Multi-dimensional feature fusion recognition method for space faint target

Abstract

The invention discloses a spatial faint target multi-dimensional feature fusion recognition method, which relates to the technical field of target recognition, and comprises the following steps of: constructing an integrated classifier by using a ROCKET architecture, an LSTM architecture and an OVO-SVM, constructing an initial basic probability distribution function by using spatial target multi-dimensional features as input quantities, and modifying an initial BPA by introducing a contraction-expansion function to obtain a spatial faint target multi-dimensional feature fusion recognition result; the multi-dimensional feature fusion recognition of the spatial faint target is completed by weakening the conflict degree between the features, performing discount, performing DST fusion operation, obtaining the fused BPA, combining the category discrimination criterion and outputting the target category, so that the problem that the recognition effect is poor due to the fact that the spatial faint target depends on a single feature for recognition can be effectively solved. The method improves the recognition accuracy of the space faint target, embodies the obvious advantages of multi-dimensional feature fusion in the aspect of improving the recognition performance, and promotes the development of a space-based infrared faint target recognition technology.

Description

Technical Field

[0001] The present invention relates to the field of target recognition technology, and in particular to a multi-dimensional feature fusion recognition method for spatial dim targets. Background Art

[0002] In recent years, aerospace technology in various countries has made great strides, and more and more space targets have been sent into orbit. On the one hand, these space targets, such as satellites and space stations, have made outstanding contributions in the fields of civil, commercial, and scientific research; on the other hand, these space activities will produce a large number of harmful space debris (such as rocket separation debris, out-of-control satellites, satellite fragments, etc.), most of which remain in orbit, making the space environment increasingly severe. Due to the long distance between the space target and the imaging surface of the infrared detector, the radiation at the entrance pupil of the detector is relatively weak, and usually only occupies one or several pixels on the infrared image, lacking information such as the shape, posture, and texture of the target. These targets are also called dark targets. Therefore, how to accurately identify the types of dark space targets and then take countermeasures is an urgent issue that needs to be considered.

[0003] The premise of identifying dim targets is to understand and obtain the difference characteristics between targets. However, since the infrared sensor can only obtain the infrared radiation time series of the target and the movement trajectory of the corresponding pixel points of the target, there is less information available, which makes it difficult to identify and distinguish the target. Therefore, it is necessary to use this known information to fully understand and explore the difference characteristics between space targets, and combine intelligent perception algorithms such as machine learning to achieve target classification and identification.

[0004] In the paper "Spatial Point Target Recognition Method Based on Multi-Sensor Data Fusion Combining Subjective Bayes Method and Neural Network" published by Li Hong in Journal of Infrared and Millimeter Waves 1997(06):49-55, the infrared radiation spectrum data of spatial point targets was used to extract the radiation flux and its change rate of spatial point targets in three different bands as effective features, and the spatial point target recognition was realized by combining Bayes and neural networks; Deng Qiuqun's paper "Research on Spatial Infrared Target Recognition Method Based on Convolutional Neural Network" published by National University of Defense Technology in 2019 used the infrared radiation of spatial targets as recognition features, combined with multi-scale convolutional neural network CNN and recursive Figure 1Deep neural network models such as convolutional neural networks (Recurrence Plot CNN, RPCNN) can be used to identify dim targets in space. These features are mostly concentrated in one or two categories of features, with a small number of features. Such single features are greatly affected by the environment and have poor stability. In addition, some space targets can use camouflage technology on some features, leading to misidentification. Multi-dimensional feature fusion recognition can avoid the shortcomings of single feature recognition and fully integrate the acquired feature information to form a complete, reasonable, and consistent perception, thereby achieving more accurate recognition.

[0005] Based on the above analysis, due to the small amount of available information for dim targets in space, misidentification often occurs when relying solely on a single feature for recognition. Therefore, it is urgent to clarify and realize how to use multi-dimensional features to process dim targets in space and give full play to the advantages of multi-dimensional feature joint recognition. It is urgent to build a suitable multi-dimensional feature fusion recognition method. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a method for multi-dimensional feature fusion recognition of spatial dim targets, comprising the following steps:

[0007] S1. Establish a dataset of infrared characteristics of space dim targets and divide it into a training set and a test set.

[0008] S2. Establish an integrated classifier, input the multi-dimensional target features into the integrated classifier, obtain the initial recognition results, and construct the initial basic probability distribution function;

[0009] S3, introducing the contraction-expansion function, modifying the initial basic probability distribution function, and then performing discount and DST fusion operations to obtain the fused basic probability distribution function;

[0010] S4. Combined with the category discrimination criteria, the target type is output to complete the multi-dimensional feature fusion recognition of spatial dim targets.

[0011] The technical solution further defined in the present invention is:

[0012] Further, in step S1, the infrared feature data set of the space dim target includes the shape of the space target, the infrared data band, the type of the space target, the type of the lighting environment of the space target flight scene, and the characteristics of the space target;

[0013] Among them, the number of space target shapes is ≥5, including flat-bottomed cones, cylindrical-bottomed cones, spherical-bottomed cones, cylinders and spheres; the number of infrared data bands is ≥2, including 8-12μm and 6-7μm; the number of space target types is ≥5, including mother cabin, real target, light decoy type one, light decoy type two and heavy decoy; the type of space target flight scene lighting environment is ≥3, including full illumination, full shadow and alternating illumination / shadow; the number of space target features is ≥6, including long-wave radiation intensity, medium- and long-wave radiation intensity, temperature, emissivity-area product, micro-motion period and speed.

[0014] As described above, in the multi-dimensional feature fusion recognition method for spatial dim targets, in step S2, the integrated classifier is composed of three classifiers: ROCKET architecture, LSTM architecture, and OVO-SVM;

[0015] The ROCKET architecture initializes a series of random convolution kernels, which are determined by five parameters: length, weight, bias, expansion, and padding. All convolution kernels are convolved with the original time series to produce a series of feature maps. Then, the positive ratio PPV and the maximum value are extracted and used as the representative of each feature map. Finally, the classifier is trained by the vector formed by the positive ratio PPV and the maximum value. The positive ratio PPV is used to characterize the proportion of the input sequence that matches the given sequence. The calculation method is shown in the following formula:

[0016]

[0017] Where L represents the length of each time series. represents the jth value of the i-th feature sequence; Indicates when hour,

[0018] The LSTM architecture consists of an input layer, LSTM layer, fully connected layer, and Softmax layer;

[0019] The OVO-SVM classifier uses the sigmoid fitting method to map the output to the posterior probability corresponding to the category:

[0020]

[0021] Among them, P represents the posterior probability value; f represents the standard output of SVM; A and B represent parameter values, which are obtained by calculating the minimum negative log-likelihood value of the parameter set.

[0022] As described above, a multi-dimensional feature fusion recognition method for space dim targets, in step S2, the multi-dimensional target features are input into the integrated classifier, and the specific method for obtaining the initial recognition result is: the long-wave radiation intensity features and the medium- and long-wave radiation intensity features are respectively input into the ROCKET architecture, the temperature features, the emissivity area features and the speed features are respectively input into the LSTM architecture, and the micro-motion period features are input into the OVO-SVM classifier. The classifier outputs the recognition probability value p = {p1, p2, ..., p M}, where M represents the total number of space target types.

[0023] As described above, in the multi-dimensional feature fusion recognition method for space dim targets, in step S2, the specific method for constructing the initial basic probability allocation function is: the recognition probability value of each type of space target output by the classifier according to each type of feature is used to form the initial basic probability allocation function

[0024]

[0025] Among them, p i· It represents the probability value of the classifier's recognition output based on the i-th feature, 1≤i≤N; M represents the total number of spatial targets, and N represents the number of target feature types used for recognition.

[0026] As described above, in a multi-dimensional feature fusion recognition method for a space dim target, in step S3, a shrink-expansion function is introduced to modify the initial basic probability allocation function. The specific method is as follows: in space target recognition, the sum of the basic probability allocation function of each feature is 1, and the average basic probability allocation function of each category is 1 / M. 1 / M is used as a threshold for scaling the basic probability allocation function value, and a shrink-expansion function is introduced to modify the basic basic probability allocation function value, as shown below:

[0027]

[0028] Among them, pi j It represents the output recognition probability value of the classifier for the jth target based on the ith feature. When the basic probability distribution function value corresponding to a certain category of the target is less than or equal to 1 / M, the basic probability distribution function value is shrunk. When the basic probability distribution function value corresponding to a certain category of the target is greater than 1 / M, the basic probability distribution function value is amplified.

[0029] As described above, in the multi-dimensional feature fusion recognition method for space dim targets, in step S3, discount and DST fusion operations are performed to obtain the fused basic probability distribution function, which specifically includes the following sub-steps:

[0030] S3.1. Normalize the modified basic probability distribution function value:

[0031]

[0032] S3.2, use the accuracy of each classifier to discount the basic probability distribution function and assign the uncertainty of the classifier to the overall recognition framework;

[0033] S3.3. Use the DST combination rule to fuse the basic probability distribution function value after discount processing. The basic probability distribution function value after fusion is:

[0034]

[0035] Among them, mm DS Represents the basic probability distribution function value after fusion.

[0036] As described above, in the multi-dimensional feature fusion recognition method for space dim targets, in step S3.2, the discount processing of the basic probability distribution function corresponding to each classifier is as follows:

[0037]

[0038] in, Represents the recognition accuracy of the classifier; m iΘ Indicates that the recognition result belongs to the basic probability distribution function value of the overall recognition framework, which is used to reflect the uncertainty of the corresponding classifier; m iM It represents the discounted basic probability distribution function value that the target belongs to class M when the target is identified based on the current features.

[0039] As described above, in the multi-dimensional feature fusion recognition method for a space dim target, in step S4, the basic probability distribution function after fusion in step S3 is m DS =(p1, p2, ..., p m , p Θ ), p Θ Represents the fusion probability value of the overall framework; assuming that the target types K1, K2∈Θ, where K1 satisfies K2 Satisfaction When the following conditions are met:

[0040]

[0041] Then the recognition type of the target to be detected is M1, where ∈1 and ∈2 represent thresholds.

[0042] As described above, a multi-dimensional feature fusion recognition method for spatial dim targets is used. In step S2, the number of random convolution kernels of the ROCKET architecture is set to 10000; the LSTM architecture includes an input layer, an LSTM layer with 200 hidden units, a fully connected layer, and a softmax layer. In the training phase, the Adam optimizer is used, the learning rate is 0.001, the smoothing constant in the first-order momentum is 0.9, the smoothing constant in the second-order momentum is 0.999, and the offset of the denominator in the gradient update relationship is 10 -8 , the maximum epoch is 120, and the loss function is cross entropy, which is used to evaluate whether the parameters learned by the current network are optimal; the kernel function of the OVO-SVM classifier is the radial basis function, the radial basis function parameter σ is 0.5, and the penalty factor is 1.

[0043] The beneficial effects of the present invention are:

[0044] In the present invention, an integrated classifier is constructed with the ROCKET architecture, the LSTM architecture and the OVO-SVM, the multi-dimensional features of the space target are used as input, an initial basic probability distribution function is constructed, a shrink-expand function is introduced to modify the initial BPA, the degree of conflict between features is reduced, discount and DST fusion operations are performed, the fused BPA is obtained, and the category discrimination criterion is combined to output the target type, and the multi-dimensional feature fusion recognition of the space dim target is completed, so that the problem of poor recognition effect of the space dim target relying on a single feature for recognition can be effectively solved, the recognition accuracy of the space dim target is improved, the obvious advantages of multi-dimensional feature fusion in improving recognition performance are reflected, and the development of space-based infrared dim target recognition technology is promoted. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0046] Figure 2 Schematic diagram of changes in long-wave radiation intensity information of five types of space targets in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram comparing the recognition performance of the method of the present invention and the comparative method at different noise levels. DETAILED DESCRIPTION

[0048] This embodiment provides a method for multi-dimensional feature fusion recognition of space dim targets, such as Figure 1 As shown, the following steps are included:

[0049] S1. Establish a dataset of infrared characteristics of faint space targets and randomly divide it into a training set and a test set in a ratio of 7:3. The dataset includes the shape of space targets, infrared data band, type of space targets, type of lighting environment of space target flight scenes and characteristics of space targets.

[0050] There are five shapes of space targets, including flat-bottomed cone, cylindrical-bottomed cone, spherical-bottomed cone, cylinder, and sphere. The attribute parameters of the targets are shown in Table 1.

[0051] Table 1

[0052]

[0053] There are two infrared data bands, including 8-12μm and 6-7μm.

[0054] There are five types of space targets, including mother cabin, real target, light decoy type one, light decoy type two, and heavy decoy.

[0055] There are three types of lighting environments for space target flight scenes, including full illumination, full shadow, and alternating illumination / shadow. Each lighting environment has two sets of flight trajectories, for a total of six sets of flight trajectories. The flight trajectory scene information is shown in Table 2 below.

[0056] Table 2

[0057]

[0058] There are six characteristics of space targets, including long-wave radiation intensity, medium- and long-wave radiation intensity, temperature, emissivity-area product, micromotion period, and speed. Figure 2 As shown, the changes in long-wave radiation intensity information of five types of space targets are displayed.

[0059] S2. Establish an integrated classifier, input the multi-dimensional target features into the integrated classifier, obtain the initial recognition results, and construct the initial basic probability assignment function (BPA); the integrated classifier consists of three classifiers: ROCKET architecture, LSTM architecture and OVO-SVM.

[0060] In this embodiment, the number of random convolution kernels of the ROCKET architecture is set to 10000. The ROCKET architecture initializes a series of random convolution kernels, which are determined by five parameters: length, weight, bias, expansion, and padding. All convolution kernels are convolved with the original time series to generate a series of feature maps. Then, the positive ratio PPV and the maximum value are extracted and used as the representative of each feature map. Finally, the classifier is trained by the vector formed by the positive ratio PPV and the maximum value. The positive ratio PPV is used to characterize the proportion of the input sequence that matches the given sequence, and its calculation method is shown in the following formula:

[0061]

[0062] Where L represents the length of each time series. represents the jth value of the i-th feature sequence; Indicates when hour,

[0063] The LSTM architecture consists of an input layer, an LSTM layer with 200 hidden units, a fully connected layer, and a Softmax layer. During the training phase, the Adam optimizer is used, the learning rate is 0.001, the smoothing constant in the first-order momentum is 0.9, the smoothing constant in the second-order momentum is 0.999, and the offset of the denominator in the gradient update relation is 10 -8 , the maximum epoch is 120, and the loss function is cross entropy, which is used to evaluate whether the parameters learned by the current network are optimal.

[0064] The kernel function of the OVO-SVM classifier is the radial basis function (RBF), the parameter σ of the radial basis kernel function is 0.5, the penalty factor is 1, and the OVO-SVM classifier uses the sigmoid fitting method to map the output to the posterior probability corresponding to the category:

[0065]

[0066] Among them, P represents the posterior probability value; f represents the standard output of SVM; A and B represent parameter values, which are obtained by calculating the minimum negative log-likelihood value of the parameter set.

[0067] In step S2, the multidimensional target features are input into the integrated classifier. The specific method for obtaining the initial recognition result is as follows: the long-wave radiation intensity features and the medium- and long-wave radiation intensity features are respectively input into the ROCKET architecture, the temperature features, the emissivity area features and the speed features are respectively input into the LSTM architecture, and the micro-motion period features are input into the OVO-SVM classifier. The classifier outputs the recognition probability value p = {p1, p2, ..., p M}, where M represents the total number of space target types.

[0068] In step S2, the specific method of constructing the initial BPA is: the recognition probability value of each type of space target output by the classifier according to each type of feature is used to form the initial basic probability allocation function

[0069]

[0070] Among them, p i· It represents the recognition output probability value of the classifier according to the i-th feature, 1≤i≤N; M represents the total number of space target types, and N represents the number of target feature types used for recognition; in this embodiment, as shown in Table 3 below, it is a schematic table of the initial BPA value obtained after processing by the integrated classifier.

[0071] Table 3

[0072]

[0073] S3, introduce the shrink-expand function to modify the initial BPA, and then perform discount and Dempster-Shafer Theory of Evidence (DST) fusion operations to obtain the fused BPA.

[0074] In step S3, the shrink-expand function is introduced to modify the initial BPA. The specific method is as follows: when there is a conflict between DST evidences, the fusion recognition result will be contrary to common sense. In this embodiment, the evidence conflict problem is solved by introducing a shrink-expand function and evidence discount processing; in spatial target recognition, the sum of the BPA of each feature (evidence) is 1, and the average BPA of each category is 1 / M. In extreme cases, when the BPA value of all categories of a feature is 1 / M, the feature is unreliable and the target cannot be successfully identified.

[0075] Therefore, 1 / M is used as a threshold for scaling the BPA value, and a shrink-expand function is introduced to modify the basic BPA value as follows:

[0076]

[0077] Among them, p ij It represents the output recognition probability value of the classifier for the jth target based on the ith feature; when the BPA value corresponding to a certain category of the target is less than or equal to 1 / M, it is considered that the probability of the test sample belonging to this category is small, and the BPA value is shrunk; when the BPA value corresponding to a certain category of the target is greater than 1 / M, it is considered that the probability of the sample belonging to this category is large, and the BPA value is amplified; in this embodiment, since the number of target categories is 5, the threshold of the shrink-expand function is 0.2.

[0078] In step S3, the discount and DST fusion operation is performed to obtain the fused BPA, which specifically includes the following sub-steps:

[0079] S3.1. Normalize the modified BPA value:

[0080]

[0081] Among them, pi j Represents the output recognition probability value of the classifier for the jth target based on the i-th feature.

[0082] S3.2. Use the accuracy of each classifier to discount the BPA, and assign the uncertainty of the classifier to the overall recognition framework. The BPA discount processing corresponding to each classifier is shown as follows:

[0083]

[0084] in, Represents the recognition accuracy of the classifier; m iΘ Indicates that the recognition result belongs to the BPA value of the overall recognition framework, which is used to reflect the uncertainty of the corresponding classifier; m iM Indicates the discounted BPA value of the target belonging to category M when the target is identified based on the current features.

[0085] S3.3. Use the DST combination rule to merge the discounted BPA values. The merged BPA values ​​are:

[0086]

[0087] Among them, m DS Indicates the BPA value after fusion.

[0088] S4. Combined with the category discrimination criteria, the target type is output to complete the multi-dimensional feature fusion recognition of spatial dim targets.

[0089] According to the fused BPA in step S3, m DS =(p1, p2, ..., p m , pα ), p Θ Represents the fusion probability value of the overall framework; assuming that the target types K1, K2∈Θ, where K1 satisfies K2 Satisfaction When the following conditions are met:

[0090]

[0091] Then the recognition type of the target to be detected is M1, where ∈1 and ∈2 represent thresholds. In this embodiment, ∈1 and ∈2 are set to 0 and 0.10 respectively.

[0092] As shown in Table 4 below, it is a comparison table of target type recognition results between the method of this embodiment and the traditional DSTT method, Murphy method and Gao method. It can be seen from the table that the traditional DSTT method and Murphy method make mistakes in identifying the target type, and the Gao method and the method of this embodiment both correctly identify the target type. However, the result of the method of this embodiment has higher identifiability, which is more conducive to the control center to make reasonable judgments on the category of the target.

[0093] Table 4

[0094]

[0095] As shown in Table 5 below, it is a comparison table of recognition performance under different noise levels between the method of this embodiment and the recognition relying on single-dimensional features and the recognition relying on four-dimensional features (temperature, emissivity area product, micro-motion period, speed).

[0096] Table 5

[0097]

[0098] From Table 5, we can see that: (1) For single-dimensional feature recognition, at any noise level, the recognition accuracy of long-wave radiation intensity and medium- and long-wave radiation intensity is higher than that of other four-dimensional features; (2) For four-dimensional features (temperature, emissivity-area product, micromotion period, speed), the effect of any recognition relying on any one of these features is worse than that relying on radiation intensity; but at low signal-to-noise ratio, the recognition effect of integrating these four-dimensional features is better than that of relying on radiation intensity; (3) The recognition effect of integrating six-dimensional features is higher than that of integrating four-dimensional features. For example, for The recognition accuracy of the fused six-dimensional features is 93.33%, which is higher than 83.89%. This shows that in space target recognition, it is necessary to use multi-dimensional features to improve the effect of target recognition.

[0099] like Figure 3 As shown in the figure, the method of this embodiment and the comparative method are shown in different noise levels. The recognition performance comparison of The recognition effect of each method is improved to varying degrees; (2) Compared with other algorithms, the algorithm in this embodiment has better recognition effect and robustness. When , the recognition accuracy of the method in this embodiment is 93.33%, which is better than 81.11%, 88.89% and 90.00% of the other three algorithms.

[0100] The method of this embodiment uses the ROCKET architecture, LSTM architecture and OVO-SVM to construct an integrated classifier, uses the multi-dimensional features of space targets as input, constructs an initial basic probability allocation function, introduces a shrink-expand function to modify the initial BPA, reduces the degree of conflict between features, performs discount and DST fusion operations, obtains the fused BPA, combines the category discrimination criterion, outputs the target type, and completes the multi-dimensional feature fusion recognition of space dim targets, thereby effectively solving the problem of poor recognition effect of space dim targets relying on a single feature for recognition, improving the recognition accuracy of space dim targets, reflecting the obvious advantages of multi-dimensional feature fusion in improving recognition performance, and promoting the development of space-based infrared dim target recognition technology.

[0101] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A multi-dimensional feature fusion recognition method for spatial dim targets, characterized by: The following steps are involved: S1. Establish a dataset of infrared characteristics of space dim targets and divide it into a training set and a test set. S2. Establish an integrated classifier, input the multi-dimensional target features into the integrated classifier, obtain the initial recognition results, and construct the initial basic probability distribution function; S3, introducing the contraction-expansion function, modifying the initial basic probability distribution function, and then performing discount and DST fusion operations to obtain the fused basic probability distribution function; S4. Combined with the category discrimination criteria, the target type is output to complete the multi-dimensional feature fusion recognition of spatial dim targets.

2. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 1 is characterized by: In step S1, the infrared feature data set of the space dim target includes the shape of the space target, the infrared data band, the type of the space target, the type of the lighting environment of the space target flight scene, and the characteristics of the space target; The number of space target shapes is ≥5, including flat-bottomed cones, cylindrical-bottomed cones, spherical-bottomed cones, cylinders, and spheres; the number of infrared data bands is ≥2, including 8-12μm and 6-7μm; the number of space target types is ≥5, including mother cabin, real target, light decoy type 1, light decoy type 2, and heavy decoy; the type of space target flight scene lighting environment is ≥3, including full illumination, full shadow, and alternating illumination / shadow; The number of target characteristics is ≥ 6, including long-wave radiation intensity, medium- and long-wave radiation intensity, temperature, emissivity-area product, micromotion period, and speed.

3. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 2 is characterized by: In step S2, the integrated classifier is composed of three classifiers: ROCKET architecture, LSTM architecture and OVO-SVM; The ROCKET architecture initializes a series of random convolution kernels, which are determined by five parameters: length, weight, bias, dilation, and padding. All convolution kernels are convolved with the original time series to produce a series of feature maps. Then, the positive ratio PPV and the maximum value are extracted and used as the representative of each feature map. Finally, the classifier is trained by the vector formed by the positive ratio PPV and the maximum value; the positive ratio PPV is used to characterize the proportion of the input sequence that matches the given sequence, and its calculation method is shown in the following formula: Where L represents the length of each time series. represents the jth value of the i-th feature sequence; Indicates when hour, The LSTM architecture consists of an input layer, LSTM layer, fully connected layer, and Softmax layer; The OVO-SVM classifier uses the sigmoid fitting method to map the output to the posterior probability corresponding to the category: Among them, P represents the posterior probability value; f represents the standard output of SVM; A and B represent parameter values, which are obtained by calculating the minimum negative log-likelihood value of the parameter set.

4. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 3 is characterized by: In step S2, the multidimensional target features are input into the integrated classifier, and the specific method for obtaining the initial recognition result is: the long-wave radiation intensity features and the medium- and long-wave radiation intensity features are respectively input into the ROCKET architecture, the temperature features, the emissivity area features and the speed features are respectively input into the LSTM architecture, and the micro-motion period features are input into the OVO-SVM classifier. The classifier outputs the recognition probability value p = {p1, p2, ..., p M }, where M represents the total number of space target types.

5. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 4 is characterized by: In step S2, the specific method of constructing the initial basic probability distribution function is: the recognition probability value of each type of space target output by the classifier according to each type of feature is used to form the initial basic probability distribution function Among them, p i· It represents the probability value of the classifier's recognition output based on the i-th feature, 1≤i≤N; M represents the total number of spatial targets, and N represents the number of target feature types used for recognition.

6. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 1 is characterized by: In step S3, the specific method of introducing the shrink-expand function to modify the initial basic probability allocation function is as follows: in space target recognition, the sum of the basic probability allocation function of each feature is 1, and the average basic probability allocation function of each category is 1 / M. 1 / M is used as a threshold for scaling the basic probability allocation function value, and a shrink-expand function is introduced to modify the basic basic probability allocation function value, as shown below: Among them, p ij It represents the output recognition probability value of the classifier for the jth target based on the ith feature. When the basic probability distribution function value corresponding to a certain category of the target is less than or equal to 1 / M, the basic probability distribution function value is shrunk. When the basic probability distribution function value corresponding to a certain category of the target is greater than 1 / M, the basic probability distribution function value is amplified.

7. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 6 is characterized by: In step S3, performing the discount and DST fusion operation to obtain the fused basic probability distribution function specifically includes the following sub-steps: S3.

1. Normalize the modified basic probability distribution function value: S3.2, use the accuracy of each classifier to discount the basic probability distribution function and assign the uncertainty of the classifier to the overall recognition framework; S3.

3. Use the DST combination rule to fuse the basic probability distribution function value after discount processing. The basic probability distribution function value after fusion is: Among them, m DS Represents the basic probability distribution function value after fusion.

8. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 7 is characterized by: In step S3.2, the discount processing of the basic probability distribution function corresponding to each classifier is as follows: in, Represents the recognition accuracy of the classifier; m iΘ Indicates that the recognition result belongs to the basic probability distribution function value of the overall recognition framework, which is used to reflect the uncertainty of the corresponding classifier; m iM It represents the discounted basic probability distribution function value that the target belongs to class M when the target is identified based on the current features.

9. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 7 is characterized by: In step S4, the basic probability distribution function after fusion in step S3 is m DS =(p1, p2, ..., p m , p Θ ), p Θ Represents the fusion probability value of the overall framework; assuming that the target types K1, K2∈Θ, where K1 satisfies K2 Satisfaction When the following conditions are met: Then the recognition type of the target to be detected is M1, where ∈1 and ∈2 represent thresholds.

10. The multi-dimensional feature fusion recognition method for spatial dim targets according to claim 3 is characterized by: In step S2, the number of random convolution kernels of the ROCKET architecture is set to 10000; the LSTM architecture includes an input layer, an LSTM layer with 200 hidden units, a fully connected layer, and a softmax layer. In the training phase, the Adam optimizer is used, the learning rate is 0.001, the smoothing constant in the first-order momentum is 0.9, the smoothing constant in the second-order momentum is 0.999, and the offset of the denominator in the gradient update relation is 10 -8 , the maximum epoch is 120, and the loss function is cross entropy, which is used to evaluate whether the parameters learned by the current network are optimal; the kernel function of the OVO-SVM classifier is the radial basis function, the radial basis function parameter σ is 0.5, and the penalty factor is 1.

Images