Fuzzy integral-based multi-modal aircraft feature decision-making level fusion method

Abstract

The invention discloses a fuzzy integral-based multi-modal aircraft feature decision-making level fusion method. The method comprises the following steps of: acquiring different types of data of an aircraft target; and performing feature recognition according to the number of types, fusing the features, and finally recognizing the fused features to obtain a recognition result. According to the characteristic that attribute differences of time, space, frequency spectrum and the like among multi-modal aircraft features are large, different fuzzy densities are distributed to the multi-modal aircraft features, and after D-S evidence fusion, the fuzzy membership degree of feature combination is obtained through a fuzzy integration method. And finally, performing weighted summation on each feature combination according to the fuzzy measure to obtain a final fusion recognition result. Compared with a traditional method, the method has the advantages that the utilization capability of the fuzzy integral on correlation of different features and the combination capability of the D-S evidence theory on feature information are combined, so that the features of each single-source aircraft are fully utilized. Experiments show that the method can effectively improve the accuracy of target recognition.

Description

Technical Field

[0001] The invention relates to a multi-modal aircraft feature decision-level fusion method based on fuzzy integral, belonging to the technical fields of artificial intelligence, fuzzy logic and information fusion, and particularly to a technology for decision support in an uncertain environment. Background Art

[0002] With the rapid development of science and technology, detection, reconnaissance and identification technology has also developed rapidly. In recent years, detection information has shown a significant growth trend. The collected aircraft feature information also presents different characteristics because it comes from different platforms and sensors. At the same time, due to complex environment and noise factors, if this information is simply superimposed, it will cause a large amount of useless information to increase and useful information to be submerged. Therefore, a method of information fusion is proposed. Information fusion is divided into three levels from low to high: pixel-level fusion, feature-level fusion, and decision-level fusion. For multiple sensors, the time, space, spectrum and other characteristics of the detected target information are different. It is difficult to use pixel-level fusion and feature-level fusion, and many unimportant information does not need to be fused. Therefore, the decision-level fusion method is adopted to obtain the single-source recognition results and then fuse the obtained target recognition results to achieve the goal of reducing the difficulty of fusion and increasing the recognition accuracy.

[0003] There are many algorithms for decision-level fusion, such as voting method, Bayesian reasoning method, DS evidence theory and fuzzy integral method. Among them, DS evidence theory is an extension and extension of Bayesian theory. Its evidence comes from the results given by each sensor, and then the credibility of each target is obtained. Finally, the Dempster merging rule is used to merge each credibility into a unified result. However, the use of DS requires that all evidence is independent of each other. However, in reality, the target information carried by multi-source evidence often overlaps with each other, which cannot fully meet this requirement. Choquet fuzzy integral assigns different weights to information from different sources, thereby reflecting the relative importance of each information source in decision-making, and can effectively handle the interdependence between attributes. It has obvious advantages in processing fuzzy and uncertain information, but Choquet fuzzy integral has the problem of complex calculation process and only a small part of the information is fully utilized when making decisions on multi-source information fusion. Summary of the invention

[0004] In order to solve the problem of difficulty in fusing target features collected by different platforms and sensors in the target fusion recognition process, this paper proposes a multi-modal aircraft feature decision-level fusion method based on fuzzy integral. The multi-source aircraft features are fused at the decision level by combining Choquet fuzzy integral and DS evidence theory method.

[0005] The overall solution of this application is:

[0006] A multi-modal aircraft feature decision-level fusion method based on fuzzy integral, the fusion method comprising the following steps:

[0007] Different types of data are acquired for aircraft targets; features are identified according to the number of types, the features are fused, and finally the fused features are identified to obtain identification results.

[0008] Furthermore, the number of types is two, and the fusion method is:

[0009] Step 1-1: Acquire two different types of data on the aircraft target;

[0010] Step 1-2: Perform feature recognition on the two aircraft features respectively, and obtain two unit recognition results respectively;

[0011] Step 1-3: Perform DS evidence synthesis on the two unit recognition results;

[0012] Step 1-4: Obtain the fuzzy density of different aircraft features respectively;

[0013] Step 1-5: Solve for the correlation parameter λ of the two aircraft features;

[0014] Step 1-6: Calculate the fuzzy measure g after feature fusion λ ;

[0015] Step 1-7: Calculate the membership of the target through the normalized fuzzy integral.

[0016] Furthermore, in the steps 1-3, the two unit recognition results {m1(w1),m1(w2),…,m1(w k )},{m2(w1),m2(w2),…,m2(w k )} to synthesize evidence, as follows:

[0017]

[0018] The recognition result {m(w1),m(w2),…,m(w k )}.

[0019] Furthermore, in the steps 1-5, the correlation parameter λ represents the correlation degree between the two aircraft features; when -1<λ<0, it indicates that the attributes are redundantly correlated, and when 0<λ, it indicates that the attributes are complementary correlated. According to the formula

[0020]

[0021] Using the fuzzy density obtained in steps 1-4, solve for the associated parameter λ.

[0022] Furthermore, in the steps 1-6, the fuzzy measure represents the importance of the combination of different aircraft features. According to the calculation formula of the fuzzy measure

[0023]

[0024] Calculate the fuzzy measure g of the feature pairwise combination based on the fuzzy density and correlation parameter λ obtained in steps 1-4 and 1-5 λ ; The pairwise combinations are calculated in turn with other fuzzy measures or fuzzy densities according to the formula to obtain the fuzzy measure results of more feature combinations.

[0025] Furthermore, in step 1-7, according to the recognition results obtained in steps 1-2 and 1-3, and the fuzzy density and fuzzy measure obtained in steps 1-4 and 1-6, the membership of the target is calculated by normalized fuzzy integral, which is recorded as {M(w1),M(w2),…,M(w k )}:

[0026]

[0027] Among them, m max (w i )=max(m1(w i ),m2(w i )), g max is m max The corresponding fuzzy density is the fusion recognition result with the largest membership as the target.

[0028] Furthermore, the number of the types is three or more, and the fusion method is:

[0029] Step 2-1: Acquire various types of data on the aircraft target;

[0030] Step 2-2: Combining multiple aircraft features to form feature pairs;

[0031] Step 2-3: Fusion recognition of each feature pair;

[0032] Step 2-4: Obtain the fusion recognition results of multiple aircraft features through weighted summation.

[0033] Furthermore, in step 2-3, for the n aircraft features, they are combined into feature pairs without duplication, and recorded as {E1, E2}, {E1, E3}, ..., {E n-1 ,E n For each feature pair, the steps 1-2 to 1-6 are respectively used to perform two aircraft feature fusion identifications to obtain the target membership and fuzzy measurement of each feature pair, which are respectively denoted as M j,l={M j,l (w1),M j,l (w2),…,M j,l (w k )} and g λ-j,l , where j and l represent the characteristic subscripts of different aircrafts respectively.

[0034] Furthermore, in step 2-4, the target membership M of each feature obtained in step 2-3 is j,l ={M j,l (w1),M j,l (w2),…,M j,l (w k )} and its corresponding fuzzy measure g λ-j,l , the target membership of multiple aircraft feature fusion recognition is obtained by weighted summation, denoted as M = {M(w1), M(w2), …, M(w k )}:

[0035]

[0036] Among them, ∑M j,l *g λ-j,l ,∑g λ-j,l Represents the product of all target membership and fuzzy measure M i,j *g λ-i,j And all fuzzy measures g λ-j,l sum;

[0037] The fusion recognition result of multiple aircraft features is taken as the target corresponding to the category with the maximum value of each item in the target membership.

[0038] Compared with the prior art, the present invention has the following technical effects:

[0039] The present invention is applicable to the problem of decision-level fusion of multi-source aircraft features. By combining Choquet fuzzy integral and DS evidence theory, the decision-level fusion algorithm of the present invention can assign different weights to different features of the target according to their characteristics, and reflect the interdependence therein with fuzzy measurement theory, so as to make full use of the feature information of each single source in the subsequent fusion process to obtain a more accurate fusion recognition result. Experimental results show that the algorithm can effectively improve the accuracy of target recognition by fusing multi-source aircraft feature information compared with single-source aircraft features. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the algorithm flow chart of two aircraft feature fusion recognition schemes;

[0041] Figure 2 It is the algorithm flow chart of multiple aircraft feature fusion recognition scheme. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0043] The multi-modal aircraft feature decision-level fusion method based on fuzzy integral proposed in the present invention can solve the problem of difficulty in fusing target features collected by different platforms and sensors during target fusion recognition. The present invention designs two aircraft feature fusion recognition schemes and multiple aircraft feature fusion recognition schemes for different numbers of target aircraft features.

[0044] The following are the definitions involved in this application:

[0045] Definition 1: Let X = {x1, x2, …, x n} is a finite non-empty set, P(X) is the power set of X, and the set function g:P(X)→[0,1] satisfies the following two conditions:

[0046]

[0047] Then g(K)≤g(T). g is called a fuzzy measure on P(X).

[0048] Definition 2: For all A,B∈P(X), and λ is the correlation parameter. If the fuzzy measure g satisfies the following conditions:

[0049] g(A∪B)=g(A)+g(B)+λg(A)g(B) (3)

[0050] Then g is called λ-fuzzy measure. Usually fuzzy measure is not additive.

[0051] For only one element x i The subset of i =g({x i}) is the fuzzy density. If X={x1,x2,…,x n} is a finite non-empty set, and x i The fuzzy density function is g(x i ), then the λ-fuzzy measure has the following properties:

[0052]

[0053] When g(X) = 1, the above formula can be simplified to:

[0054]

[0055] Definition 3: Let h:X→[0,1] be a measurable function on X, then the Choquet fuzzy integral of h on the set X with respect to the fuzzy measure g is:

[0056]

[0057] Among them, (i) is the value of h(x (1) )≤h(x (1) )≤…≤h(x (n) ) sorted subscript, X i ={x (i) ,x (i+1) ,…,x (n)},h(x (0) )=0

[0058] Definition 4: Assume that there are m possible propositions, and let Ω = {w1,w2,…,w m} is a finite and complete set, and the elements in the set are mutually exclusive, then Ω can be called the identification framework of the proposition, and the set of propositions composed of all subsets of Ω is called the power set of Ω, denoted by 2 Ω , which is expressed as follows:

[0059]

[0060] Definition 5: Let Ω be the identification frame of the proposition. If there exists Ω To a mapping function m:2 on [0,1] Ω →[0,1], and the following conditions are met:

[0061]

[0062] Then m is called the basic trust assignment function on the recognition framework Ω. If m(A)>0, then the proposition A is called a focal element.

[0063] Definition 6: Assume that under the same recognition framework Ω, there are n sets of evidence {E1,E2,…,E n},{m1,m2,…,m n} is the corresponding basic trust allocation function, then the Dempster synthesis rule is:

[0064]

[0065]

[0066] Among them, K is the conflict coefficient, which describes the conflict between evidences. The larger K is, the greater the conflict between evidences. When K = 1, {m1, m2, ..., m n}Complete conflict, in this case, the Dempster combination rule cannot be used. The main advantage of the Dempster combination rule is that it is easy to calculate and satisfies the commutative law and the associative law. Therefore, when fusing multiple pieces of evidence, the evidence can be fused one by one, and the fusion order will not have any effect on the final fusion result, thus avoiding the problem of excessive calculation when fusing multiple pieces of evidence together.

[0067] The present application is a multi-modal aircraft feature decision-level fusion method based on fuzzy integral. Two aircraft feature fusion recognition schemes and multiple aircraft feature fusion recognition schemes are designed for different feature quantities of target aircraft. The present invention uses two aircraft feature fusion schemes and multiple aircraft feature fusion schemes respectively according to the number of aircraft features.

[0068] Among them, the specific algorithm flow of the two aircraft feature fusion recognition solutions is as follows:

[0069] Step 1: Acquisition of aircraft features: Acquisition of two different types of data (aircraft visible light image, aircraft SAR image, etc.) is performed on the aircraft target to obtain two different aircraft features of the same aircraft target, recorded as {E1, E2}. There are three types of target recognition results, and the recognition framework is set to Ω = {w1, w2, w3}.

[0070] Step 2: Single-source identification of aircraft features: Perform feature identification on the two aircraft features separately. The identification algorithm is recorded as {m1, m2}, and two single-source identification results {m1(w1), m1(w2), m1(w3)}, {m2(w1), m2(w2), m2(w3)} are obtained.

[0071] Step 3 D-S evidence synthesis: For the two single-source recognition results, evidence synthesis is performed according to Formula 9 to obtain the recognition result {m(w1), m(w2), m(w3)} combining the two features.

[0072] Step 4: Obtain the fuzzy density of each feature: Obtain the fuzzy density of different aircraft features respectively, and consider the accuracy of the single-source recognition experiment, the quality of the aircraft features, and the importance of the aircraft features in experience to obtain the fuzzy density {g1,g2,…,g n}(n is the total number of features), in order to determine the degree of trust in the recognition result for a certain aircraft feature.

[0073] Step 5: Solve the association parameter λ: The association parameter λ represents the degree of association between the two aircraft features. When -1<λ<0, it means that the attributes are redundantly associated, and when 0<λ, it means that the attributes are complementary. According to the calculation formula in Definition 2, use the fuzzy density obtained in Step 4 to solve the association parameter λ.

[0074] Step 6 Calculate the fuzzy measure g after feature fusion λ: The fuzzy measure represents the importance of the combination of different aircraft features, which corresponds to the fuzzy density of a single aircraft feature. According to the calculation formula of the fuzzy measure in Definition 2, the fuzzy density and the associated parameter λ obtained in steps 4 and 5 can be used to calculate the fuzzy measure g of the feature pairwise combination. λ By calculating the pairwise combinations with other fuzzy measures or fuzzy densities in sequence according to the formula, we can get the fuzzy measure results of more feature combinations. The final result has nothing to do with the calculation order.

[0075] Step 7 calculates the membership of the target through the normalized fuzzy integral: According to the recognition results obtained in steps 2 and 3, and the fuzzy density and fuzzy measure obtained in steps 4 and 6, the membership of the target is calculated through the normalized fuzzy integral, which is recorded as {M(w1), M(w2), M(w3)}:

[0076]

[0077] Here m max (w i )=max(m1(w i ),m2(w i )), g max is m max The corresponding blur density.

[0078] The fusion recognition result with the maximum membership as the target.

[0079] The specific algorithm flow of the multiple aircraft feature fusion recognition solution is as follows:

[0080] Step 1: Acquisition of aircraft features: Acquisition of different types of data (aircraft visible light images, aircraft SAR images, etc.) is performed on the aircraft target to obtain multiple different aircraft features of the same aircraft target, recorded as {E1, E2, …, E n There are three kinds of target recognition results, and the recognition framework is assumed to be Ω = {w1, w2, w3}.

[0081] Step 2 Aircraft feature combination: For n aircraft features, combine them into feature pairs without duplication, denoted as {E1,E2}, {E1,E3}, ..., {E n-1 ,E n}.

[0082] Step 3: Fusion recognition of feature pairs: For the feature pairs {E1, E2}, {E1, E3}, ..., {E n-1 ,E n}, perform two aircraft feature fusion identification on it respectively, which corresponds to the two aircraft feature fusion steps 2 to 6, and calculate the target membership and fuzzy measurement of each feature, which are respectively recorded as M i,j ={Mi,j (w1),M i,j (w2),M i,j (w3)} and g λ-i,j , where i and j represent the characteristic subscripts of different aircrafts.

[0083] Step 4 obtains the fusion recognition result of multiple aircraft features through weighted summation: According to the target membership M of each feature obtained in step 3 i,j ={M i,j (w1),M i,j (w2),M i,j (w3)} and its corresponding fuzzy measure g λ-i,j , the target membership of multiple aircraft feature fusion recognition is obtained by weighted summation, denoted as M = {M(w1), M(w2), M(w3)}:

[0084]

[0085] Here ∑M i,j *g λ-i,j ,∑g λ-i,j Represents all M i,j *g λ-i,j All λ-i,j sum.

[0086] The fusion recognition result with the maximum membership as the target.

[0087] Embodiment 1:

[0088] The flow chart of the fusion recognition of two aircraft features is as follows: Figure 1 As shown, the specific implementation process is:

[0089] Step 1: Acquisition of aircraft features:

[0090] The aircraft characteristics of the same target obtained: the visible light geometric characteristics of the aircraft body and the visible light geometric characteristics of the aircraft contrail.

[0091] Target recognition framework: {BX, B-XX, F-XX}.

[0092] Step 2: Single-source identification of aircraft features:

[0093] Aircraft body geometric feature confidence: {0.035, 0.735, 0.233},

[0094] The result of the geometric feature recognition of the aircraft body is: B-1B, and the true value is B-1B.

[0095] Confidence of geometric features of aircraft contrails: {0.211, 0.768, 0.024},

[0096] The geometric feature recognition result of the aircraft contrail cloud is: B-1B, and the true value is B-1B.

[0097] Step 3D-S evidence synthesis:

[0098] DS fusion result: {0.01278883, 0.97752733, 0.00968384}.

[0099] Step 4: Obtain the fuzzy density of each feature:

[0100] After aircraft feature quality analysis, the feature types are {aircraft body geometric features, aircraft contrail geometric features, runway residual temperature features, aircraft loading features, aircraft contrail hyperspectral features, aircraft tail flame hyperspectral features}, and the corresponding fuzzy densities are {0.6, 0.6, 0.3, 0.25, 0.4, 0.4}.

[0101] Step 5: Solve for the associated parameter λ:

[0102] The solution is λ=-0.998.

[0103] Step 6 Calculate the fuzzy measure g after feature fusion λ :

[0104] Fusion fuzzy measure g λ =0.841.

[0105] Step 7 calculates the membership of the target through the normalized fuzzy integral:

[0106] Target membership:

[0107] {0.15424698758861904, 0.827993123903079, 0.16905877517687184},

[0108] The target fusion recognition result is: BX, which is correct.

[0109] The flow chart of multiple aircraft feature fusion recognition is as follows Figure 2 As shown, the specific implementation process is:

[0110] Step 1: Acquisition of aircraft features:

[0111] There are three characteristics of the same target aircraft:

[0112]

[0113] Target recognition framework: {BX, B-XX, F-XX}

[0114] Step 2 Aircraft feature combination:

[0115] The same target aircraft features obtained in step 1 are combined in pairs without duplication.

[0116] Get the characteristic pairs: {E1,E2}, {E1,E3}, {E2,E3}

[0117] Step 3: Fusion recognition of each feature pair

[0118] For the feature pairs obtained in step 2, perform pairwise fusion recognition to obtain the fusion recognition membership and fuzzy integral of each feature pair:

[0119] Feature pair number Membership Fuzzy integral 1,2 0.6000 0.3635 0.1291 0.4751 1,3 0.9460 0.1175 0.0842 0.7003 2,3 0.9470 0.1845 0.0195 0.7204

[0120] Step 4 obtains the fusion recognition results of multiple aircraft features through weighted summation:

[0121] According to the membership degree and fuzzy integral obtained in step 3, a weighted sum is performed to obtain the fusion recognition result of this set of aircraft features:

[0122]

[0123] The HL feature membership degree of this group M Identification results True Value Is it correct? 0.7573,0.1803,0.0624 BX BX correct

[0124] The present invention adopts the above technical scheme and conducts a verification experiment, wherein the number of aircraft unit features is 1279, the unit feature recognition accuracy is 91.39%, a total of 2362 two-item aircraft feature combinations are formed, the fusion recognition accuracy of the two-item aircraft feature combinations is 94.75%, a total of 5226 three-item aircraft feature combinations are formed, the fusion recognition accuracy of the three-item aircraft feature combinations is 96.84%, a total of 100 multi-item aircraft feature combinations are formed, and the fusion recognition accuracy of the multi-item aircraft feature combinations is 99%.

[0125] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-modal aircraft feature decision-level fusion method based on fuzzy integral, characterized by: The fusion method comprises the following steps: Different types of data are acquired for aircraft targets; features are identified according to the number of types, the features are fused, and finally the fused features are identified to obtain identification results.

2. The multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 1, characterized in that: There are two types of fusion methods: Step 1-1: Acquire two different types of data on the aircraft target; Step 1-2: Perform feature recognition on the two aircraft features respectively, and obtain two unit recognition results respectively; Step 1-3: Perform DS evidence synthesis on the two unit recognition results; Step 1-4: Obtain the fuzzy density of different aircraft features respectively; Step 1-5: Solve for the correlation parameter λ of the two aircraft features; Step 1-6: Calculate the fuzzy measure g after feature fusion λ ; Step 1-7: Calculate the membership of the target through the normalized fuzzy integral.

3. A multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 2, characterized in that: In the steps 1-3, the two unit recognition results {m1(w1),m1(w2),…,m1(w k )},{m2(w1),m2(w2),…,m2(w k )} to synthesize evidence, as follows: The recognition result {m(w1),m(w2),…,m(w k )}.

4. A multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 2, characterized in that: In the steps 1-5, the correlation parameter λ represents the correlation degree between the two aircraft features; when -1<λ<0, it indicates that the attributes are redundantly correlated, and when 0<λ, it indicates that the attributes are complementary correlated. According to the formula Using the fuzzy density obtained in steps 1-4, solve for the associated parameter λ.

5. The multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 2, characterized in that: In the steps 1-6, the fuzzy measure represents the importance of the combination of different aircraft features. According to the calculation formula of the fuzzy measure Calculate the fuzzy measure g of the feature pairwise combination based on the fuzzy density and correlation parameter λ obtained in steps 1-4 and 1-5 λ ; The pairwise combinations are calculated in turn with other fuzzy measures or fuzzy densities according to the formula to obtain the fuzzy measure results of more feature combinations.

6. A multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 2, characterized in that: In step 1-7, according to the recognition results obtained in steps 1-2 and 1-3, and the fuzzy density and fuzzy measure obtained in steps 1-4 and 1-6, the membership of the target is calculated by normalized fuzzy integral, which is recorded as {M(w1), M(w2), ..., M(w k )}: Among them, m max (w i )=max(m1(w i ),m2(w i )), g max is m max The corresponding fuzzy density is the fusion recognition result with the largest membership as the target.

7. The multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 1, characterized in that: The number of the types is three or more, and the fusion method is: Step 2-1: Acquire various types of data on the aircraft target; Step 2-2: Combining multiple aircraft features to form feature pairs; Step 2-3: Fusion recognition of each feature pair; Step 2-4: Obtain the fusion recognition results of multiple aircraft features through weighted summation.

8. The multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 7, characterized in that: In step 2-3, for the n aircraft features, they are combined into feature pairs without duplication, denoted as {E1, E2}, {E1, E3}, ..., {E n-1 ,E n For each feature pair, the steps 1-2 to 1-6 are respectively used to perform two aircraft feature fusion identifications to obtain the target membership and fuzzy measurement of each feature pair, which are respectively denoted as M j,l ={M j,l (w1),M j,l (w2),…,M j,l (w k )} and g λ-j,l , where j and l represent the characteristic subscripts of different aircrafts respectively.

9. The multi-modal aircraft feature decision-level fusion method based on fuzzy integral as claimed in claim 7, characterized in that: In step 2-4, the target membership M of each feature obtained in step 2-3 is j,l ={M j,l (w1),M j,l (w2),…,M j,l (w k )} and its corresponding fuzzy measure g λ-j,l , the target membership of multiple aircraft feature fusion recognition is obtained by weighted summation, denoted as M = {M(w1), M(w2), …, M(w k )}: Among them, ∑M j,l *g λ-j,l ,∑g λ-j,l Represents the product of all target membership and fuzzy measure M i,j *g λ-i,j And all fuzzy measures g λ-j,l sum; The fusion recognition result of multiple aircraft features is taken as the target corresponding to the category with the maximum value of each item in the target membership.

Images