the frequency band response of the sparse frequency band response feature again element-wise to obtain the intermediate value of the sparse feature weight; the input of the seventh normalization layer is connected to the output of the fifth multiplier, and its output is connected to the input of the second fully connected network, and is used to normalize the intermediate value of the sparse feature weight, and then input it into the second fully connected network for feature dimension adjustment to obtain the sparse channel weight;

[0029] The low-rank feature output unit includes a sixth multiplier, and the sparse feature output unit includes a seventh multiplier. The input of the sixth multiplier is connected to the output of the first stage and the output of the first fully connected network, respectively, and is used to multiply the first feature map element-wise with the low-rank channel weights to obtain the low-rank feature map. The input of the seventh multiplier is connected to the output of the first stage and the output of the second fully connected network, respectively, and is used to multiply the first feature map element-wise with the sparse channel weights to obtain the sparse feature map. The output of the sixth multiplier is connected to the input of the second normalization layer in the third feature fusion module, and the output of the seventh multiplier is connected to the input of the second normalization layer in the fourth feature fusion module.

[0030] Furthermore, each stage in the backbone network is formed by stacking multiple Convnext blocks. The first stage, the second stage, and the fourth stage are all stacked in 3 layers, while the third stage is stacked in 27 layers.

[0031] Furthermore, the encoder also includes a preprocessing module; the input end of the preprocessing module is used to receive externally input spacecraft component images and preprocess the input spacecraft component images; the output end of the preprocessing module is connected to the input end of the first stage.

[0032] The preprocessing includes data cropping and data augmentation, with data augmentation including random flipping and scaling.

[0033] Furthermore, the kernel size of the head convolutional layer, the third convolutional layer, the fifth convolutional layer, and the sixth convolutional layer is 1×1;

[0034] The kernel size of the first, second, fourth, and seventh convolutional layers is 3×3.

[0035] The kernel size of the block convolutional layer is 8×8.

[0036] Furthermore, the first activation layer, the second activation layer, and the third activation layer are all activated using the GELU function.

[0037] In addition, the present invention also provides a method for identifying spacecraft components, which is characterized by including the following steps:

[0038] Step 1: Construct the adaptive channel filtering and feature fusion network described above for spacecraft component identification;

[0039] Step 2: Construct a training set, which includes multiple images of spacecraft components and their corresponding original mask labels;

[0040] Step 3: Input multiple spacecraft component images and their corresponding original mask labels from the training set into the adaptive channel filtering and feature fusion network for optimization training. During optimization training, generate predicted mask labels based on the refined multi-scale fusion features output from the first convolutional layer, compare the predicted mask labels with the original mask labels, and calculate the cross-entropy loss. Finally, use the cross-entropy loss to backpropagate in the adaptive channel filtering and feature fusion network to achieve optimized training of the adaptive channel filtering and feature fusion network.

[0041] Step 4: Input the image of the spacecraft component to be tested into the optimized and trained adaptive channel filtering and feature fusion network to complete the identification of the spacecraft component.

[0042] Compared with the prior art, the present invention has the following beneficial technical effects:

[0043] 1. The adaptive channel filtering and feature fusion network for spacecraft component identification provided by this invention includes an encoder comprising a backbone network and a channel filtering module. The channel filtering module relies on the DCT algorithm and the RPCA algorithm to accurately evaluate the response performance of each channel in important frequency bands, quantify and determine the information importance of each channel, and then adaptively filter out key feature channels suitable for large and small components. This effectively eliminates redundant feature interference and significantly enhances the recognition accuracy and scene adaptability of the adaptive channel filtering and feature fusion network for components of various sizes, laying an efficient and accurate foundation for subsequent feature extraction.

[0044] 2. The adaptive channel filtering and feature fusion network for spacecraft component identification provided by this invention includes a decoder comprising four feature fusion modules, a feature splicing layer, a first convolutional layer, and a segmentation head. The feature fusion module uses a hybrid expert system as its core architecture, innovatively combining two key feature parameters: local contrast and local variance. The block gating module then performs multi-dimensional comprehensive evaluation and intelligent integration of the extracted global expert feature map, local expert feature map, and edge expert feature map based on a multi-dimensional feature weight map of block weights, achieving adaptive dynamic feature fusion. This design can effectively prevent the weak features of small components from being diluted by redundant information, significantly enhance the recognizability of small component features, and further improve the network's recognition accuracy and robustness for small components. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of an embodiment of the adaptive channel filtering and feature fusion network for spacecraft component identification according to the present invention;

[0046] Figure 2 This is a schematic diagram of the channel filtering module in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the feature fusion module in an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, advantages, and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0049] like Figure 1 As shown, an adaptive channel filtering and feature fusion network for spacecraft component identification includes an encoder and a decoder. The encoder includes a preprocessing module, a backbone network, and a channel filtering module. The decoder includes four feature fusion modules, a feature splicing layer, a first convolutional layer, and a segmentation head.

[0050] The preprocessing module receives externally input images of spacecraft components, performs preprocessing on these images, and uses the preprocessed images as input signals for the backbone network. Preprocessing includes data cropping and data augmentation, where data augmentation includes random flipping and scaling.

[0051] The backbone network consists of four stages and three downsampling layers. The four stages are sequentially designated as Stage 1, Stage 2, Stage 3, and Stage 4. The input of Stage 1 is connected to the output of the preprocessing module. The input of the first downsampling layer is connected to the output of Stage 1, and its output is connected to the input of Stage 2. The input of the second downsampling layer is connected to the output of Stage 2, and its output is connected to the input of Stage 3. The input of the third downsampling layer is connected to the output of Stage 3, and its output is connected to the input of Stage 4. The preprocessed spacecraft component image undergoes four feature extractions in these four stages, resulting in four feature maps, which are sequentially designated as the first feature map. The second feature map The third feature map and the fourth feature map Due to the downsampling operations of the three downsampling layers, the size of the four feature maps decreases sequentially.

[0052] In this embodiment, each of the four stages includes multiple residual structure Convnext blocks, which are stacked. The first, second, and fourth stages are all stacked with 3 layers (Convnext block × 3), while the third stage is stacked with 27 layers (Convnext block × 27). The convolutional kernel size of the three downsampling layers is 2 × 2.

[0053] like Figure 2 As shown, the channel filtering module includes a DCT transform unit, an RPCA decomposition filtering unit, a low-rank weight generation unit, a sparse weight generation unit, a low-rank feature output unit, and a sparse feature output unit. The input of the DCT transform unit is connected to the output of the first stage, and is used to perform a channel-by-channel discrete DCT transform on the first feature map F1 to obtain the response signal D of each feature channel in different frequency bands. hw Then, the response signals D of each characteristic channel in different frequency bands hw Expanding along the length and width dimensions yields the matrix-form response signal D. n .

[0054] The input of the RPCA decomposition and filtering unit is connected to the output of the DCT transform unit, and is used to transform the matrix-form response signal D... n Singular value decomposition yields a low-rank matrix D. ls0 and sparse matrix D sp0 Then, by filtering the low-rank and sparse matrices, the low-rank frequency band response characteristics D are obtained. ls1 and sparse frequency band response characteristics D sp1 In the case of a low-rank matrix D ls0 During the selection process, the 12 low-rank frequency bands with the largest L2 norm are retained. This is then applied to the sparse matrix D.sp0 During the screening process, the 24 sparse frequency bands with the largest L1 norm are retained.

[0055] The low-rank weight generation unit includes a first learning parameter configuration module, a first frequency band index loading module, a first frequency band response loading module, a second multiplier, a second softmax layer, a third multiplier, a sixth normalization layer, and a first fully connected network. The inputs of the first learning parameter configuration module, the first frequency band index loading module, and the first frequency band response loading module are respectively connected to the first output of the RPCA decomposition and filtering unit, used to feed the low-rank frequency band response features D... ls1 Configure the learnable parameter vector Q lr Loading low-rank frequency band response characteristics D ls1 Frequency band index, loading low-rank frequency band response characteristics D ls1 The frequency band response; the input of the second multiplier is connected to the output of the first learning parameter configuration module and the output of the first frequency band index loading module, respectively, to process the low-rank frequency band response feature D. ls1 The frequency band index and the corresponding learnable parameter vector Q lr Perform element-wise multiplication to obtain the weighted response S of the low-rank characteristic frequency band. lr The input of the second softmax layer is connected to the output of the second multiplier, and is used to weight the low-rank feature frequency band response S. lr Perform softmax function operation to realize the weighted response S of the low-rank characteristic frequency band. lr Normalization processing; the input of the third multiplier is connected to the output of the second softmax layer and the output of the first frequency band response loading module, respectively, to weight the low-rank characteristic frequency band response S after the softmax function operation. lr With low-rank frequency band response characteristics D ls1 The frequency band response is then multiplied element-wise again to obtain the low-rank feature weight intermediate value A. lr The input of the sixth normalization layer is connected to the output of the third multiplier, and its output is connected to the input of the first fully connected network. This is used to process the intermediate value A of the low-rank feature weights. lr Normalization is performed to eliminate feature distribution bias, and then the input is fed into the first fully connected network for feature dimension adjustment to obtain the low-rank channel weights W. lr .

[0056] The sparse weight generation unit includes a second learning parameter configuration module, a second frequency band index loading module, a second frequency band response loading module, a fourth multiplier, a third softmax layer, a fifth multiplier, a seventh normalization layer, and a second fully connected network. The inputs of the second learning parameter configuration module, the second frequency band index loading module, and the second frequency band response loading module are respectively connected to the second output of the RPCA decomposition and filtering unit, used to feed the sparse frequency band response features D...sp1 Configure the learnable parameter vector Q sp Loading sparse frequency band response characteristics D sp1 Frequency band index, loading sparse frequency band response characteristics D sp1 The frequency band response; the input of the fourth multiplier is connected to the output of the second learning parameter configuration module and the output of the second frequency band index loading module, respectively, to process the sparse frequency band response feature D. sp1 The frequency band index and the corresponding learnable parameter vector Q sp Perform element-wise multiplication to obtain the sparse characteristic frequency band weighted response S. sp The input of the third softmax layer is connected to the output of the fourth multiplier, and is used to weight the response S of the sparse feature frequency band. sp Perform softmax function operation to realize the sparse feature band weighted response S sp Normalization processing; the input of the fifth multiplier is connected to the output of the third softmax layer and the output of the second frequency band response loading module, respectively, to weight the sparse characteristic frequency band response S after the softmax function operation. sp With sparse frequency band response characteristics D sp1 The frequency band responses are multiplied element-wise again to obtain the sparse feature weight intermediate value A. sp The input of the seventh normalization layer is connected to the output of the fifth multiplier, and its output is connected to the input of the second fully connected network. This is used to adjust the intermediate value A of the sparse feature weights. sp Normalization is performed to eliminate feature distribution bias, and then the feature dimensions are adjusted by inputting it into the second fully connected network to obtain the sparse channel weights W. sp .

[0057] Among them, the low-rank frequency band response characteristic D ls1 The corresponding learnable parameter vector Q lr The dimension is consistent with the number of low-rank feature frequency bands after screening, and the sparse frequency band response feature D sp1 The corresponding learnable parameter vector Q sp The dimension is consistent with the number of sparse feature frequency bands after filtering. Both types of learnable parameter vectors are initialized with random values ​​and updated synchronously with network training. Both the first and second fully connected networks contain two hidden layers with 256 and 128 neurons respectively, and the activation function for both is GELU.

[0058] The low-rank feature output unit includes a sixth multiplier. The input of the sixth multiplier is connected to the output of the first stage and the output of the first fully connected network, respectively, and is used to process the first feature map. With low-rank channel weights W lr Element-wise multiplication yields the low-rank feature map D. lrThe sparse feature output unit includes a seventh multiplier. The input of the seventh multiplier is connected to the output of the first stage and the output of the second fully connected network, respectively, and is used to process the first feature map. With sparse channel weights W sp Element-wise multiplication yields the sparse feature map D. sp .

[0059] like Figure 1 As shown, the decoder includes four feature fusion modules, a feature concatenation layer, a first convolutional layer, and a segmentation head. The input of the first feature fusion module is connected to the outputs of the third and fourth stages, respectively, to fuse the third and fourth feature maps to obtain a first feature fusion map. The input of the second feature fusion module is connected to the outputs of the first and second stages, respectively, to fuse the first and second feature maps to obtain a second feature fusion map. The input of the third feature fusion module is connected to the outputs of the second and low-rank feature output units, respectively, to fuse the second and low-rank feature maps to obtain a third feature fusion map. The input of the fourth feature fusion module is connected to the outputs of the third and sparse feature output units, respectively, to fuse the third and sparse feature maps to obtain a fourth feature fusion map.

[0060] The four feature fusion modules have essentially the same structure, each including a feature fusion preprocessing module, a multi-prior extraction module, a hybrid expert feature extraction module, a block gating module, and a weighted fusion module. If the two feature maps input to the feature fusion module have different sizes, the feature fusion preprocessing module first interpolates the input low-resolution feature map to the same size as the high-resolution feature map using bilinear interpolation, and then normalizes both the interpolated and high-resolution feature maps respectively. If the two feature maps input to the feature fusion module have the same size, then normalization is performed directly on both feature maps. Therefore, since the two input sizes of the first to third feature fusion modules are different, while the two input sizes of the fourth feature fusion module are the same, the feature fusion preprocessing modules of the first to third feature fusion modules include interpolation modules, while the feature fusion preprocessing module of the fourth feature fusion module does not require an interpolation module.

[0061] like Figure 3As shown, the feature fusion preprocessing modules in the first to third feature fusion modules each include an interpolation module, a first normalization layer, a second normalization layer, and a first concatenation module; the fourth feature fusion module includes a first normalization layer, a second normalization layer, and a first concatenation module. In the first feature fusion module, the input of the interpolation module is connected to the output of the fourth stage, and the input of the first normalization layer is connected to the output of the interpolation module in the same feature fusion module, used to interpolate the fourth feature map. The size is increased to match the third feature map. The same features are obtained and normalized; the input of the second normalization layer is connected to the output of the third stage, which is used to process the third feature map. Normalization is performed; the two inputs of the first splicing module are respectively connected to the outputs of the first and second normalization layers, and are used to process the normalized fourth feature map. and the third feature map The features are then stitched together to obtain the first stitched feature map. In the second feature fusion module, the input of the interpolation module is connected to the output of the first feature fusion module, and the input of the first normalization layer is connected to the output of the interpolation module within the same feature fusion module. This is used to increase the size of the first fused feature map to match that of the second feature map through interpolation. The same features are processed and normalized; the input of the second normalization layer is connected to the output of the second stage, and is used to process the second feature map. Normalization is performed; the two inputs of the first stitching module are respectively connected to the outputs of the first and second normalization layers in the same feature fusion module, and are used to process the normalized first feature fusion map and the second feature map. The features are then concatenated to obtain a second concatenated feature map. In the third feature fusion module, the input of the interpolation module is connected to the output of the second feature fusion module, and the input of the first normalization layer is connected to the output of the interpolation module within the same feature fusion module. This is used to increase the size of the second feature fusion map to be similar to that of the low-rank feature map D through interpolation. lr The same values ​​are obtained and normalized; the input of the second normalization layer is connected to the output of the sixth multiplier, which is used to process the low-rank feature map D. lr Normalization is performed; the two inputs of the first splicing module are respectively connected to the outputs of the first and second normalization layers in the same feature fusion module, and are used to process the normalized second feature fusion map and the low-rank feature map D. lrThe features are then concatenated to obtain the third concatenated feature map. In the fourth feature fusion module, the input of the first normalization layer is connected to the output of the third feature fusion module, used to normalize the third feature map; the input of the second normalization layer is connected to the output of the seventh multiplier, used to normalize the sparse feature map D. sp Normalization is performed; the two inputs of the first stitching module are respectively connected to the outputs of the first and second normalization layers in the same feature fusion module, and are used to process the normalized third feature fusion map and sparse feature map D. sp The features are then stitched together to obtain the fourth stitched feature map.

[0062] The multi-prior extraction module is used to comprehensively evaluate the effectiveness of feature information by combining local contrast and local variance parameters. It includes a multi-prior extractor and a second stitching module. The input of the multi-prior extractor is connected to the output of the first stitching module in the same feature fusion module, and is used to perform prior extraction on the received stitched feature map to obtain the standard deviation feature map σ and the extreme value difference feature map r. The first input of the second stitching module is connected to the output of the first stitching module in the same feature fusion module, and the second input is connected to the output of the multi-prior extractor. It is used to stitch the received stitched feature map, standard deviation feature map σ, and extreme value difference feature map r along the channel dimension to obtain a multi-dimensional stitched feature map M.

[0063] The hybrid expert feature extraction module includes a global expert module, a local expert module, and a peripheral expert module.

[0064] The local expert module consists of a global average pooling layer, a multi-layer fully connected network, a first multiplier, and a head convolutional layer connected in sequence; the local expert module consists of a second convolutional layer, a third normalization layer, a first activation layer, and a third convolutional layer connected in sequence; the edge expert module consists of a grouped convolutional module, a fourth convolutional layer, a fourth normalization layer, a second activation layer, and a fifth convolutional layer connected in sequence, wherein the grouped convolutional module includes a horizontal Sobel convolutional kernel, a vertical Sobel convolutional kernel, and a fusion computation module.

[0065] The path to obtain the global expert feature map: The input of the global average pooling layer is connected to the output of the first stitching module in the same feature fusion module, which is used to perform adaptive average pooling on the received stitched feature map. In this embodiment, the stitched feature map is pooled to a 1×1 size to extract the global feature map M. g0 A multi-layer fully connected network is used to process the extracted global feature map M. g0 Channel attention weights W are generated through network computation. g In this embodiment, the multilayer fully connected network includes two linear layers, namely two GELU activation layers and two fully connected layers, with a global feature map M. g0After the first GELU activation layer, the first fully connected layer compresses its feature dimension from the number of input channels to half the number of input channels. After the second GELU activation layer, the second fully connected layer restores its feature dimension to the number of input channels. The other input of the first multiplier is connected to the output of the first concatenation module in the same feature fusion module, used to apply the channel attention weights W. g After multiplying with the concatenated feature map, it is then convolved by the head convolutional layer to obtain the global expert feature map M. g The kernel size of the head convolutional layer is 1×1.

[0066] The path to obtain the local expert feature map: The input of the second convolutional layer is connected to the output of the first stitching module in the same feature fusion module. This connection is used to convolve the received stitched feature map, adjust the number of feature channels, and obtain the local feature map M. l0 The third normalization layer, the first activation layer, and the third convolutional layer are used to process the local feature map M. l0 Normalization, activation, and feature integration are performed sequentially to obtain the local expert feature map M. l The first activation layer uses the GELU function for activation, the second convolutional layer has a kernel size of 3×3, and the third convolutional layer has a kernel size of 1×1.

[0067] The acquisition path of the edge expert feature map: Grouped convolution is used, with the number of groups matching the number of input channels. Specifically, the inputs of both the horizontal and vertical Sobel convolution kernels are connected to the output of the first concatenation module in the same feature fusion module. This output is used to extract gradients from the received concatenated feature map, resulting in the horizontal gradient feature map M. ex and vertical gradient feature map M ey The input of the fusion calculation module is connected to the output of the corresponding horizontal Sobel convolution kernel and the output of the corresponding vertical Sobel convolution kernel, respectively, for processing the horizontal gradient feature map M. ex and vertical gradient feature map M ey Perform square root and squaring operations to obtain the fused gradient feature map M. e0 The input of the fourth convolutional layer is connected to the output of the fusion calculation module, and is used to process the fused gradient feature map M. e0 Convolution is performed to obtain edge features M e1 The fourth normalization layer, the second activation layer, and the fifth convolutional layer are used to process the edge features M. e1 Normalization, activation, and feature integration are performed sequentially to obtain the edge expert feature map M. e The second activation layer uses the GELU function for activation, the fourth convolutional layer has a kernel size of 3×3, and the fifth convolutional layer has a kernel size of 1×1.

[0068] The block-gating module comprises, in sequence, a block convolutional layer, a fifth normalization layer, a third activation layer, a sixth convolutional layer, a first softmax layer, a seventh convolutional layer, a bilinear upsampling layer, and a gated weighted fusion module. The input of the block convolutional layer is connected to the output of the second stitching module within the same feature fusion module, and is used to perform skip convolution operations on the multidimensional stitched feature map M to obtain the initial block feature map. The fifth normalization layer, the third activation layer, and the sixth convolutional layer are used to sequentially normalize, activate, and integrate the initial block feature map to obtain the block logical value feature map. The first softmax layer is used to perform softmax normalization on the block logical value feature map to obtain the block weights W. G The seventh convolutional layer is used to combine the block weights W. G Further convolutions are performed, and the size is enlarged through a bilinear upsampling layer to obtain the block-weighted W. G The multidimensional feature weight map; the input of the gated weighted fusion module is connected to the outputs of the bilinear upsampling layer, the head convolutional layer, the third convolutional layer, and the fifth convolutional layer in the same feature fusion module, respectively, for passing the multidimensional feature weight map based on block weights W. G Multidimensional feature weight map guides global expert feature map M g Local expert feature map M l Edge expert feature map M e Perform a weighted summation (multiplying and summing element by element) to obtain the fusion logic feature map M. G0 The third activation layer uses the GELU function for activation. The kernel size of the block convolutional layer is 8×8 with a stride of 8. The kernel size of the sixth convolutional layer is 1×1, and the kernel size of the seventh convolutional layer is 3×3.

[0069] The input of the weighted fusion module is connected to the outputs of the gated weighted fusion module, the first normalization layer, and the second normalization layer in the same feature fusion module, respectively, for fusing the logical feature map M. G0 The feature maps output from the first and second normalization layers are weighted and summed to obtain the corresponding feature fusion map.

[0070] The input of the feature stitching layer is connected to the output of the weighted fusion module in each of the four feature fusion modules. It stitches together the feature fusion maps output by the four weighted fusion modules to obtain multi-scale fused features. The input of the first convolutional layer is connected to the output of the feature stitching layer. It performs convolution processing on the obtained multi-scale fused features to remove redundant interference, resulting in refined multi-scale fused features. The input of the segmentation head is connected to the output of the first convolutional layer. It generates predictive mask labels based on the refined multi-scale fused features to achieve spacecraft component identification. The kernel size of the first convolutional layer is 3×3.

[0071] Based on the aforementioned adaptive channel filtering and feature fusion network, the specific steps of the spacecraft component identification method of the present invention are as follows:

[0072] Step 1: Construct the training set.

[0073] The training set includes multiple images of spacecraft components and their corresponding original mask labels. All spacecraft component images undergo preprocessing using Gaussian blur and data augmentation to ensure consistent image dimensions. Then, the original mask labels for each spacecraft component image are obtained. It is important to note that the number of spacecraft component images needs to exceed the preset number of iterations for the adaptive channel filtering and feature fusion network.

[0074] Step 2: Training the adaptive channel filtering and feature fusion network.

[0075] Multiple spacecraft component images and their corresponding original mask labels from the training set are sequentially input into an adaptive channel filtering and feature fusion network for optimization training.

[0076] During optimization training, predicted mask labels are generated based on the multi-scale fusion features output from the first convolutional layer. These predicted mask labels are then compared with the original mask labels, and the cross-entropy loss is calculated. Finally, the cross-entropy loss is used for backpropagation in the adaptive channel selection and feature fusion network to achieve optimized training of the network. After reaching a preset number of iterations, the optimal network parameters are retained, resulting in the optimized adaptive channel selection and feature fusion network.

[0077] Step 3: Input the image of the spacecraft component to be tested into the optimized and trained adaptive channel filtering and feature fusion network to complete the identification of the spacecraft component.

[0078] To verify the effectiveness of the spacecraft component identification method of this invention, the Intersection over Union (IoU) and Accuracy (Accuracy) are used as evaluation metrics. The existing spacecraft component identification methods and the spacecraft component identification method based on adaptive channel filtering and feature fusion networks of this invention are tested and compared on the public dataset UESD. The existing spacecraft component identification methods include those based on Deeplabv3+, HRnet, DANet, Segformer, Swin-transformer, Uppernet, Convnext, and Defmamba.

[0079] Different methods were implemented on a PyTorch platform equipped with an NVIDIA GeForce RTX 4090 GPU (24GB RAM). To ensure consistency, the input images were resized to a uniform 512×512 and normalized. The batch size was 4, and the preset number of iterations was 160,000. The methods of this invention, the Segformer-based spacecraft component recognition method, the Swin-transformer-based spacecraft component recognition method, the Upernet-based spacecraft component recognition method, the Convnext-based spacecraft component recognition method, and the Defmamba-based spacecraft component recognition method used the Adamw optimizer to optimize the corresponding deep learning networks with an initial learning rate of 0.0001. The remaining methods used the SGD optimizer to optimize the corresponding deep learning networks with an initial learning rate of 0.01. The IoU results of different methods on the UESD dataset are shown in Table 1, and the Acc results of different methods on the UESD dataset are shown in Table 2.

[0080] Table 1. IoU results of different methods on the UESD dataset

[0081]

[0082] Table 2. Acc results for different methods on the UESD dataset.

[0083]

[0084] Generally, higher values ​​for IoU and Ac are better. Through comparative analysis, it was found that compared with existing spacecraft component identification methods, the spacecraft component identification method based on adaptive channel filtering and feature fusion network of this invention can significantly improve the identification accuracy of small components such as instruments, thrusters, and optical payloads while ensuring high accuracy.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. An adaptive channel filtering and feature fusion network for spacecraft component identification, comprising an encoder and a decoder, characterized in that: The encoder includes a backbone network and a channel filtering module; the backbone network includes four sequentially arranged stages and three downsampling layers, each downsampling layer being connected between two adjacent stages, used to extract features from the input spacecraft component image four times in sequence to obtain four feature maps of different sizes; The channel filtering module includes a DCT transform unit, an RPCA decomposition filtering unit, a low-rank weight generation unit, a sparse weight generation unit, a low-rank feature output unit, and a sparse feature output unit. The input of the DCT transform unit is connected to the output of the first stage, and is used to perform channel-by-channel discrete DCT transform and expansion on the first feature map to obtain a matrix response signal. The input of the RPCA decomposition and filtering unit is connected to the output of the DCT transform unit. It is used to perform singular value decomposition on the matrix response signal to obtain a low-rank matrix and a sparse matrix. Then, the low-rank matrix and the sparse matrix are filtered to obtain the low-rank frequency band response characteristics and the sparse frequency band response characteristics. The inputs of the low-rank weight generation unit and the sparse weight generation unit are respectively connected to the two outputs of the RPCA decomposition and screening unit. They are used to perform weight generation processing on the low-rank frequency band response features and the sparse frequency band response features to obtain the low-rank channel weights and the sparse channel weights, respectively. The input of the low-rank feature output unit is connected to the output of the first stage and the output of the low-rank weight generation unit, respectively. The input of the sparse feature output unit is connected to the output of the first stage and the output of the sparse weight generation unit, respectively. It is used to multiply the first feature map with the low-rank channel weights and the sparse channel weights element by element to obtain the low-rank feature map and the sparse feature map, respectively. The decoder includes four feature fusion modules, a feature splicing layer, a first convolutional layer, and a segmentation head; The input of the first feature fusion module is connected to the output of the third stage and the output of the fourth stage, respectively, and is used to fuse the third feature map and the fourth feature map to obtain the first feature fusion map; The input of the second feature fusion module is connected to the output of the first feature fusion module and the output of the second stage, respectively, and is used to perform feature fusion on the first feature fusion map and the second feature map to obtain the second feature fusion map; The input of the third feature fusion module is connected to the output of the second feature fusion module and the output of the low-rank feature output unit, respectively, and is used to perform feature fusion on the second feature fusion map and the low-rank feature map to obtain the third feature fusion map; The input of the fourth feature fusion module is connected to the output of the third feature fusion module and the output of the sparse feature output unit, respectively, and is used to perform feature fusion on the third feature fusion map and the sparse feature map to obtain the fourth feature fusion map. The input of the feature splicing layer is connected to the output of the four feature fusion modules respectively, and is used to splice the first to fourth feature fusion maps to obtain multi-scale fusion features; the input of the first convolutional layer is connected to the output of the feature splicing layer, and is used to perform convolution processing on the obtained multi-scale fusion features to obtain refined multi-scale fusion features. The input of the segmentation head is connected to the output of the first convolutional layer, and is used to generate predicted mask labels based on the refined multi-scale fusion features to achieve the identification of spacecraft components.

2. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 1, characterized in that: The feature fusion module includes a feature fusion preprocessing module, a multi-priority extraction module, a hybrid expert feature extraction module, a block gating module, and a weighted fusion module; In the first feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the fourth stage and the output of the third stage, respectively. It is used to normalize the fourth feature map and the third feature map respectively, and then stitch the two normalized feature maps together to obtain the first stitched feature map. In the second feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the first feature fusion module and the output of the second stage, respectively. It is used to normalize the first feature fusion map and the second feature map, and then stitch the two normalized feature maps together to obtain the second stitched feature map. In the third feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the second feature fusion module and the output of the low-rank feature output unit, respectively. It is used to normalize the second feature fusion map and the low-rank feature map, and then splice the two normalized feature maps to obtain the third spliced ​​feature map. In the fourth feature fusion module, the input of the feature fusion preprocessing module is connected to the output of the third feature fusion module and the output of the sparse feature output unit, respectively. It is used to normalize the third feature fusion map and the sparse feature map, and then splice the two normalized feature maps to obtain the fourth spliced ​​feature map. The input of the multi-priority extraction module is connected to the first output of the feature fusion preprocessing module in the same feature fusion module. It is used to extract priors from the corresponding spliced ​​feature map to obtain the standard deviation feature map and the extreme value difference feature map. Then, the spliced ​​feature map, the standard deviation feature map and the extreme value difference feature map are spliced ​​together in the channel dimension to obtain a multi-dimensional spliced ​​feature map. The hybrid expert feature extraction module includes a global expert module, a local expert module, and an edge expert module. The inputs of the global expert module, the local expert module, and the edge expert module are respectively connected to the first output of the feature fusion preprocessing module in the same feature fusion module. They are used to perform global expert feature extraction, local expert feature extraction, and edge expert feature extraction on the spliced ​​feature map respectively, so as to obtain a global expert feature map, a local expert feature map, and an edge expert feature map respectively. The input of the block gating module is connected to the output of the multi-prior extraction module in the same feature fusion module. It is used to obtain a multi-dimensional feature weight map based on block weights based on the multi-dimensional spliced ​​feature map, and guide the global expert feature map, local expert feature map, and edge expert feature map to perform weighted summation to obtain the fusion logic feature map. The input of the weighted fusion module is connected to the output of the block gating module and the second and third outputs of the feature fusion preprocessing module in the same feature fusion module, respectively. It is used to guide the two normalized feature maps to perform weighted summation through the fusion logic feature map to obtain the corresponding feature fusion map. The output of the weighted fusion module in the four feature fusion modules is connected to the input of the feature splicing layer.

3. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 2, characterized in that: The feature fusion preprocessing modules in the first to third feature fusion modules all include an interpolation module, a first normalization layer, a second normalization layer, and a first splicing module. The fourth feature fusion module includes a first normalization layer, a second normalization layer, and a first splicing module. In the first feature fusion module, the input of the interpolation module is connected to the output of the fourth stage, and the input of the first normalization layer is connected to the output of the interpolation module in the same feature fusion module. It is used to increase the size of the fourth feature map to be the same as the third feature map through interpolation and perform normalization processing. The input of the second normalization layer is connected to the output of the third stage and is used to normalize the third feature map. The two input terminals of the first splicing module are connected to the output terminals of the first normalization layer and the second normalization layer, respectively, and are used to splice the fourth and third feature maps after normalization to obtain the first spliced ​​feature map. In the second feature fusion module, the input of the interpolation module is connected to the output of the first feature fusion module, and the input of the first normalization layer is connected to the output of the interpolation module in the same feature fusion module. This is used to increase the size of the first feature fusion map to be the same as the second feature map through interpolation and to perform normalization processing. The input of the second normalization layer is connected to the output of the second stage and is used to normalize the second feature map. The two input terminals of the first splicing module are respectively connected to the output terminals of the first normalization layer and the second normalization layer in the same feature fusion module, and are used to splice the normalized first feature fusion map and the second feature map to obtain the second spliced ​​feature map. In the third feature fusion module, the input of the interpolation module is connected to the output of the second feature fusion module, and the input of the first normalization layer is connected to the output of the interpolation module in the same feature fusion module. It is used to increase the size of the second feature fusion map to be the same as the low-rank feature map through interpolation and perform normalization processing. The input of the second normalization layer is connected to the first output of the channel filtering module and is used to normalize the low-rank feature map. The two inputs of the first splicing module are respectively connected to the outputs of the first normalization layer and the second normalization layer in the same feature fusion module. They are used to splice the normalized second feature fusion map and the low-rank feature map to obtain the third spliced ​​feature map. In the fourth feature fusion module, the input of the first normalization layer is connected to the output of the third feature fusion module, and is used to normalize the third feature fusion map; the input of the second normalization layer is connected to the second output of the channel filtering module, and is used to normalize the sparse feature map. The two inputs of the first splicing module are respectively connected to the outputs of the first normalization layer and the second normalization layer in the same feature fusion module. They are used to splice the normalized third feature fusion map and the sparse feature map to obtain the fourth spliced ​​feature map. The multiple prior extraction module includes a multiple prior extractor and a second splicing module; the input end of the multiple prior extractor is connected to the output end of the first splicing module in the same feature fusion module, and is used to perform prior extraction on the received spliced ​​feature map to obtain the standard deviation feature map and the extreme value difference feature map; The first input of the second stitching module is connected to the output of the first stitching module in the same feature fusion module, and the second input is connected to the output of the multiple prior extractor. It is used to stitch the received stitched feature map, standard deviation feature map and extreme value difference feature map along the channel dimension to obtain a multi-dimensional stitched feature map. The global expert module includes a global average pooling layer, a multi-layer fully connected network, a first multiplier, and a head convolutional layer connected in sequence; the local expert module includes a second convolutional layer, a third normalization layer, a first activation layer, and a third convolutional layer connected in sequence; the edge expert module includes a grouped convolutional module, a fourth convolutional layer, a fourth normalization layer, a second activation layer, and a fifth convolutional layer connected in sequence, wherein the grouped convolutional module includes a horizontal Sobel convolutional kernel, a vertical Sobel convolutional kernel, and a fusion computation module; The input of the global average pooling layer is connected to the output of the first concatenation module in the same feature fusion module, and is used to perform adaptive average pooling on the received concatenated feature map to extract the global feature map; the multilayer fully connected network is used to generate channel attention weights based on the extracted global feature map; the other input of the first multiplier is connected to the output of the first concatenation module in the same feature fusion module, and is used to multiply the channel attention weights with the concatenated feature map to obtain the global expert feature map; The input of the second convolutional layer is connected to the output of the first stitching module in the same feature fusion module, and is used to convolve the received stitched feature map to obtain a local feature map; the third normalization layer, the first activation layer, and the third convolutional layer are used to perform normalization processing, activation processing, and feature integration on the local feature map in sequence to obtain a local expert feature map; The inputs of the horizontal and vertical Sobel convolutional kernels are both connected to the output of the first stitching module in the same feature fusion module. This is used to extract gradients from the received stitched feature maps, obtaining horizontal and vertical gradient feature maps. The input of the fusion calculation module is connected to the outputs of the corresponding horizontal and vertical Sobel convolutional kernels, respectively. This is used to perform square root and squaring operations on the horizontal and vertical gradient feature maps, obtaining a fused gradient feature map. The input of the fourth convolutional layer is connected to the output of the fusion calculation module, used to convolve the fused gradient feature map to obtain edge features. The fourth normalization layer, the second activation layer, and the fifth convolutional layer are used to perform normalization, activation, and feature integration on the edge features sequentially, obtaining an edge expert feature map. The block gating module includes a block convolutional layer, a fifth normalization layer, a third activation layer, a sixth convolutional layer, a first softmax layer, a seventh convolutional layer, a bilinear upsampling layer, and a gated weighted fusion module connected in sequence. The input of the block convolutional layer is connected to the output of the second stitching module in the same feature fusion module, and is used to perform skip convolution operations on the multidimensional stitched feature map to obtain the initial block feature map; the fifth normalization layer, the third activation layer, and the sixth convolutional layer are used to perform normalization, activation, and feature integration on the initial block feature map in sequence to obtain the block logical value feature map; the first softmax layer is used to perform softmax normalization on the block logical value feature map to obtain the block weights; the seventh convolutional layer is used to perform further convolution in combination with the block weights and to enlarge the size through the bilinear upsampling layer to obtain a multidimensional feature weight map based on the block weights; the input of the gated weighted fusion module is connected to the output of the bilinear upsampling layer, the head convolutional layer, the third convolutional layer, and the fifth convolutional layer in the same feature fusion module, and is used to guide the global expert feature map, the local expert feature map, and the edge expert feature map to perform weighted summation through the multidimensional feature weight map based on the block weights to obtain the fused logical feature map; The input of the weighted fusion module is connected to the output of the gated weighted fusion module, the first normalization layer, and the second normalization layer in the same feature fusion module. It is used to guide the feature maps output by the first normalization layer and the second normalization layer to perform weighted summation through the fusion logic feature map to obtain the corresponding feature fusion map.

4. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 3, characterized in that: When filtering low-rank matrices, the RPCA decomposition and filtering unit retains the 12 low-rank frequency bands with the largest L2 norm, and when filtering sparse matrices, it retains the 24 sparse frequency bands with the largest L1 norm.

5. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 4, characterized in that: The low-rank weight generation unit includes a first learning parameter configuration module, a first frequency band index loading module, a first frequency band response loading module, a second multiplier, a second softmax layer, a third multiplier, a sixth normalization layer, and a first fully connected network; the sparse weight generation unit includes a second learning parameter configuration module, a second frequency band index loading module, a second frequency band response loading module, a fourth multiplier, a third softmax layer, a fifth multiplier, a seventh normalization layer, and a second fully connected network. The inputs of the first learning parameter configuration module, the first frequency band index loading module, and the first frequency band response loading module are respectively connected to the first output of the RPCA decomposition and filtering unit, and are used to configure learnable parameter vectors for low-rank frequency band response features, load frequency band indices of low-rank frequency band response features, and load frequency band responses of low-rank frequency band response features. The input of the second multiplier is respectively connected to the output of the first learning parameter configuration module and the output of the first frequency band index loading module, and is used to perform element-wise multiplication of the frequency band index of the low-rank frequency band response features with the corresponding learnable parameter vector to obtain the low-rank feature frequency band weighted response. The input of the second softmax layer is connected to the output of the second multiplier, and is used to perform softmax function operation on the low-rank feature frequency band weighted response. The input of the third multiplier is respectively connected to the output of the second softmax layer and the output of the first frequency band response loading module, and is used to perform element-wise multiplication again on the low-rank feature frequency band weighted response after softmax function operation with the frequency band response of the low-rank frequency band response features to obtain the intermediate value of the low-rank feature weight. The input of the sixth normalization layer is connected to the output of the third multiplier, and its output is connected to the input of the first fully connected network. It is used to normalize the intermediate values ​​of the low-rank feature weights, and then input them into the first fully connected network to adjust the feature dimensions to obtain the low-rank channel weights. The inputs of the second learning parameter configuration module, the second frequency band index loading module, and the second frequency band response loading module are respectively connected to the second output of the RPCA decomposition and filtering unit. These modules are used to configure learnable parameter vectors for the sparse frequency band response features, load the frequency band indexes of the sparse frequency band response features, and load the frequency band responses of the sparse frequency band response features. The input of the fourth multiplier is respectively connected to the output of the second learning parameter configuration module and the output of the second frequency band index loading module. This multiplier is used to perform element-wise multiplication of the frequency band indexes of the sparse frequency band response features with the corresponding learnable parameter vectors to obtain the sparse feature frequency band weighted response. The input of the third softmax layer is connected to the output of the fourth multiplier. This multiplier is used to perform softmax function operations on the sparse feature frequency band weighted response. The input of the fifth multiplier is respectively connected to the output of the third softmax layer and the output of the second frequency band response loading module. This multiplier is used to perform element-wise multiplication again between the sparse feature frequency band weighted response after the softmax function operation and the frequency band response of the sparse frequency band response features to obtain the intermediate value of the sparse feature weight. The input of the seventh normalization layer is connected to the output of the fifth multiplier, and its output is connected to the input of the second fully connected network. It is used to normalize the intermediate values ​​of the sparse feature weights, and then input them into the second fully connected network to adjust the feature dimensions to obtain the sparse channel weights. The low-rank feature output unit includes a sixth multiplier, and the sparse feature output unit includes a seventh multiplier. The input of the sixth multiplier is connected to the output of the first stage and the output of the first fully connected network, respectively, and is used to multiply the first feature map element-wise with the low-rank channel weights to obtain the low-rank feature map. The input of the seventh multiplier is connected to the output of the first stage and the output of the second fully connected network, respectively, and is used to multiply the first feature map element-wise with the sparse channel weights to obtain the sparse feature map. The output of the sixth multiplier is connected to the input of the second normalization layer in the third feature fusion module, and the output of the seventh multiplier is connected to the input of the second normalization layer in the fourth feature fusion module.

6. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 5, characterized in that: Each stage in the backbone network is formed by stacking multiple Convnext blocks. The first, second, and fourth stages are all stacked in 3 layers, while the third stage is stacked in 27 layers.

7. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 6, characterized in that: The encoder also includes a preprocessing module; the input end of the preprocessing module is used to receive externally input spacecraft component images and preprocess the input spacecraft component images. The output of the preprocessing module is connected to the input of the first stage; The preprocessing includes data cropping and data augmentation, with data augmentation including random flipping and scaling.

8. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 7, characterized in that: The kernel size of the head convolutional layer, the third convolutional layer, the fifth convolutional layer, and the sixth convolutional layer is 1×1. The kernel size of the first, second, fourth, and seventh convolutional layers is 3×3. The kernel size of the block convolutional layer is 8×8.

9. The adaptive channel filtering and feature fusion network for spacecraft component identification according to claim 8, characterized in that: The first activation layer, the second activation layer, and the third activation layer are all activated using the GELU function.

10. A method for identifying spacecraft components, characterized in that, Includes the following steps: Step 1: Construct the adaptive channel filtering and feature fusion network for spacecraft component identification as described in any of claims 1-9; Step 2: Construct a training set, which includes multiple images of spacecraft components and their corresponding original mask labels; Step 3: Input multiple spacecraft component images and their corresponding original mask labels from the training set into the adaptive channel filtering and feature fusion network for optimization training. During optimization training, generate predicted mask labels based on the refined multi-scale fusion features output from the first convolutional layer, compare the predicted mask labels with the original mask labels, and calculate the cross-entropy loss. Finally, use the cross-entropy loss to backpropagate in the adaptive channel filtering and feature fusion network to achieve optimized training of the adaptive channel filtering and feature fusion network. Step 4: Input the image of the spacecraft component to be tested into the optimized and trained adaptive channel filtering and feature fusion network to complete the identification of the spacecraft component.

Images