Multi-scale feature fusion star image recognition method and device based on attention mechanism

Abstract

The present invention discloses a method and device for star map recognition based on a multi-scale feature fusion of an attention mechanism. When training a star map recognition model based on deep learning, a multi-scale feature map generated in the feedforward calculation process of a convolutional neural network is extracted to form a multi-scale feature hierarchical pyramid structure. In addition, a multi-scale feature fusion method based on an attention mechanism is used to replace the traditional simple linear fusion operation, and the extracted multi-scale feature maps are fused layer by layer from top to bottom according to the network depth. A multi-scale attention module is used in the fusion process. The module is implemented in the process of a channel attention mechanism, and can aggregate multi-scale context information along the channel dimension to better fuse semantically and scale-inconsistent features. The present invention takes into account that the distribution of stars in a deep space environment may present multiple scales, while emphasizing star clusters with a more global distribution and star clusters with a more local distribution, so as to improve the accuracy of the model in identifying star maps under extreme scale changes in real deep space scenes.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and star sensor deep space detection, and in particular to a multi-scale feature fusion star map recognition method and device based on an attention mechanism. Background Art

[0002] Star map recognition is a key technology for star sensors to obtain high-precision attitude information. Reliable, fast and high-precision star map recognition methods have always been an important research topic. On the other hand, with the rapid development of artificial intelligence, neural networks have achieved good results in the field of image classification applications based on the unique advantages of deep learning. Therefore, computer vision technology based on deep learning also has broad application prospects in star map recognition tasks. The star map recognition methods based on deep convolutional neural networks proposed in the past use the last layer of feature maps output by bottom-up convolution for star map recognition. This type of basic method ignores the fact that the distribution of stars in the deep space environment may present multiple scales. It only uses the strong semantic features of the last layer of the neural network for recognition, abandons the use of detailed information of the shallow neural network, and is difficult to achieve ideal recognition effects for star maps of different distribution scales generated in real scenes. Therefore, in the real deep sky star map shooting scene, simply using the basic convolutional neural network to recognize the incoming star map may result in insufficient utilization of image features, thereby affecting the model performance and reducing the accuracy of star map recognition. Summary of the invention

[0003] The present invention provides a multi-scale feature fusion star map recognition method based on the attention mechanism, which can take into account the multi-scale distribution of deep-sky star maps in real scenes, that is, no matter how big or small the stars are, the method can accurately identify all the stars in the star map. Through this method, we can improve the accuracy of the attitude measurement of the aircraft in a real vacuum environment.

[0004] The embodiment of the present invention provides a multi-scale feature fusion star map recognition method based on an attention mechanism, comprising the following steps:

[0005] Step S1, constructing a basic training set of star map images and marking them;

[0006] Step S2, perform data enhancement on the basic training set;

[0007] Step S3, constructing a complete multi-scale feature fusion model based on the attention mechanism, wherein the multi-scale feature fusion model first extracts features of different scales through the backbone network, and then fuses adjacent feature maps from top to bottom according to the network depth through a multi-scale feature fusion method based on the attention mechanism, fusing two features each time, and the resulting new feature map continues to repeat the fusion process with the lower layer features until the iteration ends, and the final fused features from top to bottom are input into the classifier;

[0008] Step S4, batch-inputting the training set obtained by data enhancement into the above-constructed network model to train and generate a star map recognition model;

[0009] Step S5, input the star map to be identified into the generated star map identification model, the star map identification model outputs the identification result, the classification result output by each star map is the main star number, and finally the remaining stars in the star map are identified according to the angular distance between the main star and the adjacent stars in the star map.

[0010] Furthermore, the data enhancement in step S2 includes adding a certain range of random position deviations to the coordinates of each star in the star map to improve the robustness of the model to position errors; adding a certain range of random magnitude deviations to each star to improve the robustness of the model to magnitude errors; adding a number of fake stars with random positions and magnitudes to each basic star map image in the training set to improve the robustness of the model to fake star interference; in addition, the basic star map image in the training set is rotated from 0° to 360°, each time rotating a certain angle, to improve the rotation invariance of the model.

[0011] Furthermore, the specific implementation of the multi-scale feature fusion method based on the attention mechanism in step S3 is:

[0012] For the extracted adjacent feature maps, the spatial resolution of the upper feature map with a smaller resolution is upsampled, and the channel dimension of the lower feature map with a larger resolution is reduced. The two processed features are added point by point for initial feature fusion;

[0013] The features obtained by initial fusion are input into the multi-scale channel attention module. The specific fusion process of the two features is as follows:

[0014]

[0015] Among them, Z is the result of the fusion of two features, X and Y represent the two features that need to be fused with the initial features. is the initial fusion operation of element-by-element addition of feature maps, S represents the initial fusion result of X and Y, M(·) represents the multi-scale channel attention module, S′ represents the output of the initial fusion feature after M(·), Represents the element-wise multiplication of two feature maps.

[0016] Furthermore, the spatial resolution of the upper feature map with a smaller resolution is upsampled to twice its original value, and the lower feature map with a larger resolution is reduced in channel dimension through a 1×1 convolution layer.

[0017] Furthermore, the specific processing process of the multi-scale channel attention module is as follows:

[0018] Step S31, for the input feature S with dimension C*H*W, first perform channel attention calculation of the local feature, and output L(S), where L(·) is the channel attention calculation formula of the local feature;

[0019] Step S32, for the input feature S with dimension C*H*W, perform channel attention calculation of the global feature, and output G(S), where G(·) is the channel attention calculation formula of the global feature;

[0020] Step S33, the final output S′ of feature S after the multi-scale channel attention module is:

[0021]

[0022] is a broadcast addition. Since the sizes of the two branches L(S) and G(S) are different, a broadcast operation must be performed here. σ is the Sigmoid activation function, which is used to calculate the weight value after the attention mechanism is processed. The two feature maps are element-wise multiplied, and the input attention map and the obtained attention weight value are element-wise multiplied to obtain the output feature.

[0023] Furthermore, the specific definition of the channel attention calculation of local features is:

[0024] L(S)=B(PWConv2(δ(B(PWConv1(S)))))

[0025] B is the Batch Normalization operation, PWConv2 is the convolution kernel size The convolution operation is , δ is the Relu activation function, and PWConv1 is the convolution kernel size. The convolution operation.

[0026] Furthermore, the specific definition of the channel attention calculation of the global feature is:

[0027] G(S)=B(PWConv2(δ(B(PWConv1(g(S))))))

[0028] B is the Batch Normalization operation, PWConv2 is the convolution kernel size The convolution operation is , δ is the Relu activation function, and PWConv1 is the convolution kernel size. The convolution operation of G(·) is a global average pooling operation. The difference from L(S) is that G(S) first performs a global average pooling operation on the input feature M. After the operation, the size of the resulting feature map is 1×1, which effectively retains the overall features and global spatial information of the original input. Because the size of the local channel attention output is H×W, the addition of the two requires a broadcast operation.

[0029] Furthermore, ResNet-50 is selected as the backbone network.

[0030] The present invention also provides a multi-scale feature fusion star map recognition device based on an attention mechanism, comprising the following modules:

[0031] The dataset construction module is used to construct a basic training set of star image and label it;

[0032] Data enhancement module, used to enhance the data of the basic training set;

[0033] A model building module is used to build a complete multi-scale feature fusion model based on the attention mechanism. The multi-scale feature fusion model first extracts features of different scales through the backbone network, and then fuses adjacent feature maps from top to bottom according to the network depth through a multi-scale feature fusion method based on the attention mechanism. Two features are fused each time, and the resulting new feature map continues to repeat the fusion process with the lower-layer features until the iteration ends, and the final fused features from top to bottom are input into the classifier;

[0034] The model training module is used to batch-input the training set obtained by data enhancement into the network model constructed above to train and generate a star map recognition model;

[0035] The star map recognition module is used to input the star map to be recognized into the generated star map recognition model. The star map recognition model outputs the recognition result. The classification result output by each star map is the main star number. Finally, the remaining stars in the star map are identified according to the angular distance between the main star and the adjacent stars in the star map.

[0036] The present invention also provides a multi-scale feature fusion star map recognition device based on an attention mechanism, comprising:

[0037] one or more processors;

[0038] A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement a multi-scale feature fusion star map recognition method based on the attention mechanism as described in the above scheme.

[0039] The following beneficial effects are achieved by implementing the present invention:

[0040] The present invention provides a method for star map recognition based on a multi-scale feature fusion of an attention mechanism. In the process of star map recognition, the method first extracts multi-scale feature maps of different convolution stages, and then performs multi-scale feature fusion based on an attention mechanism on the extracted multi-scale feature maps layer by layer from top to bottom, so that the entire model can fully utilize multi-scale features and improve the star map recognition performance in real scenes. Through multi-scale pyramid feature extraction and multi-scale feature fusion based on an attention mechanism, the recognition accuracy of real star maps in a complex real deep space environment can be improved. This can improve the accuracy and reliability of attitude positioning and orbit control of aircrafts achieved through star sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a general structure diagram of the star image recognition model based on the deep convolutional neural network method.

[0042] Figure 2 It is a multi-scale channel attention module of an example of the present invention.

[0043] Figure 3 It is a multi-scale feature fusion module based on the attention mechanism of an embodiment of the present invention.

[0044] Figure 4 This is a multi-scale feature fusion star map recognition model based on the attention mechanism of an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The present invention extracts the feature hierarchical pyramid generated in the feedforward process of the convolutional neural network, and uses a multi-scale feature fusion method based on the attention mechanism to achieve feature fusion, thereby improving the star map recognition accuracy of the deep convolutional neural network in real scenes and further improving the performance of the star sensor in the deep space detection process. Specifically, Figure 1 As shown in the figure, when the traditional deep learning-based convolutional neural network classification method is used to complete the star map recognition task, for the input star map image, only the strong semantic features of the last layer of the neural network are used for recognition. This method has a good recognition effect on the simulated ideal simulated star map with uniform distribution, but it abandons the use of detailed information of shallow neural networks and ignores the fact that the distribution of stars in the deep space environment may present multiple scales. It is difficult to achieve ideal recognition effects for star maps of different distribution scales generated in real scenes.

[0047] Therefore, the key to the invention is to extract multi-scale feature maps in a bottom-up manner during the feedforward calculation process of the backbone deep convolutional neural network framework to form a multi-scale feature hierarchical pyramid structure. The deep convolutional neural network downsamples the feature hierarchy to give it an inherent multi-scale pyramid shape. This network feature hierarchy produces feature maps of different spatial resolutions, but these feature maps have semantic differences of different depths. Therefore, the present invention also uses a multi-scale feature fusion module based on an attention mechanism to replace the simple linear fusion operation to fuse the extracted pyramid features layer by layer. This module aggregates multi-scale contextual information along the channel dimension during the channel attention mechanism process to better fuse semantically and scale-inconsistent features. Based on the above operations, the resulting network model can simultaneously emphasize star clusters with a more global distribution and star clusters with a more local distribution, making it convenient for the network to identify star maps under extreme scale changes.

[0048] refer to Figure 4 ,The multi-scale feature fusion star map recognition method based on attention mechanism, includes the following steps:

[0049] Step 1: Get an initial training set of star images of size 224×224 based on the basic star catalog. The corresponding classification label of each star image is the number of the main star in the image.

[0050] In this embodiment, the basic star catalog can also be cut into other sizes, but the best size is 224×224. If the output feature map of the classification task is too small, the information will be lost too seriously. If the size is too large, the abstract level of information is not high enough and the amount of calculation is also greater. The size of 7x7 is a good balance. When an image is reduced from a large resolution to a small resolution, the reduction factor is usually an exponential power of 2, so the input of the image must be an exponential power of 7x2. In most classification data sets represented by ImageNet, the length and width of the image are around 300 resolution. Therefore, we need to find an exponential power of 7x2 and use it as input at a resolution of around 300, where 7x2 to the power of 4 = 7x16 = 112, 7x2 to the power of 5 = 7x32 = 224, 7x2 to the power of 6 = 448, and the closest to 300 is 224.

[0051] Step 2: Perform data enhancement on the initial training set of the star map. In deep learning visual tasks, data enhancement is often used to improve the robustness of the model and reduce the model's sensitivity to the image. In the embodiment of the present invention, a random position deviation from -5 pixels to 5 pixels is added to the coordinates of each star in the star map to improve the robustness of the model to position errors; a random magnitude deviation from -0.5Mv to 0.5Mv is added to each star to improve the robustness of the model to magnitude errors; 1 to 5 fake stars with random positions and magnitudes are added to each basic star map image in the training set to improve the robustness of the model to fake star interference. In addition, since deep convolutional neural networks lack rotational invariance, this method rotates the basic star map image in the training set from 0° to 360°, with each rotation of 5°, to improve the rotational invariance of the algorithm.

[0052] Step 3: Construct a multi-scale channel attention module. The module construction process is as follows Figure 2 As shown, for an input feature S with a dimension of C*H*W, where C is the number of channels, H and W are the height and width respectively, the channel attention calculation of the local feature is first performed, which can be specifically defined as:

[0053] L(S)=B(PWConv2(δ(B(PWConv1(S)))))

[0054] B is the Batch Normalization operation, and PWConv2 is the convolution kernel size. The convolution operation is , δ is the Relu activation function, and PWConv1 is the convolution kernel size. The convolution operation. L(·) is the channel attention calculation formula of the local feature. After this operation, the resulting feature still has the same shape as the input feature, which can retain and highlight the subtle details in the underlying features.

[0055] For the input feature S with dimension C*H*W, it is necessary to perform channel attention calculation of the global feature, which can be defined as:

[0056] G(S)=B(PWConv2(δ(B(PWConv1(g(S))))))

[0057] The B, PWConv2, δ, and PWConv1 have the same meanings as those in the channel attention calculation of local features, and g(·) is the global average pooling operation. G(·) is the channel attention calculation formula for global features. The difference from L(S) is that G(S) first performs a global average pooling operation (GAP) on the input feature M. After this operation, the resulting feature map size is 1×1, which effectively retains the overall features and global spatial information of the original input. Because the local channel attention output size is H×W, the addition of the two requires a broadcast operation.

[0058] The output S′ of feature S after the multi-scale channel attention module is:

[0059]

[0060] Said is a broadcast addition. Since the sizes of the two branches L(S) and G(S) are different, a broadcast operation must be performed here. σ is the Sigmoid activation function, which is used to calculate the weight value after the attention mechanism is processed. The two feature maps are element-wise multiplied, and the input attention map and the obtained attention weight value are element-wise multiplied to obtain the output feature.

[0061] Step 4: Implement multi-scale feature fusion of adjacent feature maps based on the attention mechanism. For the extracted adjacent feature maps, upsample the spatial resolution of the upper feature map with a smaller resolution to twice the original resolution, and reduce the channel dimension of the lower feature map with a larger resolution through a 1×1 convolution layer. The two processed features are added point by point for initial feature fusion.

[0062] The features obtained by initial fusion are input into the multi-scale channel attention module constructed in step 3. The specific fusion process of the two features is as follows Figure 3 As shown, it means as follows:

[0063]

[0064] Z is the result of the fusion of two features, X and Y represent the two features that need to be fused with the initial features. is the initial fusion operation of element-by-element addition of feature maps, S represents the initial fusion result of X and Y. M(·) represents the multi-scale channel attention module constructed in step 3, S′ represents the output of the initial fusion feature after M(·), Represents the element-wise multiplication of two feature maps.

[0065] Step 5: Build a complete multi-scale feature fusion model based on attention mechanism, such as Figure 4As shown. First, select a suitable backbone network. In the example of the present invention, ResNet-50 is selected. In order to fully realize the multi-scale feature utilization in the model working process, the example of the present invention extracts the features output by the last residual block of each Stage in the ResNet-50 feedforward calculation process. The features of the four Stages are recorded as C2, C3, C4, and C5 respectively.

[0066] The multi-scale feature fusion method based on the attention mechanism of the adjacent feature maps in step 4 is used to fuse the adjacent feature maps from top to bottom according to the network depth. Two features are fused each time, and the new feature map continues to repeat the fusion process with the lower layer features until the iteration ends, and the final fused features from top to bottom are input into the classifier. For example, C2 and C3 are fused to obtain C23, and C23 is further fused with the lower layer C4, and the new features obtained are fused with C5 to obtain the final features.

[0067] Step 6: Batch input the training set obtained by data augmentation into the above constructed network model for feedforward feature pyramid extraction and back propagation training to train and generate a star map recognition model.

[0068] Step 7: Input the star map to be identified into the generated star map recognition model. The star map recognition model outputs the recognition result. The classification result output by each star map is the main star number. Finally, the remaining stars in the star map are identified according to the angular distance between the main star and the adjacent stars in the star map.

[0069] The embodiment of the present invention also provides a multi-scale feature fusion star map recognition device based on an attention mechanism, comprising the following modules:

[0070] The dataset construction module is used to construct a basic training set of star image and label it;

[0071] Data enhancement module, used to enhance the data of the basic training set;

[0072] A model building module is used to build a complete multi-scale feature fusion model based on the attention mechanism. The multi-scale feature fusion model first extracts features of different scales through the backbone network, and then fuses adjacent feature maps from top to bottom according to the network depth through a multi-scale feature fusion method based on the attention mechanism. Two features are fused each time, and the resulting new feature map continues to repeat the fusion process with the lower-layer features until the iteration ends, and the final fused features from top to bottom are input into the classifier;

[0073] The model training module is used to batch-input the training set obtained by data enhancement into the network model constructed above to train and generate a star map recognition model;

[0074] The star map recognition module is used to input the star map to be recognized into the generated star map recognition model. The star map recognition model outputs the recognition result. The classification result output by each star map is the main star number. Finally, the remaining stars in the star map are identified according to the angular distance between the main star and the adjacent stars in the star map.

[0075] It should be noted that the above-mentioned device embodiment corresponds to the method embodiment of the present invention, which can implement a multi-scale feature fusion star map recognition method based on the attention mechanism described in any method embodiment of the present invention. After comparative ablation experiments, it is found that the star map recognition accuracy of this algorithm exceeds that of traditional optimized grid algorithms, pyramid algorithms, improved logarithmic polar coordinate transformation algorithms, etc., and reaches the best in the real star map image recognition test process.

[0076] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the device provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative work.

[0077] The embodiment of the present invention further provides a multi-scale feature fusion star map recognition device based on an attention mechanism, comprising:

[0078] one or more processors;

[0079] A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement a multi-scale feature fusion star map recognition method based on the attention mechanism described in the above technical solution.

[0080] The device for implementing the multi-scale feature fusion star map recognition based on the attention mechanism may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The device for implementing the multi-scale feature fusion star map recognition based on the attention mechanism may include, but is not limited to, a processor and a memory.

[0081] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-scale feature fusion star map recognition method based on attention mechanism, characterized in that: The steps include: Step S1, constructing a basic training set of star map images and marking them; Step S2, perform data enhancement on the basic training set; Step S3, construct a complete multi-scale feature fusion model based on the attention mechanism, select ResNet-50 as the backbone network, extract the features output by the last residual block of each stage in the ResNet-50 feedforward calculation process, and then use the multi-scale feature fusion method based on the attention mechanism to fuse adjacent feature maps from top to bottom according to the network depth. Each time two features are fused, the resulting new feature map continues to repeat the fusion process with the lower layer features until the iteration ends, and the final fused features from top to bottom are input into the classifier; The specific implementation of the multi-scale feature fusion method based on the attention mechanism in step S3 is: For the extracted adjacent feature maps, upsampling the spatial resolution of the upper feature map with a smaller resolution, reducing the channel dimension of the lower feature map with a larger resolution, and adding the two processed features point by point to perform initial feature fusion; The features obtained by initial fusion are input into the multi-scale channel attention module. The specific fusion process of the two features is as follows: Among them, Z is the result of the fusion of two features, X and Y represent the two features that need to be initially fused. is the initial fusion operation of element-by-element addition of feature maps, S represents the initial fusion result of X and Y, M(·) represents the multi-scale channel attention module, S′ represents the output of the initial fusion feature after M(·), Represents the element-by-element multiplication of two feature maps; Step S4, batch-inputting the training set obtained by data enhancement into the above-constructed network model to train and generate a star map recognition model; Step S5, input the star map to be identified into the generated star map identification model, the star map identification model outputs the identification result, the classification result output by each star map is the main star number, and finally the remaining stars in the star map are identified according to the angular distance between the main star and the adjacent stars in the star map.

2. A multi-scale feature fusion star map recognition method based on attention mechanism according to claim 1, characterized in that: The data enhancement in step S2 includes adding a certain range of random position deviations to the coordinates of each star in the star map to improve the robustness of the model to position errors; adding a certain range of random magnitude deviations to each star to improve the robustness of the model to magnitude errors; adding a number of fake stars with random positions and magnitudes to each basic star map image in the training set to improve the robustness of the model to fake star interference; in addition, the basic star map image in the training set is rotated from 0° to 360°, each time rotating a certain angle, to improve the rotation invariance of the model.

3. The multi-scale feature fusion star map recognition method based on the attention mechanism according to claim 1, characterized in that: The spatial resolution of the upper feature map with a smaller resolution is upsampled to twice the original resolution, and the lower feature map with a larger resolution is reduced in channel dimension through a 1×1 convolution layer.

4. The multi-scale feature fusion star map recognition method based on the attention mechanism according to claim 1, characterized in that: The specific processing process of the multi-scale channel attention module is as follows: Step S31, for the input feature S with dimension C*H*W, first perform channel attention calculation of the local feature, and output L(S), where L(·) is the channel attention calculation formula of the local feature, where C is the number of channels, H and W are height and width respectively; Step S32, for the input feature S with dimension C*H*W, perform channel attention calculation of the global feature, and output G(S), where G(·) is the channel attention calculation formula of the global feature; Step S33, the final output S′ of feature S after the multi-scale channel attention module is: is a broadcast addition. Since the sizes of the two branches L(S) and G(S) are different, a broadcast operation must be performed here. σ is the Sigmoid activation function, which is used to calculate the weight value after the attention mechanism is processed. The two feature maps are element-wise multiplied, and the input attention map and the obtained attention weight value are element-wise multiplied to obtain the output feature.

5. A multi-scale feature fusion star map recognition method based on attention mechanism according to claim 4, characterized in that: The specific definition of the channel attention calculation of local features is: L(S)=B(PWConv2(δ(B(PWConv1(S))))) B is the Batch Normalization operation, PWConv2 is the convolution kernel size The convolution operation is , δ is the Relu activation function, and PWConv1 is the convolution kernel size. The convolution operation.

6. A multi-scale feature fusion star map recognition method based on attention mechanism according to claim 4, characterized in that: The specific definition of the channel attention calculation of the global feature is: G(S)=B(PWConv2(δ(B(PWConv1(g(S)))))) B is the Batch Normalization operation, PWConv2 is the convolution kernel size The convolution operation is , δ is the Relu activation function, and PWConv1 is the convolution kernel size. The convolution operation of G(·) is a global average pooling operation. The difference from L(S) is that G(S) first performs a global average pooling operation on the input feature M. After the operation is completed, the size of the resulting feature map is 1×1, which effectively retains the overall features and global spatial information of the original input. Because the size of the local channel attention output is H×W, the addition of the two requires a broadcast operation.

7. A multi-scale feature fusion star map recognition device based on attention mechanism, characterized in that: Includes the following modules: The dataset construction module is used to construct a basic training set of star image and label it; Data enhancement module, used to enhance the data of the basic training set; The model building module is used to build a complete multi-scale feature fusion model based on the attention mechanism. ResNet-50 is selected as the backbone network, and the features output by the last residual block of each stage in the ResNet-50 feedforward calculation process are extracted. Then, the multi-scale feature fusion method based on the attention mechanism is used to fuse adjacent feature maps from top to bottom according to the network depth. Two features are fused each time, and the resulting new feature map continues to repeat the fusion process with the lower-layer features until the iteration ends, and the final fused features from top to bottom are input into the classifier; The specific implementation of the multi-scale feature fusion method based on the attention mechanism is: For the extracted adjacent feature maps, upsampling the spatial resolution of the upper feature map with a smaller resolution, reducing the channel dimension of the lower feature map with a larger resolution, and adding the two processed features point by point to perform initial feature fusion; The features obtained by initial fusion are input into the multi-scale channel attention module. The specific fusion process of the two features is as follows: Among them, Z is the result of the fusion of two features, X and Y represent the two features that need to be initially fused. is the initial fusion operation of element-by-element addition of feature maps, S represents the initial fusion result of X and Y, M(·) represents the multi-scale channel attention module, S′ represents the output of the initial fusion feature after M(·), Represents the element-by-element multiplication of two feature maps; The model training module is used to batch-input the training set obtained by data enhancement into the network model constructed above to train and generate a star map recognition model; The star map recognition module is used to input the star map to be recognized into the generated star map recognition model. The star map recognition model outputs the recognition result. The classification result output by each star map is the main star number. Finally, the remaining stars in the star map are identified according to the angular distance between the main star and the adjacent stars in the star map.

8. A multi-scale feature fusion star map recognition device based on attention mechanism, characterized in that: include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, enables the one or more processors to implement a multi-scale feature fusion star map recognition method based on an attention mechanism as described in any one of claims 1 to 6.

Images