1. A ship automatic matching method based on optical satellite remote sensing images is characterized by comprising the following steps:

(1) defining the optical satellite remote sensing image with the shooting time before as p₁The optical satellite remote sensing image with the later shooting time is p₂To p for_iRespectively carrying out coarse classification to obtain images k of coarse classification results_iI represents the ith image, (i is 1, 2);

(2) image k of coarse classification result by rectangular template₁、k₂Matching, removing non-ship connected domains, and extracting a ship target area;

the method comprises the following specific steps:

(2.1) defining the upper left corner point of the image as an origin, and the right direction is the positive direction of an x axis and the downward direction is the positive direction of a y axis;

(2.2) calculating k₁、k₂Obtaining the length, the width and the angle of each connected domain according to the minimum circumscribed rectangle of each connected domain, wherein the length and the width are the length and the width of the corresponding minimum circumscribed rectangle, the angle is an included angle between the extension line of the long side of the minimum circumscribed rectangle and the positive direction of the x axis above the x axis, calculating the length-width ratio of the connected domains, and only keeping the connected domains with the length-width ratio of more than 1.5 and less than 15 to obtain the rest connected domains;

(2.3) calculating k separately₁、k₂The ship rectangular template of each residual connected domain is obtained to obtain k₁Template set TEM₁＝{temple_1，1，temple_1，2… } and k₂Template set TEM₂＝{temple_2，1，temple_2，2，…}；

(2.4) use of TEM₁For the roughly classified image k₁Template matching using TEM₂For the roughly classified image k₂Carrying out template matching;

(2.5) expanding the ship target area to ensure that the ship is completely contained in the ship target area;

(3) inputting the four corner coordinates of the extended ship target area into a GrabCT algorithm, and accurately extracting a remote sensing image p₁、p₂The ship in (1) to obtain the extracted image z₁And z₂；

(4) Calculating the extracted image z₁、z₂A feature vector for each vessel;

(5) remote sensing image p by using ship feature vector₁、p₂And (5) matching the ships, and outputting a ship matching result.

2. The automatic ship matching method based on optical satellite remote sensing images as claimed in claim 1, wherein the rough classification process in step (1) is as follows:

(1.1) use of a Gaussian matrix pair p with a standard deviation of 0 and a size of 5X 5_iPerforming Gaussian filtering smoothing to convert the smoothed image into a gray image g_iThe formula is as follows:

g_i＝R_i·0.299+G_i·0.587+B_i·0.114

in the formula, R_i、G_i、B_iRespectively represents p_iAnd G, smoothing the red channel, the green channel and the blue channel by Gaussian, wherein the Gaussian matrix used by the Gaussian smoothing is as follows:

(1.2) calculating the grayscale image g_iEdge gradient image Grad of_iThe formula is as follows:

in the formula, G_ixDenotes g_iGrayscale image G for lateral edge detection_iyDenotes g_iGray level images of longitudinal edge detection;

(1.3) use of the threshold T_iaFor the edge gradient image Grad obtained in the step (1.2)_iPerforming binarization processing to obtain a mask image omega_iThreshold value T_iaThe calculation formula is as follows:

T_ia＝max(Grad_i)·0.2

(1.4) mask image-based ω_iObtaining a gray image g_iTarget area a of_iCalculating a_iHistogram of gray levels H_iIs a frequency variation value v_i[r]Obtaining an image adaptive segmentation threshold T_ibBy means of T_ibFor g_iDividing to be larger than threshold value T_ibIs a foreground region, less than a threshold value T_ibDefining the pixel value of the foreground area after segmentation as 1 and the pixel value of the background area as 0 to obtain a segmentation result f_i；

Wherein, T_ibThe calculation formula is as follows:

in the formula, h_imeanRepresenting the mean value of the image grey frequencies, h_i[r]Denotes a frequency (r is 0,1, …,255) at a gray value of r, v_i[r]Indicates a frequency variation value (r ═ 0,1, …,255) at the gradation value r,representing the gray value at which the frequency variation value is minimum;

(1.5) Using 2X 2 inner core pairs f, respectively_iAnd corroding and expanding to obtain a coarse classification result image ki.

3. The automatic ship matching method based on optical satellite remote sensing images as claimed in claim 1, wherein the calculation process of the template in step (2.3) is as follows:

the length and the angle of the foreground in the template are respectively the length and the angle of the corresponding connected domain, the ship length L is taken as an interval to divide real ship data, the average length-width ratio of the ship in each interval is counted, and the length-width ratio of the template foreground is determined according to the interval c where the approximate real length of the connected domain is locatedCalculating the foreground width in the template, wherein the calculation formula of the foreground width is as follows:

in the formula, length represents the foreground length in the template, e represents the number of ships in a ship length interval c in statistical data, delta represents the ship length-width ratio in the statistical data, and res represents the spatial resolution of the remote sensing image;

roughly classified image k according to foreground length, width and angle in template₁Generating corresponding template image temples for each connected domain_1，1、temple_1，2…, obtaining an aggregate TEM₁For the coarsely classified image k₂Generating corresponding template image temples for each connected domain_2，1、temple_2，2…, obtaining an aggregate TEM₂The pixel value of the foreground region in the template is 1, and the pixel value of the background region in the template is 0.

4. The automatic ship matching method based on the optical satellite remote sensing images as claimed in claim 3, wherein the length L of the ship is 0-50 m.

5. The automatic ship matching method based on optical satellite remote sensing images as claimed in claim 1, wherein the matching method in step (2.4) is as follows:

setting a threshold T_SThe template starts to slide from the original point of the image, one pixel is slid each time, the whole image is traversed, the similarity score S between the template and the coverage area of the template is calculated at each position, and if the similarity score is larger than T, the similarity score S is calculated_SIf the matching is successful, recording coordinates of four corner points of the template when the matching is successful, marking the coordinates as a ship target area, and if the same connected domain is marked for multiple times, only keeping the highest similarity score once as the ship target area;

wherein, the similarity score S formula is calculated as follows:

in the formula, sample represents a template, X represents a connected component, and N (-) represents the number of pixels having a pixel value of 1 in the region (-).

6. The automatic ship matching method based on optical satellite remote sensing images as claimed in claim 1, wherein the expansion formula in step (2.5) is as follows:

wherein (x'₁，y′₁)、(x′₂，y′₂) Coordinates (x) representing the extended upper left corner point and lower right corner point of the ship target area, respectively₁，y₁)、(x₂，y₂) Representing the corresponding coordinates of the target area of the original vessel.

7. The automatic ship matching method based on the optical satellite remote sensing image as claimed in claim 1, wherein the accurate ship extraction process in step (3) is as follows:

(3.1) ships in the ship target area are the foreground to be divided, seawater and other targets are the background, and p is respectively paired in RGB space₁、p₂The foreground and the background are modeled by a Gaussian mixture model, pixels outside a ship target area are marked as the background, pixels inside the ship target area are marked as possible foreground or background, the Gaussian mixture model is initialized by using a K-means algorithm, K is taken to be 5, each pixel is distributed with a proper Gaussian component, and the nth pixel is distributed with proper Gaussian componentsThe Gaussian component of a pixel is denoted as k_n；

(3.2) calculating p separately₁、p₂Gibbs energy E (α, k, θ, z) of the whole image, formula is as follows:

in the formula, α represents opacity, α ∈ {1,0}, θ ═ { pi (α, K), μ (α, K), ∑ (α, K), α ═ 0,1, K ═ 1,. and K }, z represents a video gray value array, U represents a region energy function, pi represents a weight of a gaussian component, μ represents a mean vector of each gaussian component, Σ represents a covariance matrix, V represents a boundary energy function, γ ═ 50, C represents a set of a certain pixel point and an adjacent pixel point, m and n represent two neighborhood pixels, and β represents a contrast between video pixels;

(3.3) updating the remote sensing image p₁、p₂Of each pixel of the imageAnd parameters of the Gaussian mixture matrixUntil convergence, finding the minimum value of the energy function E (alpha, k, theta, z) and obtaining p₁、p₂Obtaining the accurate extracted image z of the ship by the optimal segmentation boundary of the foreground and the background₁、z₂。

8. The automatic ship matching method based on optical satellite remote sensing images as claimed in claim 1,

the calculation step in step (4) is as follows:

calculating z₁、z₂The minimum bounding rectangle of each ship in the ship, the area of the ship, the length-width ratio lwr, the angle dir and the minimum bounding rectangle area are obtained_MERThereby obtaining a remote sensing image p₁、p₂Of each ship P1_w＝(rec_w，area_w，lwr_w，dir_w) And P2_j＝(rec_j，area_j，lwr_j，dir_j) Wherein w represents p₁The w ships extracted in (j) represents p₂Extracting the jth ship, rec represents the rectangular degree;

the calculation formula of the squareness rec is as follows:

9. the automatic ship matching method based on the optical satellite remote sensing image as claimed in claim 1, wherein the matching process in the step (5) is as follows:

(5.1) calculation of p₁The feature vector and p of each ship₂Establishing a bipartite graph between two image ships according to the similarity between the feature vectors of each ship, wherein the left vertex of the bipartite graph represents p₁Middle ship, right vertex denotes p₂Of a ship of (1), the similarity is greater than a threshold value T_dConnecting the ships, wherein the weight of the connection is the similarity between the characteristic vectors of the two ships;

(5.2) assigning an initial score to each vertex of the bipartite graph, wherein the initial score of the left vertex is the maximum weight of the side connected with the left vertex, and the initial score of the right vertex is 0;

(5.3) adding the scores of the top points at the left end and the right end of the connecting line, if the added score is more than or equal to the weight of the connecting line, considering that the two sides of the connecting line are the same ship, and when p is greater than or equal to the weight of the connecting line₁Two ships are matched to p₂When the ship is the same ship, the matching conflicts, a specific value d is set to be 0.1, the left vertex score corresponding to the two conflict connecting lines is reduced by d, and the right vertex score is increased by d;

and (5.4) repeating the step (5.3), when the left vertex score corresponding to the conflict connecting line is 0, abandoning the matching of the ship, continuing to match other ships until the matching of all the ships is completed, and outputting a ship matching result.

10. The automatic ship matching method based on optical satellite remote sensing images as claimed in claim 9, wherein the similarity calculation formula in step (5.1) is as follows:

wherein R (P1)_w，P2_j) Represents p₁W of the second ship and p₂Similarity between the j-th ship in the middle, 0 ≦ R (P1)_w，P2_j)≤1，Andrespectively represent P1_wAnd P2_jIs measured.

Background

In recent years, with the launching of static optical orbit satellites and video satellites, continuous observation of ships in a large area is made possible, and a rich data source is provided for ship matching. The ship is used as a main transport carrier at sea, and ship matching has important value in both military fields and civil fields. The ship matching can find the same ship in the two images, obtain the distribution condition of the ship and the movement direction of the specific ship, and provide reliable data support for monitoring the marine traffic, monitoring illegal pollutant discharge, illegal fishing and other illegal behaviors.

In the optical satellite remote sensing image, the ship is small in size, few in texture and unclear in characteristics, and interference of sea surface noise, cloud layers and other targets on the sea brings certain difficulty to ship matching. The existing ship matching method is roughly divided into two steps of ship detection and ship matching.

The main methods for ship detection are background subtraction and saliency detection. The background difference method has a good effect on detecting moving ships under a complex background, but is sensitive to environmental noise, can only detect ships moving at a high speed under a fixed background, and has a poor effect on detecting anchored ships and ships moving at a low speed. At present, the most applied significance detection method is significance detection based on contrast, has the advantages of high calculation efficiency and capability of outputting a full-resolution significance map, is easy to generate response on local objects, and is easy to lose effectiveness when cloud cover exists or other sea surface targets interfere.

The ship matching is mainly carried out by extracting local feature points through SIFT, SURF, ORB and other methods, and then matching is carried out. The ship matching effect by adopting the method is not ideal because the ship occupies a small area in the satellite remote sensing image, the characteristics are not obvious, and redundant data generated by calculating the characteristic points of the whole image is more.

Disclosure of Invention

Based on the problems in the prior art, the invention provides an automatic ship matching method based on optical satellite remote sensing images, which can realize ship detection from rough to fine, can accurately match target ships and realizes automatic ship matching of the optical satellite remote sensing images.

The invention adopts the following technical scheme:

a ship automatic matching method based on optical satellite remote sensing images comprises the following steps:

(1) defining the optical satellite remote sensing image with the shooting time before as p₁The optical satellite remote sensing image with the later shooting time is p₂To p for_iRespectively carrying out coarse classification to obtain images k of coarse classification results_iI represents the ith image, (i is 1, 2);

further, the coarse classification process in step (1) is as follows:

(1.1) use of a Gaussian matrix pair p with a standard deviation of 0 and a size of 5X 5_iPerforming Gaussian filtering smoothing to convert the smoothed image into a gray image g_iThe formula is as follows:

g_i＝R_i·0.299+G_i·0.587+B_i·0.114

in the formula, R_i、G_i、B_iRespectively represents p_iAnd G, smoothing the red channel, the green channel and the blue channel by Gaussian, wherein the Gaussian matrix used by the Gaussian smoothing is as follows:

(1.2) calculating the grayscale image g_iEdge gradient image Grad of_iThe formula is as follows:

in the formula, G_ixDenotes g_iGrayscale image of lateral edge detection, G_iyDenotes g_iGray level images of longitudinal edge detection;

(1.3) use of the threshold T_iaFor the edge gradient image Grad obtained in the step (1.2)_iPerforming binarization processing to obtain a mask image omega_iThreshold value T_iaThe calculation formula is as follows:

T_ia＝max(Grad_i)·0.2

(1.4) mask image-based ω_iObtaining a gray image g_iTarget area a of_iCalculating a_iHistogram of gray levels H_iIs a frequency variation value v_i[r]Obtaining an image adaptive segmentation threshold T_ibBy means of T_ibFor g_iDividing to be larger than threshold value T_ibIs a foreground region, less than a threshold value T_ibDefining the pixel value of the foreground area after segmentation as 1 and the pixel value of the background area as 0 to obtain a segmentation result f_i；

In the technical scheme, the foreground comprises ships and other targets, and the background comprises seawater and other targets. Other objects include clouds, islands, noise, etc.

Wherein, T_ibThe calculation formula is as follows:

in the formula, h_imeanRepresenting the mean value of the image grey frequencies, h_i[r]Denotes a frequency (r is 0,1, …,255) at a gray value of r, v_i[r]Indicates a frequency variation value (r ═ 0,1, …,255) at the gradation value r,representing the gray value at which the frequency variation value is minimum;

(1.5) Using 2X 2 inner core pairs f, respectively_iCorroding and expanding to obtain a coarse classification result image k_i。

(2) Image k of coarse classification result by rectangular template₁、k₂Matching, removing non-ship connected domains, and extracting a ship target area;

the method comprises the following specific steps:

(2.1) defining the upper left corner point of the image as an origin, and the right direction is the positive direction of an x axis and the downward direction is the positive direction of a y axis;

(2.2) calculating k₁、k₂Obtaining the length, the width and the angle of each connected domain according to the minimum circumscribed rectangle of each connected domain, wherein the length and the width are the length and the width of the corresponding minimum circumscribed rectangle, the angle is an included angle between the extension line of the long side of the minimum circumscribed rectangle and the positive direction of the x axis above the x axis, calculating the length-width ratio of the connected domains, and only keeping the connected domains with the length-width ratio of more than 1.5 and less than 15 to obtain the rest connected domains;

(2.3) calculating k separately₁、k₂The ship rectangular template of each residual connected domain is obtained to obtain k₁Template set TEM₁＝{temple_1,1,temple_1,2… } and k₂Template set TEM₂＝{temple_2,1,temple_2,2,…}；

Further, the calculation process of the template in step (2.3) is as follows:

the length and the angle of the foreground in the template are respectively the length and the angle of the corresponding connected domain, the ship length L is taken as an interval to divide real ship data, the average length-width ratio of the ship in each interval is counted, and the length-width ratio of the template foreground is determined according to the interval c where the approximate real length of the connected domain is locatedCalculating the foreground width in the template, wherein the calculation formula of the foreground width is as follows:

in the formula, length represents the foreground length in the template, e represents the number of ships in a ship length interval c in statistical data, delta represents the ship length-width ratio in the statistical data, and res represents the spatial resolution of the remote sensing image;

roughly classified image k according to foreground length, width and angle in template₁Each of which is connected withTemplate image template corresponding to domain generation_1,1、temple_1,2…, obtaining an aggregate TEM₁For the coarsely classified image k₂Generating corresponding template image temples for each connected domain_2,1、temple_2,2…, obtaining an aggregate TEM₂The pixel value of the foreground region in the template is 1, and the pixel value of the background region in the template is 0.

Further, the length L of the ship is 0-50 m.

When the L is 50 m, the reaction time is,the specific values of (a) are as follows:

(2.4) use of TEM₁For the roughly classified image k₁Template matching using TEM₂For the roughly classified image k₂Carrying out template matching;

further, the matching method in step (2.4) is as follows:

setting a threshold T_SThe template starts to slide from the original point of the image, one pixel is slid each time, the whole image is traversed, the similarity score S between the template and the coverage area of the template is calculated at each position, and if the similarity score is larger than T, the similarity score S is calculated_SIf the matching is successful, recording coordinates of four corner points of the template when the matching is successful, marking the coordinates as a ship target area, and if the same connected domain is marked for multiple times, only keeping the highest similarity score once as the ship target area;

wherein, the similarity score S formula is calculated as follows:

in the formula, sample represents a template, X represents a connected component, and N (-) represents the number of pixels having a pixel value of 1 in the region (-).

Further, the threshold T in step (2.4)_S＝0.8。

(2.5) expanding the ship target area to ensure that the ship is completely contained in the ship target area;

further, the extended formula in step (2.5) is as follows:

wherein (x'₁,y'₁)、(x'₂,y'₂) Coordinates (x) representing the extended upper left corner point and lower right corner point of the ship target area, respectively₁,y₁)、(x₂,y₂) Representing the corresponding coordinates of the target area of the original vessel.

(3) Inputting the four corner coordinates of the extended ship target area into a GrabCT algorithm, and accurately extracting a remote sensing image p₁、p₂The ship in (1) to obtain the extracted image z₁And z₂；

Further, the accurate extraction process of the ship in the step (3) is as follows:

(3.1) ships in the ship target area are the foreground to be divided, seawater and other targets are the background, and p is respectively paired in RGB space₁、p₂The foreground and the background are modeled by a Gaussian mixture model, pixels outside a ship target area are marked as the background, pixels inside the ship target area are marked as possible foreground or background, the Gaussian mixture model is initialized by using a K-means algorithm, K is taken to be 5, each pixel is distributed with a proper Gaussian component, and the Gaussian component of the nth pixel is expressed as K_n；

(3.2) calculating p separately₁、p₂Gibbs energy E (α, k, θ, z) of the whole image, formula is as follows:

in the formula, α represents opacity, α ∈ {1,0}, θ ═ { pi (α, K), μ (α, K), ∑ (α, K), α ═ 0,1, K ═ 1,. and K }, z represents a video gray value array, U represents a region energy function, pi represents a weight of a gaussian component, μ represents a mean vector of each gaussian component, Σ represents a covariance matrix, V represents a boundary energy function, γ ═ 50, C represents a set of a certain pixel point and an adjacent pixel point, m and n represent two neighborhood pixels, and β represents a contrast between video pixels;

(3.3) updating the remote sensing image p₁、p₂Of each pixel of the imageAnd parameters of the Gaussian mixture matrixUntil convergence, finding the minimum value of the energy function E (alpha, k, theta, z) and obtaining p₁、p₂Obtaining the accurate extracted image z of the ship by the optimal segmentation boundary of the foreground and the background₁、z₂。

(4) Calculating the extracted image z₁、z₂A feature vector for each vessel;

further, the calculating step in the step (4) is as follows:

calculating z₁、z₂The minimum bounding rectangle of each ship in the ship, the area of the ship, the length-width ratio lwr, the angle dir and the minimum bounding rectangle area are obtained_MERThereby obtaining a remote sensing image p₁、p₂Of each ship P1_w＝(rec_w,area_w,lwr_w,dir_w) And P2_j＝(rec_j,area_j,lwr_j,dir_j) Wherein w represents p₁The w ships extracted in (j) represents p₂Extracting the jth ship, rec represents the rectangular degree;

the calculation formula of the squareness rec is as follows:

(5) remote sensing image p by using ship feature vector₁、p₂And (5) matching the ships, and outputting a ship matching result.

Further, the matching process in step (5) is as follows:

(5.1) calculation of p₁The feature vector and p of each ship₂Establishing a bipartite graph between two image ships according to the similarity between the feature vectors of each ship, wherein the left vertex of the bipartite graph represents p₁Middle ship, right vertex denotes p₂Of a ship of (1), the similarity is greater than a threshold value T_dConnecting the ships, wherein the weight of the connection is the similarity between the characteristic vectors of the two ships;

further, the threshold value T in step (5.1)_d＝0.6。

Further, the similarity calculation formula in step (5.1) is as follows:

wherein R (P1)_w,P2_j) Represents p₁W of the second ship and p₂Similarity between the j-th ship in the middle, 0 ≦ R (P1)_w,P2_j)≤1，Andrespectively represent P1_wAnd P2_jIs measured.

(5.2) assigning an initial score to each vertex of the bipartite graph, wherein the initial score of the left vertex is the maximum weight of the side connected with the left vertex, and the initial score of the right vertex is 0;

(5.3) adding the scores of the top points at the left end and the right end of the connecting line, if the added score is more than or equal to the weight of the connecting line, considering that the two sides of the connecting line are the same ship, and when p is greater than or equal to the weight of the connecting line₁Two in middleA ship is matched to p₂When the ship is the same ship, the matching conflicts, a specific value d is set to be 0.1, the left vertex score corresponding to the two conflict connecting lines is reduced by d, and the right vertex score is increased by d;

and (5.4) repeating the step (5.3), when the left vertex score corresponding to the conflict connecting line is 0, abandoning the matching of the ship, continuing to match other ships until the matching of all the ships is completed, and outputting a ship matching result.

(III) advantageous effects

The method can effectively eliminate the interference of sea surface noise, cloud layers and other sea surface targets, can accurately extract the ship under a moving background and a fixed background, and can realize the automatic matching of the ship under the conditions of few ship characteristics and unclear texture;

the method for self-adaptive threshold value is provided for roughly classifying the optical remote sensing image, extracting possible foreground targets, then automatically generating a template to accurately position ships in the foreground targets, accurately extracting the ships through a GrabCT algorithm, reducing the detection range from rough to fine, avoiding interactive operation of the GrabCT algorithm, eliminating interference of cloud layers, other targets on the sea and noise on the sea, and realizing automatic detection of the ships; and constructing a bipartite graph between two image ships, and empowering the bipartite graph through the similarity between the feature vectors, so that the problem of few feature points of the ship is solved, and the automatic matching of the ship is realized.

Description of the drawings:

FIG. 1 is a schematic flow chart of the steps performed in the present invention;

FIG. 2 is a diagram of the results of the coarse classification according to the present invention;

FIG. 3 is a statistical chart of the ship length and the aspect ratio according to the statistics of the present invention;

FIG. 4 is a schematic view of a captain interval template made in accordance with the present invention;

FIG. 5 is a graph of the results of the ship extraction according to the present invention;

FIG. 6 is a schematic diagram of a ship match bipartite view of the present invention.

The specific implementation mode is as follows:

in order to make the objects, contents and advantages of the present invention clearer, the following detailed description of the embodiments of the present invention is made with reference to the accompanying drawings and examples:

referring to fig. 1, the method comprises the following specific steps:

(1) defining the optical satellite remote sensing image with the shooting time before as p₁The optical satellite remote sensing image with the later shooting time is p₂To p for_iRespectively carrying out coarse classification to obtain images k of coarse classification results_iI represents the ith video (i is 1, 2);

the rough classification process is as follows:

1.1) use of a Gaussian matrix pair p with a standard deviation of 0 and a size of 5X 5_iPerforming Gaussian filtering smoothing to convert the smoothed image into a gray image g_iThe formula is as follows:

g_i＝R_i·0.299+G_i·0.587+B_i·0.114

in the formula, R_i、G_i、B_iRespectively represents p_iAnd G, smoothing the red channel, the green channel and the blue channel by Gaussian, wherein the Gaussian matrix used by the Gaussian smoothing is as follows:

1.2) calculating the grayscale image g_iEdge gradient image Grad of_iThe formula is as follows:

in the formula, G_ixDenotes g_iGrayscale image of lateral edge detection, G_iyDenotes g_iGray level images of longitudinal edge detection;

1.3) Using the threshold T_iaFor the edge gradient image Grad obtained in the step 2)_iPerforming binarization processing to obtain a mask image omega_iThreshold value T_iaThe calculation formula is as follows:

T_ia＝max(Grad_i)·0.2

1.4) based on the mask image ω_iObtaining a gray image g_iTarget area a of_iCalculating a_iHistogram of gray levels H_iIs a frequency variation value v_i[r]Obtaining an image adaptive segmentation threshold T_ibBy means of T_ibFor g_iAnd (3) carrying out segmentation, wherein the pixel value of the foreground region is 1 after segmentation, the pixel value of the background region is 0, the foreground comprises ships and other targets, and the background comprises seawater and other targets, so that a segmentation result f is obtained_i；

Wherein, T_ibThe calculation formula is as follows:

in the formula, h_imeanRepresenting the mean value of the image grey frequencies, h_i[r]Denotes a frequency (r is 0,1, …,255) at a gray value of r, v_i[r]Indicates a frequency variation value (r ═ 0,1, …,255) at the gradation value r,representing the gray value at which the frequency variation value is minimum;

1.5) checking f with 2X 2 kernels, respectively_iCorroding and expanding to obtain a coarse classification result image k_iAn example is shown in fig. 2.

(2) Image k of coarse classification result by rectangular template₁、k₂Matching, removing non-ship connected domains, and extracting a ship target area;

the method comprises the following specific steps:

2.1) defining the upper left corner point of the image as an original point, the right direction as the positive direction of an x axis and the downward direction as the positive direction of a y axis;

2.2) calculating k₁、k₂And acquiring the length, the width and the angle of each connected domain according to the minimum external rectangle of each connected domain, wherein the length and the width are the length and the width of the corresponding minimum external rectangle, and the angle is the square of the long-side extension line of the minimum external rectangle and the x axisCalculating the length-width ratio of the connected domain at an included angle above the x axis, and only keeping the connected domain with the length-width ratio of more than 1.5 and less than 15 to obtain the residual connected domain;

2.3) calculating k separately₁、k₂The ship rectangular template of each residual connected domain is obtained to obtain k₁Template set TEM₁＝{temple_1,1,temple_1,2… } and k₂Template set TEM₂＝{temple_2,1,temple_2,2,…}；

The calculation process of the template is as follows:

the length and the angle of the foreground in the template are respectively the length and the angle of the corresponding connected domain, the ship length of 50 meters is taken as an interval to divide real ship data, the average length-width ratio of the ship of each interval is counted, and the length-width ratio of the template foreground is determined according to the interval c where the approximate real length of the connected domain is locatedCalculating the foreground width in the template, wherein the calculation formula of the foreground width is as follows:

in the formula, length represents the foreground length in the template, e represents the number of ships in the ship length interval c in the statistical data, delta represents the ship length-width ratio in the statistical data, res represents the spatial resolution of the remote sensing image,the values of (c) are shown in fig. 3, specifically as follows:

roughly classified image k according to length, width and angle of foreground in template₁Generating corresponding template image temples for each connected domain_1,1、temple_1,2…, obtaining an aggregate TEM₁To coarseClassified image k₂Generating corresponding template image temples for each connected domain_2,1、temple_2,2…, obtaining an aggregate TEM₂The pixel value of the foreground region in the template is 1, the pixel value of the background region in the template is 0, and the template example of each captain section is shown in fig. 4;

2.4) Using TEM₁For the roughly classified image k₁Template matching using TEM₂For the roughly classified image k₂Carrying out template matching, wherein the matching method comprises the following steps: setting a threshold T_SAnd (0.8), sliding the template from the origin of the image, sliding one pixel at a time, traversing the whole image, calculating a similarity score S between the template and the coverage area of the template at each position, and if the similarity score is more than T_SIf the matching is successful, recording coordinates of four corner points of the template when the matching is successful, marking the coordinates as a ship target area, and if the same connected domain is marked for multiple times, only keeping the highest similarity score once as the ship target area;

wherein, the similarity score S formula is calculated as follows:

in the formula, sample represents a template, X represents a connected domain, and N (-) represents the number of pixels with a pixel value of 1 in a region (-) and;

2.5) expanding the target area of the ship to ensure that the ship is completely contained in the target area of the ship, wherein the expansion formula is as follows:

wherein (x'₁,y'₁)、(x'₂,y'₂) Coordinates (x) representing the extended upper left corner point and lower right corner point of the ship target area, respectively₁,y₁)、(x₂,y₂) Representing the corresponding coordinates of the target area of the original vessel.

(3) Inputting the four corner coordinates of the extended ship target area into a GrabCT algorithm, and accurately extracting a remote sensing image p₁、p₂The ship in (1) to obtain the extracted image z₁And z₂；

The accurate ship extraction process comprises the following steps:

3.1) ships in the ship target area are the foreground to be divided, the seawater and other targets are the background, and p is respectively paired in RGB space₁、p₂The foreground and the background are modeled by a Gaussian mixture model, pixels outside a ship target area are marked as the background, pixels inside the ship target area are marked as possible foreground or background, the Gaussian mixture model is initialized by using a K-means algorithm, K is taken to be 5, each pixel is distributed with a proper Gaussian component, and the Gaussian component of the nth pixel is expressed as K_n；

3.2) calculating p separately₁、p₂Gibbs energy E (α, k, θ, z) of the whole image, formula is as follows:

in the formula, α represents opacity, α ∈ {1,0}, θ ═ { pi (α, K), μ (α, K), ∑ (α, K), α ═ 0,1, K ═ 1,. and K }, z represents a video gray value array, U represents a region energy function, pi represents a weight of a gaussian component, μ represents a mean vector of each gaussian component, Σ represents a covariance matrix, V represents a boundary energy function, γ ═ 50, C represents a set of a certain pixel point and an adjacent pixel point, m and n represent two neighborhood pixels, and β represents a contrast between video pixels;

3.3) updating the remote sensing image p₁、p₂Of each pixel of the imageAnd parameters of the Gaussian mixture matrixUntil convergence, finding the minimum value of the energy function E (alpha, k, theta, z) and obtaining p₁、p₂Obtaining the accurate extracted image z of the ship by the optimal segmentation boundary of the foreground and the background₁、z₂An example is shown in fig. 5.

(4) Calculating the extracted image z₁、z₂The feature vector of each ship in the method comprises the following calculation steps:

calculating z₁、z₂The minimum bounding rectangle of each ship in the ship, the area of the ship, the length-width ratio lwr, the angle dir and the minimum bounding rectangle area are obtained_MERThereby obtaining a remote sensing image p₁、p₂Of each ship P1_w＝(rec_w,area_w,lwr_w,dir_w) And P2_j＝(rec_j,area_j,lwr_j,dir_j) Wherein w represents p₁The w ships extracted in (j) represents p₂Extracting the jth ship, rec represents the rectangular degree;

the calculation formula of the squareness rec is as follows:

(5) remote sensing image p by using ship feature vector₁、p₂And (5) matching the ships, and outputting a ship matching result.

Wherein, the matching process is as follows:

5.1) calculating p₁The feature vector and p of each ship₂Establishing a bipartite graph between two image ships according to the similarity between the feature vectors of each ship, wherein the left vertex of the bipartite graph represents p₁Middle ship, right vertex denotes p₂Of a ship of (1), the similarity is greater than a threshold value T_dConnecting the vessels of which the number is 0.6, wherein the weight of the connecting line is the similarity between the feature vectors of the two vessels, and the similarity calculation formula is as follows:

wherein R (P1)_w,P2_j) Represents p₁W of the second ship and p₂Similarity between the j-th ship in the middle, 0 ≦ R (P1)_w,P2_j)≤1，Andrespectively represent P1_wAnd P2_jThe mean value of (a);

5.2) assigning an initial score to each vertex of the bipartite graph, wherein the initial score of the left vertex is the maximum weight of the side connected with the left vertex, and the initial score of the right vertex is 0, and the constructed bipartite graph is shown in FIG. 6;

5.3) adding the scores of the top points at the left end and the right end of the connecting line, if the added score is more than or equal to the weight of the connecting line, considering that the two sides of the connecting line are the same ship, and when p is greater than or equal to the weight of the connecting line₁Two ships are matched to p₂When the ship is the same ship, the matching conflicts, a specific value d is set to be 0.1, the left vertex score corresponding to the two conflict connecting lines is reduced by d, and the right vertex score is increased by d;

the case of matching conflict is illustrated: in fig. 6, when a3 selects a connection line with a weight of 0.9 to match b1, the connection line conflicts with a1 with a weight of 0.8, at this time, the scores of a1 and a3 are respectively reduced by 0.1, the score of b1 is increased by 0.1, and the connection line with a weight of 0.75 between a1 and b2 is found to meet the matching requirement, so that a1 is paired with b2, and a3 is paired with b1, so that the conflicts are solved;

5.4) repeating the step 5.3), when the left vertex score corresponding to the conflict connecting line is 0, abandoning the matching of the ship, continuing to match other ships until the matching of all ships is completed, and outputting the ship matching result.

After the matching operation, the scores of the left vertices in fig. 6 become 0.7, 0.75, 0.8, and 0.77, respectively, and the scores of the right vertices in fig. 6 become 0.1, 0, 0.1, and 0, respectively, and the ships that have succeeded in the final matching are (a1, b2), (a2, b4), (a3, b1), and (a4, b 3).