A distributed image fusion method based on average consensus

Abstract

The present invention belongs to the field of image fusion technology, and in particular, relates to a distributed image fusion method based on average consensus. The main steps of the present invention are: each flight platform collects target image information from different perspectives; obtains the average of the initial values ​​of all nodes according to the average consensus algorithm; calculates the eigenvector of the covariance matrix of the total output of the node by a distributed method; and obtains the image fusion result after distributed processing. In the method of the present invention, each platform of the flight cluster only needs to transmit data information with adjacent nodes, avoiding the problem of communication congestion of the central node, greatly reducing the requirements for communication capacity between platforms, and improving the real-time processing capability of the flight cluster platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of image fusion, and in particular relates to a distributed image fusion method based on average consensus. Background Art

[0002] Flying swarms have attracted widespread attention due to their efficient resource sharing efficiency and close coordination capabilities. In a flying swarm, the collaboration between nodes (or individuals) does not need to rely on a single central control unit to guide their behavior and decision-making process, which can greatly improve the efficiency of completing tasks such as target detection. Flying swarm multi-platform image fusion refers to the process of integrating image information from multiple sources to generate new images with more information, which can enhance image information, improve image quality or extract more useful data.

[0003] However, the existing flight cluster image processing is centralized processing, which has the following disadvantages:

[0004] 1. The core of a centralized system is the central server or database. Once this central component fails, the service of the entire system will be interrupted.

[0005] 2. When the number of nodes increases, all data needs to be transmitted to the central node when centralized processing is used, which poses a serious challenge to the communication capacity of the fusion center node. When a node is far away from the central node or is blocked, there are also problems such as communication interruption.

[0006] 3. When processing high-resolution images or video streams, the computing power problem of the central node is more prominent. High-resolution images and video streams contain more data and require very high computing resources. At the same time, it also exacerbates memory requirements and latency issues.

[0007] Although existing distributed processing methods have been applied in the fields of radar signal processing and tracking and positioning, they have not yet been applied in the field of cluster image fusion. Summary of the invention

[0008] In view of the shortcomings in the current image fusion field, the present invention provides a distributed image fusion method based on average consensus.

[0009] The technical solution adopted by the present invention is:

[0010] A distributed image fusion method based on average consensus includes the following steps:

[0011] S1. Each flight platform collects target image information from different perspectives. Specifically, M flight platforms with cameras are arranged in a distributed manner, each flight platform is used as a node, and each node is allowed to communicate only with its adjacent nodes to construct a node connectivity graph, and the image data obtained by node m is represented by a Z-dimensional vector xm Indicates that m = 0, 1, ..., M-1;

[0012] S2. Obtain the average of the initial values ​​of all nodes according to the average consensus algorithm, specifically:

[0013] Let u m is the mean of node m, u(0) is u0,u1,…u M-1 The value of the combined M-dimensional vector at iteration 0, that is, the initial value, is ac After 2 times, the average consensus is obtained:

[0014] u(I ac )=11 T u(0) / M

[0015] Among them, 1 represents a vector of all 1s, u(I ac ) represent node iteration I ac The value of times, I ac £M-1; the calculation method for each iteration is:

[0016] u(t+1)=W(t)u(t)

[0017]

[0018] W(t)=(Ll t+1 I)t=1,2,…,R-2

[0019] Where t represents the number of iterations, L is the Laplacian matrix of the node connectivity graph, I is the identity matrix, and l i are different eigenvalues ​​of L, i = 1, 2, …, R, where R represents the total number of eigenvalues ​​of L;

[0020] Iterate I ac After the u(I ac ) is regarded as u m The exact average value of m (u m );

[0021] S3. Calculate the eigenvector of the covariance matrix of the total output of the node in a distributed manner, specifically:

[0022] Calculate the first eigenvector of the matrix, randomly initialize the MZ-dimensional vector r1(0), and iterate:

[0023]

[0024] in, represents the portion of r1(n) stored in node m, K is the total number of snapshots, and x m[k] represents the output of the kth snapshot of node m, AC m (u m ), set the number of iterations and then update r1(n) by the method in S2;

[0025] After I m After iterations, the obtained vector is normalized to obtain an estimate of the eigenvector:

[0026]

[0027] Get the first eigenvector After that, the other eigenvectors are calculated in turn. Specifically, the calculation method of the qth eigenvector is, q≥2:

[0028]

[0029] S4. Obtain the image fusion result after distributed processing, specifically:

[0030] Arrange the obtained q eigenvectors in descending order of their eigenvalues ​​to form a transformation matrix F, and then transform the output vector as follows:

[0031] y=Fx

[0032] Among them, the transformation matrix F is a q′MZ-dimensional matrix, x is an MZ-dimensional vector, and y is a q-dimensional vector, which realizes the compression and fusion of image data from MZ dimension to q dimension.

[0033] Beneficial effects of the present invention:

[0034] 1) Each platform in the flight cluster only needs to transmit data information with adjacent nodes, avoiding the problem of communication congestion in the central node, greatly reducing the requirements for communication capacity between platforms, and improving the real-time processing capability of the flight cluster platform;

[0035] 2) A certain flying node only needs to establish a link with a certain node in the cluster to complete image fusion, which improves the adaptability of cluster detection to the environment, such as in mountains and forests, urban buildings, where it is impossible to guarantee that any node can have a communication link;

[0036] 3) Computing resources can be evenly distributed across the nodes in the cluster, achieving multi-platform parallel computing, avoiding the problem of insufficient computing at the central node, and effectively improving processing efficiency;

[0037] 4) It can be applied to flexible spatial cluster formations because only adjacent nodes need to transmit information to each other, so it has high reliability and scalability, and the cluster detection system is more robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of a distributed image fusion method based on average consensus proposed by the present invention.

[0039] Figure 2 It is the distributed connectivity graph of the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0041] like Figure 1 As shown, the method of the present invention comprises the following main steps:

[0042] S1. Each flight platform collects target image information from different perspectives;

[0043] S2. Obtain the average of the initial values ​​of all nodes according to the average consensus algorithm;

[0044] S3, calculating the eigenvector of the covariance matrix of the total output of the node in a distributed manner;

[0045] S4. Obtain the image fusion result after distributed processing.

[0046] In step S1, it is defined that M camera platforms are arranged in a distributed manner to capture images from different perspectives. The communication network can be modeled using an undirected graph G = (V, E), where V is a set of nodes and E is a set of edges. Each node represents a camera platform and an edge represents a communication link. The image data acquired by node m (m = 0, 1, ..., M-1) is represented by a Z-dimensional vector x m Indicates; suppose a total of K snapshots of data are collected, and the output of node m in the kth snapshot is x m [k]. Figure 2 As shown, each node communicates only with its adjacent nodes, and the image information output by each node only needs to be transmitted to the adjacent nodes.

[0047] In step S2, the node updates the value through distributed linear iteration in the following form:

[0048]

[0049] Where t represents the number of iterations, m (= 0, 1, ... M-1) represents the node, and u m (t) represents the scalar value of node m at t iterations, W mm (t) represents u m The weight of (t); y m ={i{m,i}∈e} represents the adjacent nodes of node m, e represents the neighboring node set of node m, W mi (t) represents u iThe above formula represents the updating process of node m, that is, the value of node m at the next moment is obtained by weighting the values ​​of node m and its adjacent nodes at the previous moment. The iteration is performed simultaneously at each node, and the mean value of all nodes can be obtained after a certain number of iterations, thus realizing parallel operation.

[0050] Writing formula (1) in matrix form, we can get

[0051] u(t+1)=W(t)u(t) (2)

[0052] The weight matrix is

[0053]

[0054] W(t)=(Ll t+1 I)t=1,2,…,R-2

[0055] Where L is the Laplacian matrix of the camera node connectivity graph, I is the identity matrix, and l i (i=1,2,…,R) are different eigenvalues ​​of L.

[0056] The time-limited average consensus is the average consensus in iteration I. ac (￡M-1) times later, we get [1]

[0057] u(I ac )=11 T u(0) / M (4)

[0058] Among them, 1 represents a vector of all 1s, u(0), u(I ac ) represent the node initial value and iteration I respectively ac The value of times.

[0059] In step S3, assuming that the eigenvalues ​​of the covariance matrix output by the camera nodes are different, the eigenvectors of the matrix are calculated next: first, the first eigenvector of the matrix is ​​calculated, and then the other eigenvectors are calculated one by one using the calculated eigenvectors.

[0060] The first eigenvector is calculated as follows:

[0061] First, randomly initialize the MZ-dimensional vector r1(0);

[0062] Then, iterate according to the following formula

[0063]

[0064] in in represents a portion r1(n) stored in node m; K is the total number of snapshots, x m[k] represents the output image of node m in the kth snapshot; AC m (u m ) represents the scalar value u after sufficient iterations m The average consistency of the output is u m The exact average value of in is the mean of all nodes, that is, So there is

[0065]

[0066] So we can calculate r of all nodes 1,m (n+1), and get the next iteration value r1(n+1).

[0067] Finally, after a certain number of iterations, the obtained vector is normalized to obtain an estimate of the feature vector. The normalization method is as follows

[0068]

[0069] Among them I m Represents the total number of iterations that make r1 converge, r1(I m ) represents the final convergence value of the iteration, represents the estimate of the first eigenvector.

[0070] The other eigenvectors are calculated as follows:

[0071] To estimate the qth eigenvector, first calculate the first q-1 eigenvectors;

[0072] because is a matrix The largest eigenvector of is the estimate of the i-th eigenvector, which has the following iterative formula

[0073]

[0074] Since the above formula is the same as the iterative form for calculating the first eigenvector, the estimate of the qth eigenvector can be obtained according to the same processing method as above

[0075] In step S4, the q eigenvectors obtained in S3 are arranged in descending order of their eigenvalues ​​to form a transformation matrix F, and then the output vector is transformed as follows:

[0076] y=Fx (8)

[0077] The transformation matrix F is a q′MZ-dimensional matrix, x is an MZ-dimensional vector, and y is a q-dimensional vector. This realizes the compression fusion of image data from MZ dimensions to q dimensions. Similarly, if greater compression is required, the first m (m

Claims

1. A distributed image fusion method based on average consensus, characterized in that: The following steps are involved: S1. Each flight platform collects target image information from different perspectives. Specifically, M flight platforms with cameras are arranged in a distributed manner, each flight platform is used as a node, and each node is allowed to communicate only with its adjacent nodes to construct a node connectivity graph, and the image data obtained by node m is represented by a Z-dimensional vector x m Indicates that m = 0, 1, ..., M-1; S2. Obtain the average of the initial values ​​of all nodes according to the average consensus algorithm, specifically: Let u m is the mean of node m, u(0) is u0,u1,…u M-1 The value of the combined M-dimensional vector at iteration 0, that is, the initial value, is ac After 2 times, the average consensus is obtained: u(I ac )=11 T u(0) / M Among them, 1 represents a vector of all 1s, u(I ac ) represent node iteration I ac The value of times, I ac ≤M-1; the calculation method for each iteration is: u(t+1)=W(t)u(t) Where t represents the number of iterations, L is the Laplacian matrix of the node connectivity graph, I is the identity matrix, and λ i are different eigenvalues ​​of L, i = 1, 2, …, R, where R represents the total number of eigenvalues ​​of L; Iterate I ac After the u(I ac ) is considered as u m The exact average value of m (u m ); S3. Calculate the eigenvector of the covariance matrix of the total output of the node in a distributed manner, specifically: Calculate the first eigenvector of the matrix, randomly initialize the MZ-dimensional vector r1(0), and iterate: in, represents the portion of r1(n) stored in node m, K is the total number of snapshots, and x m [k] represents the output of the kth snapshot of node m, AC m (u m ), set the number of iterations and then update r1(n) by the method in S2; After I m After iterations, the obtained vector is normalized to obtain an estimate of the eigenvector: Get the first eigenvector After that, the other eigenvectors are calculated in turn. Specifically, the calculation method of the qth eigenvector is, q≥2: in is the estimate of the i-th eigenvector; S4. Obtain the image fusion result after distributed processing, specifically: Arrange the obtained q eigenvectors in descending order of their eigenvalues ​​to form a transformation matrix F, and then transform the output vector as follows: y=Fx Among them, the transformation matrix F is a q×MZ dimensional matrix, x is an MZ dimensional vector, and y is a q dimensional vector.

Images