A cluster ship formation identification method based on topological structure similarity

Abstract

The application belongs to the technical field of image processing, and discloses a cluster ship formation identification method based on topological structure similarity, which comprises the following steps: step one, constructing a single-stage anchor-free ship target key point detection model, estimating a cluster ship target in any direction by using a center point, and obtaining the coordinate position of the ship center point; step two, using a Delaunay triangulation network to realize expression of the peripheral contour of the cluster ship target, and drawing the topological structure of the peripheral contour of the cluster ship formation; step three, using the topological structure information of the formation to be identified to calculate the similarity with a public standard ship formation, so as to realize cluster ship formation identification; in the process, the topological structure, distance relationship, distribution range, distribution density and cluster ship target are used to calculate the similarity. The application can effectively identify the cluster ship formation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a cluster ship formation recognition method based on topological structure similarity. BACKGROUND

[0002] In a large range of high-resolution optical remote sensing images, the background of the ship target is complex and changeable, and multiple types of ship targets are densely arranged in any direction, with small intra-class differences, which leads to limitations of the general detection and recognition model. Therefore, cluster ship formation recognition in high-resolution visible light remote sensing images is a great challenge.

[0003] A group is a target collection with organization and mutual connection in a broad sense. Although a target group is composed of multiple targets, the main characteristics of the target group are the mutual relationship between the targets, the organizational structure between the groups, and the context relationship between the single target and the whole. Cluster ships are different from isolated ships, and often show group characteristics in the process of movement. The movement of the group members is constrained by the other members in the group, and the movement characteristics are consistent with the whole group. The movement mode of the cluster is relatively similar, and the spatial position is relatively fixed within a certain time. There are relatively few methods for cluster ship formation recognition at present, which can be roughly divided into three categories: cluster ship formation recognition method based on structural unit, cluster ship formation recognition method based on graph model, and cluster ship formation recognition method based on shape context. According to the arrangement mode of the group ship, a plurality of basic formations can be obtained, and the basic unit in the formation can be understood to analyze the whole ship group formation, so as to intelligently perceive the formation.

[0004] However, this kind of method is more dependent on the basic unit constituting the formation, and the recognition result of the basic unit directly affects the efficiency of the formation recognition. In order to solve the problems such as great change of target cluster ship target scale and dense arrangement of ship targets in any direction, some scholars use graph model to recognize group formation, extract some key points in the graph, and then find the corresponding points in the image to be matched, and connect them in turn to constitute the graph model. This simple graph model for recognizing group formation only considers the position information of the single target node, and does not consider the close relationship of the context connection of the group behavior, and lacks rich high-level semantic information. At the same time, for cluster ship formation recognition, the local feature descriptor cannot be directly used in the graph matching method, and the close relationship between the ship formation targets cannot be described by a single node (ship) to describe the relationship between the whole graph (formation). Although the target group recognition method based on shape context uses a histogram to count the sampling points of the target edge and obtains a local descriptor of the shape context, this kind of method needs a large number of sampling points, and for cluster ship formation, the number of ship targets is limited, so this kind of method cannot obtain effective ship formation recognition result. SUMMARY

[0005] The application intends to provide a cluster ship formation identification method based on topological structure similarity.

[0006] To achieve the above-mentioned purpose, the application provides the following technical scheme, a cluster ship formation identification method based on topological structure similarity, comprising the following steps:

[0007] Step one, constructing a single-stage anchor-free ship target key point detection model, using the center point to estimate the cluster ship target in any direction, and obtaining the coordinate position of the ship center point;

[0008] Step two, using the Delaunay triangulation network to realize the expression of the peripheral contour of the cluster ship target, and drawing the topological structure of the peripheral contour of the cluster ship formation;

[0009] Step three, using the topological structure information of the to-be-identified formation to calculate the similarity with the public standard ship formation, so as to realize the cluster ship formation identification; in the process, the topological structure, distance relationship, distribution range, distribution density and cluster ship target are used to calculate the similarity.

[0010] In another preferred embodiment of the application, in step one, the single-stage anchor-free ship target key point detection model is constructed, the rotating frame is used to detect the ship target arranged in any direction and with extreme length-width ratio, and the Centernet network is used to realize the single-stage anchor-free key point ship target detection;

[0011] The DLA backbone network is used to fuse semantic and spatial information and extract target feature vectors; at the same time, the CA attention mechanism is added to capture the position and channel information of the network; the CA attention mechanism includes information embedding and coordinate attention generation, and the output is,

[0012]

[0013] In another preferred embodiment of the application, in step one, the coordinate position of the ship center point is obtained by inputting the input image into the backbone network to generate a heat map of the ship target center point W, H and C are the width, height and category of the to-be-detected target; and the center point is mapped to , using the Gaussian kernel function,

[0014]

[0015] In another preferred embodiment of the application, in step one, the focal loss is used to alleviate the imbalance problem of positive and negative samples in the training process, and the specific formula is,

[0016]

[0017] In another preferred embodiment of the present application, in step two, the constraint condition of the Delaunay triangle is that the circumcircle of each constructed triangle does not contain other vertices of the triangle at the same time, and the line segments connected by the endpoints do not intersect each other.

[0018] In another preferred embodiment of the present application, in step three, the calculation of the similarity of the cluster topology structure uses the point neighbor representation using the topology relationship of the cluster, uses the fixed-length distance neighbor as the parameter for describing the topology information of the cluster, tries different fixed-length distances d, obtains the number of neighbors of each point in the topology structure, and the similarity of the topology structure is calculated as,

[0019]

[0020] In another preferred embodiment of the present application, in step three, the similarity calculation of the distribution range of the cluster uses the area S of the topology structure to reflect the approximate distribution range of the cluster, and the calculation formula is as follows,

[0021]

[0022] In another preferred embodiment of the present application, in step three, the similarity calculation of the distribution density of the cluster uses the area S of the topology structure to reflect the approximate distribution range of the cluster, and the calculation formula is as follows,

[0023]

[0024] In another preferred embodiment of the present application, in step three, the similarity calculation of the distance relationship of the cluster uses the area S of the topology structure to reflect the approximate distribution range of the cluster, and the calculation formula is as follows,

[0025]

[0026] In another preferred embodiment of the present application, in step three, the similarity calculation of the cluster ship target uses the area S of the topology structure to reflect the approximate distribution range of the cluster, and the calculation formula is as follows,

[0027]

[0028] In summary, the present application designs a cluster ship formation recognition method based on topology structure similarity, obtains ship node information, draws a cluster topology structure, and improves the robustness of description. First, a single-stage anchor-free ship target key point detection algorithm is used to obtain the coordinate position of the ship center point; second, the Delaunay triangle net is modeled to build the cluster ship target outer contour, and the topology structure of the cluster ship target is drawn; finally, the similarity to the standard formation is calculated, the cluster ship formation recognition is realized, and the cluster formation and change are judged.

[0029] Additional aspects and advantages of the present application will be made apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0030] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which:

[0031] Figure 1 is a flow chart of the embodiment of the present application.

[0032] Figure 2 is a schematic diagram of the position of the center point of a ship in the embodiment of the present application.

[0033] Figure 3 is a Delaunay triangulation of a cluster ship formation in the embodiment of the present application.

[0034] Figure 4 is a standard aircraft carrier battle group formation in the embodiment of the present application.

[0035] Figure 5 is a convex hull of a cluster ship formation in the embodiment of the present application. DETAILED DESCRIPTION

[0036] Embodiments of the present application are described in detail below with reference to the attached drawing figures, wherein the same or like reference numerals are used throughout the drawing figures to refer to the same or like elements or elements having the same or similar functionality. The embodiments described below are merely exemplary for the purposes of explanation and are not intended to limit the application, which is limited only by the claims.

[0037] In the description of the present application, it needs to be understood that the orientation or positional relationship indicated by the terms "longitudinal", "transverse", "vertical", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0038] In the description of the present application, unless otherwise specified and limited, it needs to be explained that the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be mechanical connection or electrical connection, it can be the communication between the two elements, it can be direct connection, or indirect connection through intermediate medium, and the specific meaning of the above terms can be understood by the person skilled in the art according to the specific circumstances.

[0039] The application provides a cluster ship formation identification method based on topological structure similarity. Figure 1 As shown in the preferred embodiment of the application, it comprises the following steps:

[0040] Step one, construct a single-stage anchor-free frame ship target key point detection model, estimate any direction cluster ship target by using the center point, and obtain the coordinate position of the ship center point.

[0041] Construct a single-stage anchor-free frame ship target key point detection model:

[0042] The rotating frame is used to detect the ship target arranged in any direction and with extreme length-width ratio, and the Centernet network is used to realize the single-stage anchor-free frame key point ship target detection, considering that the center point of the air-to-ground target is not affected by the direction change.

[0043] Firstly, the DLA backbone network is used to fuse semantic and spatial information and extract target feature vectors; and the CA attention mechanism is added, which can capture the position and channel information of the network, enhance the attention degree to the ship target, and further improve the expression ability of the network to the target feature extraction.

[0044] The CA attention mechanism includes information embedding and coordinate attention generation, and the output is:

[0045]

[0046] Wherein x c (i,j) is the input feature map, is the attention weight of the input feature in the height direction, is the attention weight of the input feature in the width direction.

[0047] The center point position estimation network obtains the position of the ship center point:

[0048] The input image is input into the feature extraction backbone network proposed in the above steps to generate the heat map of the ship target center point W, H, and C are the width, height, and category of the target to be detected. The center point is mapped to by using a Gaussian kernel function, and the Gaussian kernel function is

[0049]

[0050] Wherein p is the center point position of the target, (x,y) is the coordinate position of the target, is the position of the center point after down-sampling.

[0051] Most of the samples on the heat map are negative samples, and a small number of center points exist as positive samples. Focal loss is used to alleviate the imbalance between positive and negative samples during training. The specific formula is,

[0052]

[0053] Where N is the number of ship targets in the network, and a, b are the penalty coefficients in the training process, is the center point prediction map, and Y xyc is the true value heat map.

[0054] Step two, use the Delaunay triangulation network to realize the expression of the peripheral contour of the cluster ship target, and draw the topological structure of the peripheral contour of the cluster ship formation. Delaunay triangle can represent the boundary of any shape area, and can also well represent the cluster point group target with random interference factors. Compared with single target or multiple targets, group behavior is the whole line. Using the simple structure and small data redundancy of Delaunay triangle method to construct the peripheral contour (convex hull) of the cluster ship target is helpful to the effective expression of the shape distribution of the point group.

[0055] First, the center point of the ship target is regarded as the node of the graph structure, and the data set is V=P∪L, P is the center point of the ship (regarded as discrete points), and these points are mutually visible; L is the end point set of the line segment connecting the center points. The constraint condition of Delaunay triangle is that the circumcircle of each constructed triangle does not contain other vertices of the triangle at the same time, and the line segments connected by the end points do not intersect each other.

[0056] The ship is analyzed as a discrete point, and the center point coordinate position of the cluster formation ship target is obtained by using step one. The center point of the ship target is regarded as the node of the graph structure, as shown in Figure 2 .

[0057] Then the peripheral contour (convex hull) of the cluster ship target is constructed. The peripheral contour of the group target is often calculated by the convex hull of the convex polygon. On the basis of having obtained the group nodes (formation ship center points), the Delaunay triangle method is used to connect each node between the groups to construct the peripheral contour (convex polygon) of the point group, as shown in Figure 3 .

[0058] Step three, the distribution of the cluster ship formation has a close relationship with the convex hull. In step two, the Delaunay triangulation network establishes the topological structure of the cluster ship target. The topological structure information of the to-be-recognized formation is used to calculate the similarity with the public standard ship formation, so as to realize the recognition of the cluster ship formation; the topological structure, distance relationship, distribution range, distribution density and cluster ship target are used to calculate the similarity in the process.

[0059] Cluster topological structure similarity calculation:

[0060] The topology relationship of the cluster is used to represent the neighbors of the point, the neighbors with a fixed distance are used as parameters to describe the topology information of the cluster, different fixed distances d are tried, the number of neighbors of each point in the topology structure is obtained, and the similarity calculation of the topology structure is,

[0061]

[0062] Wherein, N is the number of cluster ship target neighbors, and n is the number of ship targets.

[0063] The similarity calculation of the distribution range of the cluster is:

[0064] The area S of the convex hull is used to reflect the approximate distribution range of the cluster, and the calculation formula is as follows,

[0065]

[0066] The similarity calculation of the distribution density of the cluster is:

[0067] The calculation of the distribution density of the cluster is the ratio of the number n of cluster ships and the area S of the convex hull, which reflects the sparsity, and the calculation formula is as follows,

[0068]

[0069] The similarity calculation of the distance relationship of the cluster is:

[0070] The circumscribed minimum matrix (length X, width Y) is used to calculate the distance similarity of the cluster ship target, which reflects the shape distribution of the cluster, and the calculation formula is as follows,

[0071]

[0072] The similarity calculation of the cluster ship target is:

[0073] In order to realize the stability of the geometric mean value when processing the difference problem, the overall calculation of the similarity of the cluster ship target is carried out, and the calculation formula is as follows,

[0074]

[0075] Experiment: using the technical scheme provided by the application, simulation experiment is carried out.

[0076] 1. Simulation condition

[0077] In order to test the effectiveness of the application and verify the effectiveness of the cluster ship target recognition based on the similarity calculation, the following experiments are carried out. 200 cluster ship formation data are generated by using a simulation platform, and the application is based on Ubuntu 18.04 system and deep learning Pytorch framework to carry out experiments.

[0078] Among them,Figure 4 To determine the formation and change degree of the cluster ship target by calculating the similarity of the cluster ship and the public cluster ship target formation; Figure 5 To cluster the ship formation and the peripheral contour.

[0079] 2, simulation results

[0080] (1) The calculation factor statistics results required by the cluster ship formation similarity calculation are shown in Table 1

[0081]

[0082]

[0083] (2) The similarity of each factor of the cluster and the overall similarity calculation results are shown in Table 2

[0084] Clustered ships SIM Topo ]]> SIM Area ]]> SIM thic ]]> SIM sp ]]> SIM Cluster 1, 2 0.772 0.271 0.560 0.864 0.564 Cluster 1, 3 0.860 0.876 0.861 0.914 0.877 Cluster 2, 3 0.862 0.309 0.604 0.896 0.616

[0085] Figure 4 The single-double aircraft carrier departure formation diagrams of cluster 2 (b) and cluster 3 (c) are Figure 3 By calculating the similarity of the cluster ship and the public cluster ship target formation, the formation and change degree of the cluster ship target are determined, and the experimental results are shown in Table 1 and Table 2. Figure 4 From Table 1, the similarity results of cluster 1 and cluster 3 are higher, and in combination with the public cluster ship (

[0086] , Figure 3 (d)) Single-double aircraft carrier departure formation comparison, the formation of cluster ships 1 and 3 are similar to double aircraft carrier departure formation, and the change is not large in a certain time period. Figure 4

[0087] In summary, the ship target detection model based on the key point is adopted to obtain the coordinate position of the ship center point; secondly, the ship is analyzed as a discrete point, the ship target center point is regarded as the node of the graph structure, and the Delaunay triangle method is used to draw the topological structure (convex hull) of the cluster ship formation peripheral contour; finally, the density distribution of the cluster ship is calculated by using the area of the convex hull and the number of target points in the cluster, and then compared and identified with the public standard formation.

[0088] ​In the description of the specification, the description of the terms "preferred embodiment", "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the particular feature, structure, material or characteristic being described in connection with the embodiment or example is included in at least one embodiment or example of the application. The illustrative appearance of the above-mentioned terms in various places in the specification are not necessarily referred to the same embodiment or example. Moreover, the particular features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0089] Although embodiments of the present application have been shown and described, it would be appreciated by those skilled in the art that changes, modifications, alternatives and variations to these embodiments could be made without departing from the principles and spirit of the application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for recognizing naval formations based on topological similarity, characterized in that, Includes the following steps: Step 1: Construct a single-stage, anchor-free ship target key point detection model, use the center point to estimate the cluster of ship targets in any direction, and obtain the coordinate position of the ship's center point. A single-stage, frameless ship target key point detection model is constructed. A rotating frame is used to detect ship targets arranged in arbitrary directions with extreme aspect ratios. The Centernet network is used to realize the single-stage, frameless key point ship target detection. The DLA backbone network is used to fuse semantic and spatial information to extract target feature vectors. Simultaneously, a CA attention mechanism is incorporated to capture the network's positional and channel information. The CA attention mechanism includes information embedding and coordinate attention generation, with the output being... ;in For the input feature map, Assign attention weights to the input features in the height direction. Assign attention weights to the input features in the width direction; Specifically, to obtain the coordinates of the ship's center point, the input image is fed into the backbone network to generate a heatmap of the ship's target center point. , The width, height, and category of the target to be detected are defined respectively; and a Gaussian kernel function is used to map the center point to... Above, the Gaussian kernel function is, ;in The center point of the target The target coordinates are... This represents the position of the center point after downsampling. Step 2: Use the Deloitte triangulation to represent the outer contour of the cluster of ships and draw the topological structure of the outer contour of the cluster of ships. Step 3: Utilize the topological structure information of the formation to be identified to calculate the similarity with publicly available standard ship formations, thereby achieving cluster ship formation identification; the similarity is calculated using topological structure, distance relationships, distribution range, distribution density, and cluster ship targets.

2. The method for recognizing naval formations based on topological similarity according to claim 1, characterized in that: Step one uses focal loss to alleviate the imbalance between positive and negative samples during training. The specific formula is as follows: ； in The number of ship targets in the network. This is the penalty coefficient during the training process. For center point prediction map, This is a heatmap of the true values.

3. The method for recognizing clustered ship formations based on topological similarity according to claim 2, characterized in that: In step two, the constraints of the Deloitte triangle are that the circumcircle of each constructed triangle does not contain any other vertices of the triangle, and the line segments connecting the endpoints do not intersect each other.

4. The method for recognizing clustered ship formations based on topological similarity according to claim 3, characterized in that: Step three, cluster topology similarity calculation, utilizes the cluster's topological relationships using a neighbor representation. Fixed-distance neighbors are used as parameters to describe the cluster's topology information, and different fixed-distance neighbors are explored. The number of neighbors for each point in the topology is obtained, and the similarity of the topology is calculated as follows: ;in, The number of target neighbors for the cluster of ships. The number of ship targets.

5. The method for recognizing clustered ship formations based on topological similarity according to claim 4, characterized in that: In step three, the similarity calculation of the cluster distribution range uses the area S of the topological structure to reflect the approximate distribution range of the clusters. The calculation formula is as follows. 。 6. The method for recognizing naval formations based on topological similarity according to claim 5, characterized in that: The similarity calculation of cluster distribution density in step three is performed using the following formula: 。 7. The method for recognizing clustered ship formations based on topological similarity according to claim 6, characterized in that: The similarity calculation of cluster distance relationships in step three has a length of , width is The calculation formula is as follows: 。 8. The method for recognizing clustered ship formations based on topological similarity according to claim 7, characterized in that: The similarity calculation for the cluster of ship targets in step three is calculated using the following formula: 。

Images